Alaamer/medium-articles-posts-with-content

0

null

0

null

2018-08-18

2018-08-18 09:12:11

2018-08-18

2018-08-18 09:12:23

0

false

en

2018-08-18

2018-08-18 09:12:23

1

1b9019a01aa8

2.69434

0

[Download] [PDF] Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) PDF EPUB KINDLE…

1

Download eBook Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Douglas C. Montgomery Epub #EPUB [Download] [PDF] Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) PDF EPUB KINDLE By Douglas C. Montgomery Link http://topmagazines.club/?book=1118780574 . . . . . . . . . . . . . . . . . . . Read Online PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Download PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Download Full PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Download PDF and EPUB Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read PDF ePub Mobi Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Reading PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read Book PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read online Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Download Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Douglas C. Montgomery pdf, Download Douglas C. Montgomery epub Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read pdf Douglas C. Montgomery Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Download Douglas C. Montgomery ebook Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read pdf Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Online Download Best Book Online Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read Online Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Book, Read Online Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) E-Books, Read Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Online, Read Best Book Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Online, Read Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Books Online Download Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Full Collection, Download Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Book, Read Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Ebook Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) PDF Read online, Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) pdf Download online, Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Read, Download Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Full PDF, Read Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) PDF Online, Read Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Books Online, Read Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Full Popular PDF, PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Read Book PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read online PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Download Best Book Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Read PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Collection, Read PDF Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) Full Online, Read Best Book Online Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics), Download Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in Probability and Statistics) PDF files

Download eBook Introduction to Linear Regression Analysis, Fifth Edition Set (Wiley Series in…

0

download-ebook-introduction-to-linear-regression-analysis-fifth-edition-set-wiley-series-in-1b9019a01aa8

2018-08-18

2018-08-18 09:12:23

https://medium.com/s/story/download-ebook-introduction-to-linear-regression-analysis-fifth-edition-set-wiley-series-in-1b9019a01aa8

false

714

null

Machine Learning

machine-learning

Machine Learning

51,320

fdadi31108

null

6a72c51f36c2

fdadi31108_30894

0

1

20,181,104

null

0

null

0

null

2018-09-20

2018-09-20 13:49:06

2018-09-20

2018-09-20 13:51:48

2

false

en

2018-09-20

2018-09-20 15:25:16

6

1b9058ae60c0

2.92673

1

0

By Daniel Burrus and Neil Smith

5

A Strategic Imperative for Anticipatory Leaders Is Cognitive Performance By Daniel Burrus and Neil Smith (In this blog series on how elevating cognitive performance is a game changer for organizations, I’ve invited Neil Smith, CTO at Think Outcomes, to join me in writing on this important topic due to his expertise and the cognitive performance software his firm has created.) AI and cognitive computing have grabbed headlines. Yet, anticipatory leaders know that the elevation of cognitive performance among teams is key to maximize results. Leaders need to help their teams of professionals improve how they envision opportunities, manage downside risks and achieve greater results. Cognitive computing has to do more than deliver data-driven insights to their minds. It must help teams shape outcomes, act on implications and professionalize role-based, cerebral processes in the form of software processes. That’s where cognitive performance is front and center. Cognitive performance involves how well professionals perform their cognitive work. Specifically, how they: establish vision identify problems ask questions of uncertainty arrive at critical thoughts analyze situations synthesize information reason judge solve problems communicate collaborate define follow-on actions They perform these cerebral activities with their thoughts and their communications. These mini processes in their minds are nondeterministic and lead to decisions within organizations. As machine learning and deep learning move into organizations, professionals who want to increase their cognitive performance must step up their game at the same time. They must center their attention on addressing uncertainties and advance their abilities to identify and create greater certainty. In doing so, they must raise their levels of quality in decision-making processes and stakeholder communication processes that take place in their minds. Their stakeholders, customers, suppliers, employees and their industries depend on it. The status quo of gut-based decision making and misunderstandings among viewpoints leads to operational inefficiencies and monetary waste in downstream activities. Change is accelerating in business, which creates more uncertainties that find their way into enterprises across all functional responsibilities — in strategies, integrations, operations, supply chains, human resources, research, engineering, finance, process management, product management and consulting, to name a few. Today, cognitive performance is based on role-based experience, learning, frequency, recency and luck — all of which vary from role to role and person to person. The cognitive activities in the minds of professionals are ripe for optimization. Optimization is possible by learning anticipatory skills and applying cognitive performance technologies. The human mind is limited when it is engaged to: structure decision data process situational information store organized knowledge recall situations with specificity understand alternative viewpoints engineer outcomes with greater clarity Although these are human limitations, the mind is extendable through the use of computing, which does a very good job of augmenting the mind for these activities. In today’s era of cognitive computing, the human mind can benefit from a digital extension to achieve the cognitive capabilities it cannot — and does not — realize on its own. At work, professionals who think for a living formulate how to execute their work in their minds. They’ve built their cognitive expertise over time through on-the-job experiences and homegrown cerebral processes. Business operations are both transactional and cognitive Before transactional software systems codified the operating processes of transactional work into the business infrastructure — i.e., ERP, SCM and CRM processes — organizations created homegrown processes and systems to manage their transactional operations. ERP, SCM and CRM systems optimized task-oriented processes before, during and after a customer transaction within organizations, in supply chains and in demand chains. As a result, the transactional side of the operating model has become relatively frictionless. Today’s friction exists within the minds of professionals on the cognitive side of the operating model. A key to future success is to eliminate this friction. That’s where anticipatory skills, combined with cognitive performance software, comes into play. Learn how to elevate your planning, accelerate innovation and transform results with The Anticipatory Learning System and how to maximize the cognitive performance of your team with Cognitive Performance Software.

A Strategic Imperative for Anticipatory Leaders Is Cognitive Performance

5

a-strategic-imperative-for-anticipatory-leaders-is-cognitive-performance-1b9058ae60c0

2018-09-20

2018-09-20 15:25:16

https://medium.com/s/story/a-strategic-imperative-for-anticipatory-leaders-is-cognitive-performance-1b9058ae60c0

false

674

null

Machine Learning

machine-learning

Machine Learning

51,320

Daniel Burrus

#1 Bestselling Author, Global Futurist, Innovation Expert and Keynote Speaker. One of the World’s Leading Futurists on Global Trends and Innovation.

a08cb0ce5697

DanielBurrus

5,573

1,878

20,181,104

null

0

null

0

75c6fcb1f120

2018-06-25

2018-06-25 08:59:44

2018-06-25

2018-06-25 09:27:39

4

false

no

2018-06-26

2018-06-26 18:54:53

0

1b910bec37db

2.118868

9

0

Hva blockchain er, er delt inn i tre hovedretninger. Det er forskjellig bruk av teknologien, men også faktisk forskjellige tekniske…

5

Blockchain og AI er direkte konkurrenter Hva blockchain er, er delt inn i tre hovedretninger. Det er forskjellig bruk av teknologien, men også faktisk forskjellige tekniske implementasjoner. Disse tre retningene har forskjellige verdiforslag, bruksområder og er rett og slett tre helt forskjellige ting. Dette gjør nok at mange synes at blockchain er forvirrende. De tre forskjellige retningene er kryptovaluta, smartkontrakt-plattformer og private Distributed Ledger Technologies (DLT). Kryptovalutaer, med Bitcoin i spissen, handler hovedsakelig om å kunne overføre digitale verdier mellom parter på en desentralisert og effektiv måte. Smartkontrakt-plattformer på den andre siden, er slik man ser for seg ‘clouden’ før man lærer at ‘clouden’ bare er servere hos Amazon. Ethereum, som den største smartkontakt-plattformen, er på en måte en desentralisert versjon av Amazon Web Services. Mens Amazon bygger opp under Web 2.0-tankegangen om sentralisert eierskap av data i siloer, bygger smartkontrakt-plattformene opp under Web 3.0, hvor data demokratiseres og frigjøres. Jeg holdt nettopp et foredrag om dette på Revisorforeningens DnR-dag. Hele opptaket er tilgjengelig på bunnen av siden. På scenen under Revisorforeningens DnR-dag Jeg argumenterer videre i foredraget at AI og blockchain er direkte konkurrenter. Da er det viktig å poengtere at jeg mener AI som “å konvertere ustrukturert data til strukturert data”, som er det mesteparten av AI idag handler om. Jeg snakker altså ikke om AI som brukes i selvkjørende biler osv. — det er noe annet. Så hvorfor er AI og blockchain konkurrenter? Hvis man ser på en handel, så vil AI kun endre en fraksjon av hvordan ting gjøres idag: bilags- og regnskapsføring blir enklere. Visualisering av en handel idag, med uthevet hva AI kan forbedre. Det kan ikke argumenteres for at denne bruken av AI er distrupsjon. Hvis man derimot ser på hvordan blockchain vil endre handel, så ser vi hvordan hele grunnlaget for hva AI fikser, forsvinner. Hvordan en handel vil se ut med blockchain Med blockchain er handelen og kvitteringen det samme. Med smartkontrakter gjøres regnskapsføringen automatisk, fordi handelen og regnskapet er også det samme. Innrapportering blir unødvendig, fordi Staten kontinuerlig har tilgang til alle handler som gjennomføres. Pengeoverføring gjennom et betalingsnettverk er innebygget i blockchain, så dette leddet forsvinner også. En krone til AI, er en krone til et gammelt system. Jeg mener derfor det er essensielt å bruke innovasjonsbudsjettene på den distruptive blockchain-teknologien, istedenfor enkle chatbots og bilder av kvitteringer. Se hele foredraget her:

Blockchain og AI er direkte konkurrenter

74

blockchain-og-ai-er-direkte-konkurrenter-1b910bec37db

2018-06-26

2018-06-26 18:54:53

https://medium.com/s/story/blockchain-og-ai-er-direkte-konkurrenter-1b910bec37db

false

376

Blockchangers is Norway's leading blockchain company, helping other's both understand and utilize this game changing phenomena. We do so through advisory, lectures, workshops and developing Proof of Concepts for our clients.

null

blockchangers

null

Blockchangers

mail@blockchangers.com

blockchangers

TECHNOLOGY,BLOCKCHAIN,IPFS,ETHEREUM,WEB3

blockchangers

Norwegian

norwegian

Norwegian

204

Jon Ramvi

CEO and founder of Blockchangers

9d8eb9833d81

ramvi

155

198

20,181,104

null

0

null

0

null

2018-08-29

2018-08-29 10:28:55

2018-08-29

2018-08-29 10:29:22

1

false

en

2018-08-29

2018-08-29 10:29:22

2

1b93a3e7f75d

4.166038

0

A few years ago, artificial intelligence was just the stuff of sci-fi. It was a novel and cool innovation that appeared to dwell in the…

1

Pros & cons of artificial intelligence in mobile app development A few years ago, artificial intelligence was just the stuff of sci-fi. It was a novel and cool innovation that appeared to dwell in the future. But today, it is a part of the mainstream, with its innovations included in almost every modern device. Starting from the chat bots to predictive analytics, organizations and developers are investigating innovative approaches to use artificial intelligence to deliver better client services, new items and reconsider business processes. AI and machine learning, both are causing an extraordinary change in the way that engineers, organizations, and clients consider modeling algorithms and intelligent communications with applications. There are different classifications or levels of AI in Mobile App Development Companies. It can be classified into strong and weak. Strong AI depicts programming that could mimic a human, in thought and activities. Weak just mimic human in activities. There is also restricted and general AI, this characterization depends on the extent of the capacity of the programming. And additionally, there are advantages and disadvantages of AI you should consider so your team can make an informed decision. Pros of Artificial intelligence Complete tedious tasks easily: There are unique classifications or ranges of AI in cellular app development companies. It may be labeled into strong and weak. Strong AI depicts programming that might mimic a human, in concept and activities. Weak just mimic human in activities. There is also constrained and preferred AI; this characterization relies upon on the extent of the ability of the programming. And additionally, there are advantages and disadvantages of AI you have to consider so your team could make an informed decision. Daily application: AI has observed use in our daily lives. In journey and tourism, AI programs can be utilized to locate areas a great deal results easily by making use of augmented truth. Augmented truth superimposes computer-generated pics on the actual international view of the customer, enhancing his/her view of the surrounding. This is moreover utilized in museums, making an allowance for greater vigorous customer enjoy. AI can likewise be utilized for predictive composition and revision of human spelling. In the posting on pictures via social networking sites, AI programming can apprehend and become aware of a man’s face and may successfully tag the person. Less error: Artificial intelligence machines are especially important in fields that require a very high degree of accuracy and precision. One of such location is area exploration. In space exploration, there aren’t any margins for blunders and any wrongly located number should spoil the complete procedure. Hence the want for synthetic intelligence machines to assist with the crunching of very complex numbers. Also, artificially clever robots are programmed to help with area exploration. These robots are constructed to resist the cruel atmosphere of space. They are acclimatized such that they cannot be damaged down, changed or disfigured with the aid of the opposed area environment. Daily application: Artificial intelligence has located use in our everyday lives. In tour and tourism, synthetic clever apps can be used to detect places lots without problems using augmented fact. Augmented truth superimposes pc-generated images on the real global view of the user, improving his/her perception of the encircling. This is likewise used in museums, allowing for extra sturdy purchaser enjoy. Artificial intelligence can also be used for predictive writing and correction of human spelling. In the posting on pictures on social media, synthetic intelligence programming can help to pick out and discover a person’s face and correctly tag the person. Medication: AI’s have an impact on in medication and hospitals are brilliant. Hospitals can leverage clever robots to treat and diagnose the ailment together with growing the awareness of ability side outcomes. They may even simulate surgical treatment process for training and education. Mining Process: Investigating for gasoline can be extra specific and green with help from AI robots. Such robots are constructed to manual miners through hard, grueling work by using presenting real-time instructions to improve the performance of people. Cons of Artificial intelligence: The absence of original creativity: When it involves creativeness and creativity for mobile application development companies, you could just find it in human beings. AI machines are capable of assist in designing and developing but would no longer be able to suppose. So the creativity part can be continually missing in AI generation and can simplest be observed in human beings. The AI devices do now not have the sensitivity and feeling this is just determined in people. Cost: Another major disadvantage of artificially intelligent machines for mobile application development organizations is cost. They are very expensive to build and significantly more expensive to manage and maintain. There is always the requirement for updates so they can coordinate the changing needs and conditions. Lack of judgment calls: No be counted how smart an AI robotic is, it will usually lack the capability to make decisions related to people and based on feelings. There are instances while human judgment is critical and selections can’t be based totally on calculations and algorithms. AI Implementation is challenging: Despite all its astonishing blessings, its miles true that nearly, imposing AI solutions and smart machines is a major pain point. For a mobile app developer, the usage of and adding AI capabilities to several dynamic elements might be a baffling and grueling task. The reason is the requirement of ideal harmony and coordination among moving elements to make sure enticing, seamless person experience. No chance in experience: In general people, mastering comes after experience. But, this isn’t so with AI machines in mobile utility improvement enterprise. Obviously, they are able to save loads of information; but they couldn’t get admission to these facts within the similar manner as people do. They can’t adjust their responses and reactions as in line with the changing surroundings. They can’t make a distinction between somebody who’s hardworking and any person who isn’t always. They do no longer have the human touch which is critical for powerful residing. Hire best AI developer with a leading company, Data EximIT who are serving the clients with their best services. With immense experience and skilled software developers the company do incredible job. It is a clever decision to get in touch with them for best solutions!

Pros & cons of artificial intelligence in mobile app development

0

pros-cons-of-artificial-intelligence-in-mobile-app-development-1b93a3e7f75d

2018-08-29

2018-08-29 10:29:22

https://medium.com/s/story/pros-cons-of-artificial-intelligence-in-mobile-app-development-1b93a3e7f75d

false

1,051

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Data EximIT

Web Development, Mobile App Design and Development, UI/UX Design, Blockchain Development, M-Commerce and Digital Marketing

b22dcc08f707

dataeximit

0

2

20,181,104

null

0

null

0

null

2017-10-09

2017-10-09 16:53:22

2017-10-09

2017-10-09 17:16:22

9

false

en

2017-10-24

2017-10-24 09:08:24

0

1b945d169827

4.856604

4

0

First, we load the data:

5

Political analysis using R: correlation between polity and female power in the world Courtesy of Transcendence Systems First, we load the data: irdata <- read.csv(“IRdata.csv”, header=TRUE) We can determine the header names using the names function as shown: names(irdata) The output is: [1] “country” “year” “ifs” “region” “capital” “income” “agri.land” [8] “arable.land” “birth.rate” “fertility” “ado.fertility” “CO2” “death.rate” “life” [15] “life.f” “life.m” “pop” “pop.density” “pop.y” “pop.s” “pop.work” [22] “pop.f” “pop.growth” “pop.rural” “pop.urban” “pop.urban.growth” “oil” “forest” [29] “savings” “land” “export” “FDI” “GDP” “GDPpc” “growth” [36] “gov.expen” “infant.mort” “inflation” “mobile” “resources” “aid” “reserves” [43] “telephone” “trade” “journal” “remittances” “trade.service” “trademark” “cinc” [50] “milex” “milper” “energy.con” “xrate” “bkcrisis” “dc.private” “fs.deposit” [57] “deposit” “pliberty” “cliberty” “demo” “auto” “polity” “durable” [64] “xconst” “polcomp” “cspart” “female.power” “corruption” “freexp” “clean.election” [71] “system” “yrsoffc” “yrcurnt” “finittrm” “military” “legis.seats” “intviol” [78] “intwar” “civviol” “civwar” “ethviol” “ethwar” “actotal” “nborder” [85] “peacekeeping” “BITs” “NATO” “WTO” “IMF” “OECD” “EU” [92] “discrimpop” “regautpop” “lang1” “lang2” “islands” “dist_US” “Christian” [99] “Muslim” “Hindu” “Buddhist” “Taoist” “no.religion” To show the first few columns and the header, we use head as shown: head(irdata) Sample output is: country year ifs region capital income agri.land arable.land 1 Afghanistan 1960 AFG South Asia Kabul Low income NA NA 2 Afghanistan 1961 AFG South Asia Kabul Low income 377000 11.71767 3 Afghanistan 1962 AFG South Asia Kabul Low income 377600 11.79426 4 Afghanistan 1963 AFG South Asia Kabul Low income 378100 11.87085 5 Afghanistan 1964 AFG South Asia Kabul Low income 378730 11.94743 6 Afghanistan 1965 AFG South Asia Kabul Low income 378750 11.94743 birth.rate fertility ado.fertility CO2 death.rate life 1 51.276 7.45 145.321 414.3710000000000 32.403 32.32851 2 51.374 7.45 145.321 491.3780000000000 31.902 32.77744 3 51.464 7.45 145.321 689.3960000000000 31.415 33.21990 4 51.544 7.45 145.321 707.7310000000000 30.937 33.65788 5 51.614 7.45 145.321 839.7430000000001 30.464 34.09288 6 51.668 7.45 145.321 1008.4250000000000 29.992 34.52539 life.f life.m pop pop.density pop.y pop.s pop.work pop.f 1 33.105 31.589 8994793 NA 42.17061 2.798986 55.03040 48.30661 2 33.557 32.035 9164945 14.03815 42.47375 2.808451 54.71780 48.39668 3 34.001 32.476 9343772 14.31206 42.64217 2.803814 54.55402 48.48131 4 34.440 32.913 9531555 14.59969 42.73415 2.785002 54.48085 48.56059 5 34.875 33.348 9728645 14.90158 42.82173 2.751894 54.42638 48.63462 6 35.307 33.781 9935358 15.21821 42.93593 2.705066 54.35901 48.70356 pop.growth pop.rural pop.urban pop.urban.growth oil forest 1 1.816077 91.77900000000000 8.221000000000000 5.256151 NA NA 2 1.876528 91.49200000000000 8.507999999999999 5.307999 NA NA 3 1.934999 91.19499999999999 8.805000000000000 5.366358 NA NA 4 1.992521 90.89000000000000 9.109999999999999 5.397755 NA NA 5 2.049423 90.57400000000000 9.426000000000000 5.459349 NA NA 6 2.105369 90.25000000000000 9.750000000000000 5.484886 NA NA In order to get a better grasp and view of the data, we use the fix function. This enables us to view the data and explore it in an easier manner. fix(irdata) Sample output is: Empirical analysis, descriptive analysis and exploratory data analysis The next step of Exploratory Data Analysis (EDA), is to determine the occurrences polity levels. To start off, concentration of the analysis shall be on the individual variables. After the variables have been examined independently, they are their relationship with each other is examined. Univariate variable empirical, descriptive and exploratory data science Polity variable The polity value is the independent variable in determining the correlation between polity and female power in a country. Then source of the data is from the provided IRdata dataset. First off, we can an overview of the data using the summary function as shown: summary(irdata$polity) The output is: Min. 1st Qu. Median Mean 3rd Qu. Max. NA’s -10.0000 -7.0000 1.0000 0.8937 8.0000 10.0000 2769 To represent the values visually, we use the boxplot function as shown: boxplot(irdata$polity,ylab=”Polity”) The output is: For the mean, the code is: polity.cleaned<-na.omit(irdata$polity) polity.dens<-density(polity.cleaned,from=0) plot(polity.dens,main=”Polity”) abline(v=mean(polity.cleaned),col=”red”) Here is the output: To better view the distribution of the actual values, a histogram should be plotted. This can be done using the command: hist(irdata$polity,xlab=”polity”,main=””) This produces the output: Additionally, we can determine the dispersion of data using variance and standard deviation from the functions var and sd as shown: var (polity.cleaned) sd(polity.cleaned) The output is: 55.34104 7.439156 From the above figures and results, the variable polity is positively skewed with both a relatively high variance and standard deviation. Female power variable First, the summary of the data is provided as shown: summary(irdata$female.power) The output is: Min. 1st Qu. Median Mean 3rd Qu. Max. NA’s 0.059 0.418 0.593 0.583 0.769 0.969 3655 Then, we plot the values of female power. Since it contains NA values, it needs to be cleaned to work with boxplot. This is done by using the na.omit function and assigning it to a variable named female.power.cleaned. The summary of the variable distribution can then be plotted as shown: female.power.cleaned=na.omit(irdata$female.power) boxplot(female.power.cleaned,ylab=”Female Power”) The output is: To plot the mean, the code is as shown: irdata<-read.csv(“IRdata.csv”) female.power.cleaned<-na.omit(irdata$female.power) female.power.dens<-density(female.power.cleaned,from=0) cat(female.power.dens) plot(female.power.dens,main=”Female power”) abline(v=mean(female.power.cleaned),col=”red”) The output is: Then we run a command to determine the frequency of female power values: hist(irdata$female.power,xlab=”female power”,main=””) The output is: To determine the dispersion of data, the variance and standard deviation is calculated using the var and sd functions as shown: var(female.power.cleaned) sd(female.power.cleaned) The output is: 0.05003542 0.223686 From the above figures and results, the variable female power is uniformly distributed with low variance and standard deviation. Bivariable empirical, descriptive and exploratory data analysis First off, it is important to determine the correlation between the independent variable polity and the dependent variable female power. This can be done by subsetting the data frame holding the file values and omitting the rows where either or both polity and female power have an NA value. This is then used to determine the correlation between variables as shown: polity.female.power<-subset(irdata,select=c(country,polity,female.power) polity.female.power<-na.omit(polity.female.power) cor(polity.female.power$polity,polity.female.power$female.power) The output is: 0.7023152 Next, we can plot the data of polity versus female power using the plot function as shown: plot(x=irdata$polity+10,y=irdata$female.power,xlab=”Polity (0–20)”,ylab=”Female Power”) The output is: From this, we can determine that polity and female power are correlated. Thank you for your time.

Political analysis using R: correlation between polity and female power in the world

8

political-analysis-using-r-correlation-between-polity-and-female-power-in-the-world-1b945d169827

2018-04-13

2018-04-13 18:21:30

https://medium.com/s/story/political-analysis-using-r-correlation-between-polity-and-female-power-in-the-world-1b945d169827

false

969

null

Data Science

data-science

Data Science

33,617

David Kabii

null

ccfafa4b74bb

david.kabii

21

25

20,181,104

null

0

null

0

43a68f94549b

2018-03-06

2018-03-06 20:15:10

2018-03-29

2018-03-29 10:47:16

4

false

en

2018-04-28

2018-04-28 17:04:38

1

1b94a0ca429b

3.503774

3

0

How emerging technology is empowering people in new ways.

5

Technology, and its power to empower. How emerging technology is empowering people in new ways. In 2007, Bob Offereins started losing his sight as his retinal cells began to degenerate. It was a hereditary disease and not something he was expecting or had planned for. He was only 23 at the time. 90% of all information transmitted to and processed by our brain is visual information. This includes texts, images, faces, objects and practically every other thing that you are currently seeing. Close your eyes for just a moment, and imagine all the things you did in your day upto this point. And now imagine how would you have done them all with your eyes closed. If you had to shop in supermarket, catch a train, cook a meal or get a coffee with friend, without sight. And that is what it comes down to. Losing vision is losing access to 90% of all information that you otherwise have access to. When you lose access to information, you lose independence. Header Image showing a view of a street but blocked with a brush stroke. As soul-crushing as that experience was, he decided to not let it dictate how he lives. He fuelled his dreams with determination and used technology as a tool to propel him forward. The technology at the time was relatively primitive, but useful. He mainly depended on magnifiers and monoculars to read and see things he could not. Then came the advent of camera phones, which allowed him to take pictures of things on the go and zoom into them to see details. Slowly, the phones started getting smarter and that’s when things started to get really interesting. Bob’s old magnifier and monocular on a table with an iPhone. The technology of the past decade facilitated Bob to be a consumer of information around him, not a creator. He still had a feeling of not contributing enough. He recalls an old Dutch proverb: “achter de geraniums zitten” which translates as “to sit behind the geraniums”. Geraniums are flowers that are typically put in boxes, below the window at the facade of a Dutch house. So if you are behind these geraniums, you are inside the house, not being active. Something Bob just did not want to do or be. That’s where new age technology comes in, because it provides users the opportunity to move from being a mere consumer to a creator. They are no longer products, but tools. Tools that can give the users control and empower them to be better versions of themselves. As the saying goes, “Give a man a fish, and you feed him for a day. Teach a man to fish, and you feed him for a lifetime”. The technology solutions of today are proving to be a fishing rod in this equation. Illustration showing transition from product to tool, consumer to creator and fish to fishing rod. That’s what enticed Bob most about Envision ever since he first heard of it. It provides him with an opportunity to step out of the house and control the way he accesses information. He thinks it’s impractical and unrealistic to expect every piece of visual information out there to be structurally built for visually impaired people. There are only so many braille signboards that one can put up. A tool like Envision, however, flips the perspective on it by giving the visually impaired people the ability to access any kind of visual information that is out there. With Envision Bob can now read everything, from timetables at the train station to the chalkboard at the cafe down the street, without any of them having to make any special accommodations for him. When world is covered with thorns, it’s better to put on shoes than try covering the whole world in carpet. A photo of a person looking down at his shoes. “In the end it’s not about more diversity, but more inclusiveness”, he says, “I do not need special treatment because of my impairment. All I want is a tool and an opportunity to have equal access to information that everyone else has.” A video by Bob Offereins summarising his thoughts about inclusiveness and Envision. Thank you for reading. This post was written by Karthik Mahadevan and Bob Offereins for the inaugural edition of Envision User Stories. Envision is a visual recognition tool that enable visually impaired users to read texts from any surface in multiple languages, recognise faces and objects, get a description of scenes they are in and more. With Envision, visually impaired users can shop in supermarkets, use public transport, read menu cards in restaurants, recognise their friends, find their belongings and so much more, all on their own. You can get a free trial of Envision by downloading it from the AppStore: https://itunes.apple.com/app/envision-ai/id1268632314

Technology, and its power to empower.

105

technology-and-its-power-to-empower-1b94a0ca429b

2018-04-28

2018-04-28 17:04:39

https://medium.com/s/story/technology-and-its-power-to-empower-1b94a0ca429b

false

743

Enabling vision for all

null

letsenvision

null

LetsEnvision

karthik@letsenvision.com

letsenvision

ARTIFICIAL INTELLIGENCE,ACCESSIBILITY,INCLUSIVE DESIGN,VISUAL IMPAIRMENT,ASSISTIVE TECHNOLOGY

letsenvision

Technology

technology

Technology

166,125

Lets Envision

Enabling vision for all

5c74b56b83d

karthik_51596

4

1

20,181,104

null

0

null

0

6a619176a69f

2018-06-13

2018-06-13 09:37:41

2018-06-13

2018-06-13 10:26:26

5

false

en

2018-06-13

2018-06-13 10:26:26

0

1b94c2e7c3d

3.331447

0

IOE (Internet Of Everything) is the new IOT (Internet Of Things)

5

Why Is It (IoT) Internet Of Everything (IoE)? IOE (Internet Of Everything) is the new IOT (Internet Of Things) IOE (Internet Of Everything) is the new IOT (Internet Of Things) With the whole world flocking towards IoT market, businesses are rushing themselves to produce more and more smart devices. Even though its first foundation was laid in the late 1980s by students from Carnegie Mellon University, the actual boom began when the thought model for future interconnection environment was proposed, in 2004. Today, it’s not just our phones, desktops or tablets can be interlinked but the other multitude of devices like washing machines, electric toasters, vacuum cleaners, and whatever you name it. The kind of life we were thinking was real when as a kid were shown Sym-bionic Titan and Futurama? has begun to become real now indeed. As a kid I loved Zoidberg. I called him ‘Fingers On The Lip Guy’! Well, none of it are crazy anymore. Want to see what’s happening inside your home? even when you’re 1000 miles away, turn on the app, surf around your home with 360 degree VR! This IoT landscape is growing faster and has already covered most of the verticals we humans have created — from home appliances to facilities management. After gauging its growth and usage for several years now, some of the researchers have made its astounding future predictions. 1. Breakthrough in e-commerce industry Once smart technology gets to be at the helm, every retail stores will become more accessible and more trouble-free for customers. The key idea behind supporting such technological disruption is that we all believe in its capacity to give us a tranquil shopping experience. Source: Grazitti Interactive As the customer ourselves even after reading and knowing all about IoT, we won’t be able to figure out when and how the changes shall appear. “It mostly will be in a way that something is done different but can’t see what it is” It is soon evolving into the cornerstone of retail, that all we would feel are the products and appliances ‘coming live’ into invariably beautiful and innovative experiences. Also, the real revolution it is when we have ‘indirect smart technology’ existing around the us. Source: lukehannon.ie/ Less the friction is in a customer’s journey, greater the conversion rate would be. Thus to convert/conquer more consumer, the shopping journey has to be made hassle free by employing essential amount of energy, resources, people, and products within the store. Mainly by bringing in things like high-fidelity IR arrays and UHF RFID. 2. Advancements in consumer-facing industries As the consumer interests in connected devices, applications, and services are slowly touching the roof, the demand for IoT and connectivity will keep driving the fundamental shift in the way the consumer-facing industries interact with their visitors/customers. “This will completely shift some of the familial spaces like retail and insurance” For ex: Amazon being the pioneer within retail industry to implement IoT in its native device Alexa, it will continue to overpower other retailers. This will be the greatest challenge for other emerging retailers to gain back their ROI that they necessarily have to shift from product into services to keep their businesses afloat. And the insurance sector have to quickly adapt to IoT since the data that are to be generated by IoT will fundamentally redefine the way they calculate and mitigate risk. 3. EDGE (Enhanced Data For Global Evolution) will become the king! Reason being some of the IoT devices have higher potential to ground existing networks, it is high-time that EDGE starts working on processing, reducing, and analysing data before it hits the internet. Source: Mobile Communications Our cameras collect relatively high volume of data and the new jet engines are loaded with sensors that generate more than 10 gigabytes/s when running and terabytes per flight. Even our cars are now recording several informations. Thus, one part of the global IoT network that has to be upgraded double quick is EDGE (the margin between IoT devices and the computers on the internet).

Why Is It (IoT) Internet Of Everything (IoE)?

0

why-is-it-iot-internet-of-everything-ioe-1b94c2e7c3d

2018-06-13

2018-06-13 10:26:28

https://medium.com/s/story/why-is-it-iot-internet-of-everything-ioe-1b94c2e7c3d

false

662

AI-based automation solution. Skitter helps companies reduce cost and increase productivity by providing cognitive and automation solutions.

null

CogniDesk

null

CogniDesk

Info@skitter.in

cognidesk

ARTIFICIAL INTELLIGENCE,AUTOMATION,SERVICE DESK,HELPDESK,ITSM

CogniDesk

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Steffi S

Technology writer | AI practitioner.

3f4160f48193

steffi.Netizen

1,334

140

20,181,104

null

0

class Char3Model(nn.Module): def __init__(self, vocab_size, n_fac): super().__init__() self.e = nn.Embedding(vocab_size, n_fac) # The 'green arrow' from our diagram self.l_in = nn.Linear(n_fac, n_hidden) # The 'orange arrow' from our diagram self.l_hidden = nn.Linear(n_hidden, n_hidden) # The 'blue arrow' from our diagram self.l_out = nn.Linear(n_hidden, vocab_size) def forward(self, c1, c2, c3): in1 = F.relu(self.l_in(self.e(c1))) in2 = F.relu(self.l_in(self.e(c2))) in3 = F.relu(self.l_in(self.e(c3))) h = V(torch.zeros(in1.size()).cuda()) h = F.tanh(self.l_hidden(h+in1)) h = F.tanh(self.l_hidden(h+in2)) h = F.tanh(self.l_hidden(h+in3)) return F.log_softmax(self.l_out(h)) class CharLoopModel(nn.Module): # This is an RNN! def __init__(self, vocab_size, n_fac): super().__init__() self.e = nn.Embedding(vocab_size, n_fac) self.l_in = nn.Linear(n_fac, n_hidden) self.l_hidden = nn.Linear(n_hidden, n_hidden) self.l_out = nn.Linear(n_hidden, vocab_size) def forward(self, *cs): bs = cs[0].size(0) h = V(torch.zeros(bs, n_hidden).cuda()) for c in cs: inp = F.relu(self.l_in(self.e(c))) h = F.tanh(self.l_hidden(h+inp)) return F.log_softmax(self.l_out(h), dim=-1) class CharRnn(nn.Module): def __init__(self, vocab_size, n_fac): super().__init__() self.e = nn.Embedding(vocab_size, n_fac) self.rnn = nn.RNN(n_fac, n_hidden) self.l_out = nn.Linear(n_hidden, vocab_size) def forward(self, *cs): bs = cs[0].size(0) h = V(torch.zeros(1, bs, n_hidden)) inp = self.e(torch.stack(cs)) outp,h = self.rnn(inp, h) return F.log_softmax(self.l_out(outp[-1]), dim=-1) class CharSeqStatefulRnn(nn.Module): def __init__(self, vocab_size, n_fac, bs): self.vocab_size = vocab_size super().__init__() self.e = nn.Embedding(vocab_size, n_fac) self.rnn = nn.RNN(n_fac, n_hidden) self.l_out = nn.Linear(n_hidden, vocab_size) self.init_hidden(bs) def forward(self, cs): bs = cs[0].size(0) if self.h.size(1) != bs: self.init_hidden(bs) outp,h = self.rnn(self.e(cs), self.h) self.h = repackage_var(h) return F.log_softmax(self.l_out(outp), dim=-1).view(-1, self.vocab_size) def init_hidden(self, bs): self.h = V(torch.zeros(1, bs, n_hidden)) def repackage_var(h): return Variable(h.data) if type(h) == Variable else tuple(repackage_var(v) for v in h) from torchtext import vocab, data from fastai.nlp import * from fastai.lm_rnn import * PATH='data/nietzsche/' TRN_PATH = 'trn/' VAL_PATH = 'val/' TRN = f'{PATH}{TRN_PATH}' VAL = f'{PATH}{VAL_PATH}' %ls {PATH} models/ nietzsche.txt trn/ val/ %ls {PATH}trn trn.txt TEXT = data.Field(lower=True, tokenize=list) bs=64; bptt=8; n_fac=42; n_hidden=256 FILES = dict(train=TRN_PATH, validation=VAL_PATH, test=VAL_PATH) md = LanguageModelData.from_text_files(PATH, TEXT, **FILES, bs=bs, bptt=bptt, min_freq=3) len(md.trn_dl), md.nt, len(md.trn_ds), len(md.trn_ds[0].text) (963, 56, 1, 493747) class CharSeqStatefulRnn(nn.Module): def __init__(self, vocab_size, n_fac, bs): self.vocab_size = vocab_size super().__init__() self.e = nn.Embedding(vocab_size, n_fac) self.rnn = nn.RNN(n_fac, n_hidden) self.l_out = nn.Linear(n_hidden, vocab_size) self.init_hidden(bs) def forward(self, cs): bs = cs[0].size(0) if self.h.size(1) != bs: self.init_hidden(bs) outp,h = self.rnn(self.e(cs), self.h) self.h = repackage_var(h) return F.log_softmax(self.l_out(outp), dim=-1).view(-1, self.vocab_size) def init_hidden(self, bs): self.h = V(torch.zeros(1, bs, n_hidden)) m = CharSeqStatefulRnn(md.nt, n_fac, 512).cuda() opt = optim.Adam(m.parameters(), 1e-3) fit(m, md, 4, opt, F.nll_loss) def RNNCell(input, hidden, w_ih, w_hh, b_ih, b_hh): return F.tanh(F.linear(input, w_ih, b_ih) + F.linear(hidden, w_hh, b_hh)) class CharSeqStatefulRnn2(nn.Module): def __init__(self, vocab_size, n_fac, bs): super().__init__() self.vocab_size = vocab_size self.e = nn.Embedding(vocab_size, n_fac) self.rnn = nn.RNNCell(n_fac, n_hidden) self.l_out = nn.Linear(n_hidden, vocab_size) self.init_hidden(bs) def forward(self, cs): bs = cs[0].size(0) if self.h.size(1) != bs: self.init_hidden(bs) outp = [] o = self.h for c in cs: o = self.rnn(self.e(c), o) outp.append(o) outp = self.l_out(torch.stack(outp)) self.h = repackage_var(o) return F.log_softmax(outp, dim=-1).view(-1, self.vocab_size) def init_hidden(self, bs): self.h = V(torch.zeros(1, bs, n_hidden)) def GRUCell(input, hidden, w_ih, w_hh, b_ih, b_hh): gi = F.linear(input, w_ih, b_ih) gh = F.linear(hidden, w_hh, b_hh) i_r, i_i, i_n = gi.chunk(3, 1) h_r, h_i, h_n = gh.chunk(3, 1) resetgate = F.sigmoid(i_r + h_r) inputgate = F.sigmoid(i_i + h_i) newgate = F.tanh(i_n + resetgate * h_n) return newgate + inputgate * (hidden - newgate) class CharSeqStatefulGRU(nn.Module): def __init__(self, vocab_size, n_fac, bs): super().__init__() self.vocab_size = vocab_size self.e = nn.Embedding(vocab_size, n_fac) self.rnn = nn.GRU(n_fac, n_hidden) self.l_out = nn.Linear(n_hidden, vocab_size) self.init_hidden(bs) def forward(self, cs): bs = cs[0].size(0) if self.h.size(1) != bs: self.init_hidden(bs) outp,h = self.rnn(self.e(cs), self.h) self.h = repackage_var(h) return F.log_softmax(self.l_out(outp), dim=-1).view(-1, self.vocab_size) def init_hidden(self, bs): self.h = V(torch.zeros(1, bs, n_hidden)) from fastai import sgdr n_hidden=512 class CharSeqStatefulLSTM(nn.Module): def __init__(self, vocab_size, n_fac, bs, nl): super().__init__() self.vocab_size,self.nl = vocab_size,nl self.e = nn.Embedding(vocab_size, n_fac) self.rnn = nn.LSTM(n_fac, n_hidden, nl, dropout=0.5) self.l_out = nn.Linear(n_hidden, vocab_size) self.init_hidden(bs) def forward(self, cs): bs = cs[0].size(0) if self.h[0].size(1) != bs: self.init_hidden(bs) outp,h = self.rnn(self.e(cs), self.h) self.h = repackage_var(h) return F.log_softmax(self.l_out(outp), dim=-1).view(-1, self.vocab_size) def init_hidden(self, bs): self.h = (V(torch.zeros(self.nl, bs, n_hidden)), V(torch.zeros(self.nl, bs, n_hidden))) m = CharSeqStatefulLSTM(md.nt, n_fac, 512, 2).cuda() lo = LayerOptimizer(optim.Adam, m, 1e-2, 1e-5) on_end = lambda sched, cycle: save_model(m, f'{PATH}models/cyc_{cycle}') cb = [CosAnneal(lo, len(md.trn_dl), cycle_mult=2, on_cycle_end=on_end)] fit(m, md, 2**4-1, lo.opt, F.nll_loss, callbacks=cb) def get_next(inp): idxs = TEXT.numericalize(inp) p = m(VV(idxs.transpose(0,1))) r = torch.multinomial(p[-1].exp(), 1) return TEXT.vocab.itos[to_np(r)[0]] def get_next_n(inp, n): res = inp for i in range(n): c = get_next(inp) res += c inp = inp[1:]+c return res print(get_next_n('for thos', 400)) for those the skemps), or imaginates, though they deceives. it should so each ourselvess and new present, step absolutely for the science." the contradity and measuring, the whole! 293. perhaps, that every life a values of blood of intercourse when it senses there is unscrupulus, his very rights, and still impulse, love? just after that thereby how made with the way anything, and set for harmless philos from fastai.conv_learner import * PATH = "data/cifar10/" os.makedirs(PATH,exist_ok=True) classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') stats = (np.array([ 0.4914 , 0.48216, 0.44653]), np.array([ 0.24703, 0.24349, 0.26159])) def get_data(sz,bs): tfms = tfms_from_stats(stats, sz, aug_tfms=[RandomFlipXY()], pad=sz//8) return ImageClassifierData.from_paths(PATH, val_name='test', tfms=tfms, bs=bs) bs=256 data = get_data(32,bs) lr=1e-2 class SimpleNet(nn.Module): def __init__(self, layers): super().__init__() self.layers = nn.ModuleList([ nn.Linear(layers[i], layers[i + 1]) for i in range(len(layers) - 1)]) def forward(self, x): x = x.view(x.size(0), -1) for l in self.layers: l_x = l(x) x = F.relu(l_x) return F.log_softmax(l_x, dim=-1) learn = ConvLearner.from_model_data(SimpleNet([32*32*3, 40,10]), data) learn, [o.numel() for o in learn.model.parameters()] (SimpleNet( (layers): ModuleList( (0): Linear(in_features=3072, out_features=40) (1): Linear(in_features=40, out_features=10) ) ), [122880, 40, 400, 10]) learn.summary() OrderedDict([('Linear-1', OrderedDict([('input_shape', [-1, 3072]), ('output_shape', [-1, 40]), ('trainable', True), ('nb_params', 122920)])), ('Linear-2', OrderedDict([('input_shape', [-1, 40]), ('output_shape', [-1, 10]), ('trainable', True), ('nb_params', 410)]))]) learn.lr_find() learn.sched.plot() %time learn.fit(lr, 2) A Jupyter Widget [ 0. 1.7658 1.64148 0.42129] [ 1. 1.68074 1.57897 0.44131] CPU times: user 1min 11s, sys: 32.3 s, total: 1min 44s Wall time: 55.1 s %time learn.fit(lr, 2, cycle_len=1) A Jupyter Widget [ 0. 1.60857 1.51711 0.46631] [ 1. 1.59361 1.50341 0.46924] CPU times: user 1min 12s, sys: 31.8 s, total: 1min 44s Wall time: 55.3 s class ConvNet(nn.Module): def __init__(self, layers, c): super().__init__() self.layers = nn.ModuleList([ nn.Conv2d(layers[i], layers[i + 1], kernel_size=3, stride=2) for i in range(len(layers) - 1)]) self.pool = nn.AdaptiveMaxPool2d(1) self.out = nn.Linear(layers[-1], c) def forward(self, x): for l in self.layers: x = F.relu(l(x)) x = self.pool(x) x = x.view(x.size(0), -1) return F.log_softmax(self.out(x), dim=-1) learn = ConvLearner.from_model_data(ConvNet([3, 20, 40, 80], 10), data) learn.summary() OrderedDict([('Conv2d-1', OrderedDict([('input_shape', [-1, 3, 32, 32]), ('output_shape', [-1, 20, 15, 15]), ('trainable', True), ('nb_params', 560)])), ('Conv2d-2', OrderedDict([('input_shape', [-1, 20, 15, 15]), ('output_shape', [-1, 40, 7, 7]), ('trainable', True), ('nb_params', 7240)])), ('Conv2d-3', OrderedDict([('input_shape', [-1, 40, 7, 7]), ('output_shape', [-1, 80, 3, 3]), ('trainable', True), ('nb_params', 28880)])), ('AdaptiveMaxPool2d-4', OrderedDict([('input_shape', [-1, 80, 3, 3]), ('output_shape', [-1, 80, 1, 1]), ('nb_params', 0)])), ('Linear-5', OrderedDict([('input_shape', [-1, 80]), ('output_shape', [-1, 10]), ('trainable', True), ('nb_params', 810)]))]) learn.lr_find(end_lr=100) learn.sched.plot() %time learn.fit(1e-1, 2) A Jupyter Widget [ 0. 1.72594 1.63399 0.41338] [ 1. 1.51599 1.49687 0.45723] CPU times: user 1min 14s, sys: 32.3 s, total: 1min 46s Wall time: 56.5 s %time learn.fit(1e-1, 4, cycle_len=1) A Jupyter Widget [ 0. 1.36734 1.28901 0.53418] [ 1. 1.28854 1.21991 0.56143] [ 2. 1.22854 1.15514 0.58398] [ 3. 1.17904 1.12523 0.59922] CPU times: user 2min 21s, sys: 1min 3s, total: 3min 24s Wall time: 1min 46s class ConvLayer(nn.Module): def __init__(self, ni, nf): super().__init__() self.conv = nn.Conv2d(ni, nf, kernel_size=3, stride=2, padding=1) def forward(self, x): return F.relu(self.conv(x)) class ConvNet2(nn.Module): def __init__(self, layers, c): super().__init__() self.layers = nn.ModuleList([ConvLayer(layers[i], layers[i + 1]) for i in range(len(layers) - 1)]) self.out = nn.Linear(layers[-1], c) def forward(self, x): for l in self.layers: x = l(x) x = F.adaptive_max_pool2d(x, 1) x = x.view(x.size(0), -1) return F.log_softmax(self.out(x), dim=-1) class BnLayer(nn.Module): def __init__(self, ni, nf, stride=2, kernel_size=3): super().__init__() self.conv = nn.Conv2d(ni, nf, kernel_size=kernel_size, stride=stride, bias=False, padding=1) self.a = nn.Parameter(torch.zeros(nf,1,1)) self.m = nn.Parameter(torch.ones(nf,1,1)) def forward(self, x): x = F.relu(self.conv(x)) x_chan = x.transpose(0,1).contiguous().view(x.size(1), -1) if self.training: self.means = x_chan.mean(1)[:,None,None] self.stds = x_chan.std (1)[:,None,None] return (x-self.means) / self.stds *self.m + self.a class ConvBnNet(nn.Module): def __init__(self, layers, c): super().__init__() self.conv1 = nn.Conv2d(3, 10, kernel_size=5, stride=1, padding=2) self.layers = nn.ModuleList([BnLayer(layers[i], layers[i + 1]) for i in range(len(layers) - 1)]) self.out = nn.Linear(layers[-1], c) def forward(self, x): x = self.conv1(x) for l in self.layers: x = l(x) x = F.adaptive_max_pool2d(x, 1) x = x.view(x.size(0), -1) return F.log_softmax(self.out(x), dim=-1) class ConvBnNet2(nn.Module): def __init__(self, layers, c): super().__init__() self.conv1 = nn.Conv2d(3, 10, kernel_size=5, stride=1, padding=2) self.layers = nn.ModuleList([BnLayer(layers[i], layers[i+1]) for i in range(len(layers) - 1)]) self.layers2 = nn.ModuleList([BnLayer(layers[i+1], layers[i + 1], 1) for i in range(len(layers) - 1)]) self.out = nn.Linear(layers[-1], c) def forward(self, x): x = self.conv1(x) for l,l2 in zip(self.layers, self.layers2): x = l(x) x = l2(x) x = F.adaptive_max_pool2d(x, 1) x = x.view(x.size(0), -1) return F.log_softmax(self.out(x), dim=-1) learn = ConvLearner.from_model_data((ConvBnNet2([10, 20, 40, 80, 160], 10), data) %time learn.fit(1e-2, 2) A Jupyter Widget [ 0. 1.53499 1.43782 0.47588] [ 1. 1.28867 1.22616 0.55537] CPU times: user 1min 22s, sys: 34.5 s, total: 1min 56s Wall time: 58.2 s %time learn.fit(1e-2, 2, cycle_len=1) A Jupyter Widget [ 0. 1.10933 1.06439 0.61582] [ 1. 1.04663 0.98608 0.64609] CPU times: user 1min 21s, sys: 32.9 s, total: 1min 54s Wall time: 57.6 s class ResnetLayer(BnLayer): def forward(self, x): return x + super().forward(x) class Resnet(nn.Module): def __init__(self, layers, c): super().__init__() self.conv1 = nn.Conv2d(3, 10, kernel_size=5, stride=1, padding=2) self.layers = nn.ModuleList([BnLayer(layers[i], layers[i+1]) for i in range(len(layers) - 1)]) self.layers2 = nn.ModuleList([ResnetLayer(layers[i+1], layers[i + 1], 1) for i in range(len(layers) - 1)]) self.layers3 = nn.ModuleList([ResnetLayer(layers[i+1], layers[i + 1], 1) for i in range(len(layers) - 1)]) self.out = nn.Linear(layers[-1], c) def forward(self, x): x = self.conv1(x) for l,l2,l3 in zip(self.layers, self.layers2, self.layers3): x = l3(l2(l(x))) x = F.adaptive_max_pool2d(x, 1) x = x.view(x.size(0), -1) return F.log_softmax(self.out(x), dim=-1) learn = ConvLearner.from_model_data(Resnet([10, 20, 40, 80, 160], 10), data) wd=1e-5 %time learn.fit(1e-2, 2, wds=wd) A Jupyter Widget [ 0. 1.58191 1.40258 0.49131] [ 1. 1.33134 1.21739 0.55625] CPU times: user 1min 27s, sys: 34.3 s, total: 2min 1s Wall time: 1min 3s %time learn.fit(1e-2, 3, cycle_len=1, cycle_mult=2, wds=wd) A Jupyter Widget [ 0. 1.11534 1.05117 0.62549] [ 1. 1.06272 0.97874 0.65185] [ 2. 0.92913 0.90472 0.68154] [ 3. 0.97932 0.94404 0.67227] [ 4. 0.88057 0.84372 0.70654] [ 5. 0.77817 0.77815 0.73018] [ 6. 0.73235 0.76302 0.73633] CPU times: user 5min 2s, sys: 1min 59s, total: 7min 1s Wall time: 3min 39s %time learn.fit(1e-2, 8, cycle_len=4, wds=wd) A Jupyter Widget [ 0. 0.8307 0.83635 0.7126 ] [ 1. 0.74295 0.73682 0.74189] [ 2. 0.66492 0.69554 0.75996] [ 3. 0.62392 0.67166 0.7625 ] [ 4. 0.73479 0.80425 0.72861] [ 5. 0.65423 0.68876 0.76318] [ 6. 0.58608 0.64105 0.77783] [ 7. 0.55738 0.62641 0.78721] [ 8. 0.66163 0.74154 0.7501 ] [ 9. 0.59444 0.64253 0.78106] [ 10. 0.53 0.61772 0.79385] [ 11. 0.49747 0.65968 0.77832] [ 12. 0.59463 0.67915 0.77422] [ 13. 0.55023 0.65815 0.78106] [ 14. 0.48959 0.59035 0.80273] [ 15. 0.4459 0.61823 0.79336] [ 16. 0.55848 0.64115 0.78018] [ 17. 0.50268 0.61795 0.79541] [ 18. 0.45084 0.57577 0.80654] [ 19. 0.40726 0.5708 0.80947] [ 20. 0.51177 0.66771 0.78232] [ 21. 0.46516 0.6116 0.79932] [ 22. 0.40966 0.56865 0.81172] [ 23. 0.3852 0.58161 0.80967] [ 24. 0.48268 0.59944 0.79551] [ 25. 0.43282 0.56429 0.81182] [ 26. 0.37634 0.54724 0.81797] [ 27. 0.34953 0.54169 0.82129] [ 28. 0.46053 0.58128 0.80342] [ 29. 0.4041 0.55185 0.82295] [ 30. 0.3599 0.53953 0.82861] [ 31. 0.32937 0.55605 0.82227] CPU times: user 22min 52s, sys: 8min 58s, total: 31min 51s Wall time: 16min 38s class Resnet2(nn.Module): def __init__(self, layers, c, p=0.5): super().__init__() self.conv1 = BnLayer(3, 16, stride=1, kernel_size=7) self.layers = nn.ModuleList([BnLayer(layers[i], layers[i+1]) for i in range(len(layers) - 1)]) self.layers2 = nn.ModuleList([ResnetLayer(layers[i+1], layers[i + 1], 1) for i in range(len(layers) - 1)]) self.layers3 = nn.ModuleList([ResnetLayer(layers[i+1], layers[i + 1], 1) for i in range(len(layers) - 1)]) self.out = nn.Linear(layers[-1], c) self.drop = nn.Dropout(p) def forward(self, x): x = self.conv1(x) for l,l2,l3 in zip(self.layers, self.layers2, self.layers3): x = l3(l2(l(x))) x = F.adaptive_max_pool2d(x, 1) x = x.view(x.size(0), -1) x = self.drop(x) return F.log_softmax(self.out(x), dim=-1) learn = ConvLearner.from_model_data(Resnet2([16, 32, 64, 128, 256], 10, 0.2), data) wd=1e-6 %time learn.fit(1e-2, 2, wds=wd) %time learn.fit(1e-2, 3, cycle_len=1, cycle_mult=2, wds=wd) %time learn.fit(1e-2, 8, cycle_len=4, wds=wd) log_preds,y = learn.TTA() preds = np.mean(np.exp(log_preds),0) metrics.log_loss(y,preds), accuracy(preds,y) (0.44507397166057938, 0.84909999999999997) PATH = "data/dogscats/" sz = 224 arch = resnet34 # <-- Name of the function bs = 64 m = arch(pretrained=True) # Get a model w/ pre-trained weight loaded m ResNet( (conv1): Conv2d (3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (maxpool): MaxPool2d(kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), dilation=(1, 1)) (layer1): Sequential( (0): BasicBlock( (conv1): Conv2d (64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (conv2): Conv2d (64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True) ) (1): BasicBlock( (conv1): Conv2d (64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (conv2): Conv2d (64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True) ) (2): BasicBlock( (conv1): Conv2d (64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (conv2): Conv2d (64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True) ) ) (layer2): Sequential( (0): BasicBlock( (conv1): Conv2d (64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (conv2): Conv2d (128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) (downsample): Sequential( (0): Conv2d (64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) ) ) (1): BasicBlock( (conv1): Conv2d (128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (conv2): Conv2d (128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) ) (2): BasicBlock( (conv1): Conv2d (128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (conv2): Conv2d (128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) ) (3): BasicBlock( (conv1): Conv2d (128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) (relu): ReLU(inplace) (conv2): Conv2d (128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True) ) ) ... (avgpool): AvgPool2d(kernel_size=7, stride=7, padding=0, ceil_mode=False, count_include_pad=True) (fc): Linear(in_features=512, out_features=1000) ) m = nn.Sequential(*children(m)[:-2], nn.Conv2d(512, 2, 3, padding=1), nn.AdaptiveAvgPool2d(1), Flatten(), nn.LogSoftmax()) tfms = tfms_from_model(arch, sz, aug_tfms=transforms_side_on, max_zoom=1.1) data = ImageClassifierData.from_paths(PATH, tfms=tfms, bs=bs) learn = ConvLearner.from_model_data(m, data) learn.freeze_to(-4) learn.fit(0.01, 1) learn.fit(0.01, 1, cycle_len=1) f2=np.dot(np.rollaxis(feat,0,3), py) f2-=f2.min() f2/=f2.max() f2 sf = SaveFeatures(m[-4]) py = m(Variable(x.cuda())) sf.remove() py = np.exp(to_np(py)[0]); py array([ 1., 0.], dtype=float32) feat = np.maximum(0, sf.features[0]) feat.shape class SaveFeatures(): features=None def __init__(self, m): self.hook = m.register_forward_hook(self.hook_fn) def hook_fn(self, module, input, output): self.features = to_np(output) def remove(self): self.hook.remove()

106

null

2018-01-08

2018-01-08 16:45:13

2018-01-10

2018-01-10 20:01:06

13

false

en

2018-07-27

2018-07-27 15:57:25

60

1b9503aff0c

33.856604

56

1

0

My personal notes from fast.ai course. These notes will continue to be updated and improved as I continue to review the course to “really”…

4

Deep Learning 2: Part 1 Lesson 7 My personal notes from fast.ai course. These notes will continue to be updated and improved as I continue to review the course to “really” understand it. Much appreciation to Jeremy and Rachel who gave me this opportunity to learn. Lessons: 1 ・ 2 ・ 3 ・ 4 ・ 5 ・ 6 ・ 7 ・ 8 ・ 9 ・ 10 ・ 11 ・ 12 ・ 13 ・ 14 Lesson 7 The theme of Part 1 is: classification and regression with deep learning identifying and learning best and established practices focus is on classification and regression which is predicting “a thing” (e.g. a number, a small number of labels) Part 2 of the course: focus is on generative modeling which means predicting “lots of things” — for example, creating a sentence as in neural translation, image captioning, or question answering while creating an image such as in style transfer, super-resolution, segmentation and so forth. not as much best practices but a little more speculative from recent papers that may not be fully tested. Review of Char3Model [02:49] Reminder: RNN is not in any way different or unusual or magical — just a standard fully connected network. Standard fully connected network Arrows represent one or more layer operations — generally speaking a linear followed by a non-linear function, in this case matrix multiplications followed by relu or tanh Arrows of the same color represent exactly the same weight matrix being used. One slight difference from previous is that there are inputs coming in at the second and third layers. We tried two approaches — concatenating and adding these inputs to the current activations. By using nn.Linear we get both the weight matrix and the bias vector wrapped up for free for us. To deal with the fact that there is no orange arrow coming in for the first ellipse , we invented an empty matrix Almost identical except for the for loop PyTorch version — nn.RNN will create the loop and keep track of h as it goes along. We are using white section to predict the green character — which seems wasteful as the next section mostly overlaps with the current section. We then tried splitting it into non-overlapping pieces in multi-output model: In this approach, we are throwing away our h activation after processing each section and started a new one. In order to predict the second character using the first one in the next section, it has nothing to go on but a default activation. Let’s not throw away h . Stateful RNN [08:52] One additional line in constructor. self.init_hidden(bs) sets self.h to bunch of zeros. Wrinkle #1 [10:51] — if we were to simply do self.h = h , and we trained on a document that is a million characters long, then the size of unrolled version of the RNN has a million layers (ellipses). One million layer fully connected network is going to be very memory intensive because in order to do a chain rule, we have to multiply one million layers while remembering all one million gradients every batch. To avoid this, we tell it to forget its history from time to time. We can still remember the state (the values in our hidden matrix) without remembering everything about how we got there. Grab the tensor out of Variable h (remember, a tensor itself does not have any concept of history), and create a new Variable out of that. The new variable has the same value but no history of operations, therefore when it tries to back-propagate, it will stop there. forward will process 8 characters, it then back propagate through eight layers, keep track of the values in out hidden state, but it will throw away its history of operations. This is called back-prop through time (bptt). In other words, after the for loop, just throw away the history of operations and start afresh. So we are keeping our hidden state but we are not keeping our hidden state history. Another good reason not to back-propagate through too many layers is that if you have any kind of gradient instability (e.g. gradient explosion or gradient banishing), the more layers you have, the harder the network gets to train (slower and less resilient). On the other hand, the longer bptt means that you are able to explicitly capture a longer memory and more state. Wrinkle #2 [16:00] — how to create mini-batches. We do not want to process one section at a time, but a bunch in parallel at a time. When we started looking at TorchText for the first time, we talked about how it creates these mini-batches. Jeremy said we take a whole long document consisting of the entire works of Nietzsche or all of the IMDB reviews concatenated together, we split this into 64 equal sized chunks (NOT chunks of size 64). For a document that is 64 million characters long, each “chunk” will be 1 million characters. We stack them together and now split them by bptt — 1 mini-bach consists of 64 by bptt matrix. The first character of the second chunk(1,000,001th character) is likely be in the middle of a sentence. But it is okay since it only happens once every million characters. Question: Data augmentation for this kind of dataset? [20:34] There is no known good way. Somebody recently won a Kaggle competition by doing data augmentation which randomly inserted parts of different rows — something like that may be useful here. But there has not been any recent state-of-the-art NLP papers that are doing this kind of data augmentation. Question: How do we choose the size of bptt? [21:36] There are a couple things to think about: the first is that mini-batch matrix has a size of bs (# of chunks) by bptt so your GPU RAM must be able to fit that by your embedding matrix. So if you get CUDA out of memory error, you need reduce one of these. If your training is unstable (e.g. your loss is shooting off to NaN suddenly), then you could try decreasing your bptt because you have less layers to gradient explode through. If it is too slow [22:44], try decreasing your bptt because it will do one of those steps at a time. for loop cannot be parallelized (for the current version). There is a recent thing called QRNN (Quasi-Recurrent Neural Network) which does parallelize it and we hope to cover in part 2. So pick the highest number that satisfies all these. Stateful RNN & TorchText [23:23] When using an existing API which expects data to be certain format, you can either change your data to fit that format or you can write your own dataset sub-class to handle the format that your data is already in. Either is fine, but in this case, we will put our data in the format TorchText already support. Fast.ai wrapper around TorchText already has something where you can have a training path and validation path, and one or more text files in each path containing bunch of text that are concatenated together for your language model. Made a copy of Nietzsche file, pasted into training and validation directory. Then deleted the last 20% of the rows from training set, and deleted everything but the last 20% from the validation set [25:15]. The other benefit of doing it this way is that it seems like it is more realistic to have a validation set that was not a random shuffled set of rows of text, but it was totally separate part of the corpus. When you are doing a language model, you do not really need separate files. You can have multiple files but they just get concatenated together anyway. In TorchText, we make this thing called Field and initially Field is just a description of how to go about pre-processing the text. lower — we told it to lowercase the text tokenize — Last time, we used a function that splits on whitespace that gave us a word model. This time, we want a character model, so use list function to tokenize strings. Remember, in Python, list('abc') will return ['a', 'b', 'c'] . bs : batch size, bptt : we renamed cs , n_fac : size of embedding, n_hidden : size of our hidden state We do not have a separate test set, so we’ll just use validation set for testing TorchText randomize the length of bptt a little bit each time. It does not always give us exactly 8 characters; 5% of the time, it will cut it in half and add on a small standard deviation to make it slightly bigger or smaller than 8. We cannot shuffle the data since it needs to be contiguous, so this is a way to introduce some randomness. Question [31:46]: Does the size remain constant per mini-batch? Yes, we need to do matrix multiplication with h weight matrix, so mini-batch size must remain constant. But sequence length can change no problem. len(md.trn_dl) : length of data loader (i.e. how many mini-batches), md.nt : number of tokens (i.e. how many unique things are in the vocabulary) Once you run LanguageModelData.from_text_files , TEXT will contain an extra attribute called vocab. TEXT.vocab.itos list of unique items in the vocabulary, and TEXT.vocab.stoi is a reverse mapping from each item to number. Wrinkle #3 [33:51]: Jeremy lied to us when he said that mini-batch size remains constant. It is very likely that the last mini-batch is shorter than the rest unless the dataset is exactly divisible by bptt times bs . That is why we check whether self.h ‘s second dimension is the same as bs of the input. If it is not the same, set it back to zero with the input’s bs . This happens at the end of the epoch and the beginning of the epoch (setting back to the full batch size). Wrinkle #4 [35:44]: The last wrinkle is something that slightly sucks about PyTorch and maybe somebody can be nice enough to try and fix it with a PR. Loss functions are not happy receiving a rank 3 tensor (i.e. three dimensional array). There is no particular reason they ought to not be happy receiving a rank 3 tensor (sequence length by batch size by results — so you can just calculate loss for each of the two initial axis). Works for rank 2 or 4, but not 3. .view will reshape rank 3 tensor into rank 2 of -1 (however big as necessary) by vocab_size. TorchText automatically changes the target to be flattened out, so we do not need to do that for actual values (when we looked at a mini-batch in lesson 4, we noticed that it was flattened. Jeremy said we will learn about why later, so later is now.) PyTorch (as of 0.3), log_softmax requires us to specify which axis we want to do the softmax over (i.e. which axis we want to sum to one). In this case we want to do it over the last axis dim = -1. Let’s gain more insight by unpacking RNN [42:48] We remove the use of nn.RNN and replace it with nn.RNNCell . PyTorch source code looks like the following. You should be able to read and understand (Note: they do not concatenate the input and the hidden state, but they sum them together — which was our first approach): Question about tanh [44:06]: As we have seen last week, tanh is forcing the value to be between -1 and 1. Since we are multiplying by this weight matrix again and again, we would worry that relu (since it is unbounded) might have more gradient explosion problem. Having said that, you can specify RNNCell to use different nonlineality whose default is tanh and ask it to use relu if you wanted to. for loop is back and append the result of linear function to a list — which in end gets stacked up together. fast.ai library actually does exactly this in order to use regularization approaches that are not supported by PyTorch. Gated Recurrent Unit (GRU) [46:44] In practice, nobody really uses RNNCell since even with tanh , gradient explosions are still a problem and we need use low learning rate and small bptt to get them to train. So what we do is to replace RNNCell with something like GRUCell . http://www.wildml.com/2015/10/recurrent-neural-network-tutorial-part-4-implementing-a-grulstm-rnn-with-python-and-theano/ Normally, the input gets multiplied by a weight matrix to create new activations h and get added to the existing activations straight away. That is not wha happens here. Input goes into h˜ and it doesn’t just get added to the previous activations, but the previous activation gets multiplied by r (reset gate) which has a value of 0 or 1. r is calculated as below — matrix multiplication of some weight matrix and the concatenation of our previous hidden state and new input. In other words, this is a little one hidden layer neural net. It gets put through the sigmoid function as well. This mini neural net learns to determine how much of the hidden states to remember (maybe forget it all when it sees a full-stop character — beginning of a new sentence). z gate (update gate) determines what degree to use h˜ (the new input version of hidden states) and what degree to leave the hidden state the same as before. http://colah.github.io/posts/2015-08-Understanding-LSTMs/ Linear interpolation Above is what GRUCell code looks like, and our new model that utilize this is below: As a result, we can lower the loss down to 1.36 (RNNCell one was 1.54). In practice, GRU and LSTM are what people uses. Putting it all together: Long Short-Term Memory [54:09] LSTM has one more piece of state in it called “cell state” (not just hidden state), so if you do use a LSTM, you have to return a tuple of matrices in init_hidden (exactly the same size as hidden state): The code is identical to GRU one. The one thing that was added was dropout which does dropout after each time step and doubled the hidden layer — in a hope that it will be able to learn more and be resilient as it does so. Callbacks (specifically SGDR) without Learner class [55:23] After creating a standard PyTorch model, we usually do something like opt = optim.Adam(m.parameters(), 1e-3). Instead, we will use fast.ai LayerOptimizer which takes an optimizer optim.Adam , our model m , learning rate 1e-2 , and optionally weight decay 1e-5 . A key reason LayerOptimizer exists is to do differential learning rates and differential weight decay. The reason we need to use it is that all of the mechanics inside fast.ai assumes that you have one of these. If you want to use callbacks or SGDR in code you are not using the Learner class, you need to use this. lo.opt returns the optimizer. When we call fit, we can now pass the LayerOptimizer and also callbacks. Here, we use cosine annealing callback — which requires a LayerOptimizer object. It does cosine annealing by changing learning rate in side the lo object. Concept: Create a cosine annealing callback which is going to update the learning rates in the layer optimizer lo . The length of an epoch is equal to len(md.trn_dl) — how many mini-batches are there in an epoch is the length of the data loader. Since it is doing cosine annealing, it needs to know how often to reset. You can pass in cycle_mult in usual way. We can even save our model automatically just like we did with cycle_save_name in Learner.fit. We can do callback at a start of a training, epoch or a batch, or at the end of a training, an epoch, or a batch. It has been used for CosAnneal (SGDR), and decoupled weight decay (AdamW), loss-over-time graph, etc. Testing [59:55] In lesson 6, when we were testing CharRnn model, we noticed that it repeated itself over and over. torch.multinomial used in this new version deals with this problem. p[-1] to get the final output (the triangle), exp to convert log probability to probability. We then use torch.multinomial function which will give us a sample using the given probabilities. If probability is [0, 1, 0, 0] and ask it to give us a sample, it will always return the second item. If it was [0.5, 0, 0.5], it will give the first item 50% of the time, and second item . 50% of the time (review of multinomial distribution) To play around with training character based language models like this, try running get_next_n at different levels of loss to get a sense of what it looks like. The example above is at 1.25, but at 1.3, it looks like a total junk. When you are playing around with NLP, particularly generative model like this, and the results are kind of okay but not great, do not be disheartened because that means you are actually very VERY nearly there! Back to computer vision: CIFAR 10 [1:01:58] CIFAR 10 is an old and well known dataset in academia — well before ImageNet, there was CIFAR 10. It is small both in terms of number of images and size of images which makes it interesting and challenging. You will likely be working with thousands of images rather than one and a half million images. Also a lot of the things we are looking at like in medical imaging, we are looking at a specific area where there is a lung nodule, you are probably looking at 32 by 32 pixels at most. It also runs quickly, so it is much better to test our your algorithms. As Ali Rahini mentioned in NIPS 2017, Jeremy has the concern that many people are not doing carefully tuned and throught-about experiments in deep learning, but instead, they throw lots of GPUs and TPUs or lots of data and consider that a day. It is important to test many versions of your algorithm on dataset like CIFAR 10 rather than ImageNet that takes weeks. MNIST is also good for studies and experiments even though people tend to complain about it. CIFAR 10 data in image format is available here classes — image labels stats —When we use pre-trained models, you can call tfms_from_model which creates the necessary transforms to convert our data set into a normalized dataset based on the means and standard deviations of each channel in the original model that was trained in. Since we are training a model from scratch, we ned to tell it the mean and standard deviation of our data to normalize it. Make sure you can calculate the mean and the standard deviation for each channel. tfms — For CIFAR 10 data augmentation, people typically do horizontal flip and black padding around the edge and randomly select 32 by 32 area within the padded image. From this notebook by our student Kerem Turgutlu: nn.ModuleList — whenever you create a list of layers in PyTorch, you have to wrap it in ModuleList to register these as attributes. Now we step up one level of API higher — rather than calling fit function, we create a learn object from a custom model. ConfLearner.from_model_data takes standard PyTorch model and model data object. With a simple one hidden layer model with 122,880 parameters, we achieved 46.9% accuracy. Let’s improve this and gradually build up to a basic ResNet architecture. CNN [01:12:30] Let’s replace a fully connected model with a convolutional model. Fully connected layer is simply doing a dot product. That is why the weight matrix is big (3072 input * 40 = 122880). We are not using the parameters very efficiently because every single pixel in the input has a different weight. What we want to do is a group of 3 by 3 pixels that have particular patterns to them (i.e. convolution). We will use a filter with three by three kernel. When there are multiple filters, the output will have additional dimension. Replace nn.Linear with nn.Conv2d First two parameters are exactly the same as nn.Linear — the number of features coming in, and the number of features coming out kernel_size=3 , the size of the filter stride=2 will use every other 3 by 3 area which will halve the output resolution in each dimension (i.e. it has the same effect as 2 by 2 max-pooling) ConvNet([3, 20, 40, 80], 10) — It start with 3 RGB channels, 20, 40, 80 features, then 10 classes to predict. AdaptiveMaxPool2d — This followed by a linear layer is how you get from 3 by 3 down to a prediction of one of 10 classes and is now a standard for state-of-the-art algorithms. The very last layer, we do a special kind of max-pooling for which you specify the output activation resolution rather than how big of an area to poll. In other words, here we do 3 by 3 max-pool which is equivalent of 1 by 1 adaptive max-pool. x = x.view(x.size(0), -1) — x has a shape of # of the features by 1 by 1, so it will remove the last two layers. This model is called “fully convolutional network” — where every layer is convolutional except for the very last. The default final learning rate lr_find tries is 10. If the loss is still getting better at that point, you can overwrite by specifying end_lr . It flattened out around 60% accuracy. Considering it uses about 30,000 parameters (compared to 47% with 122k parameters) Time per epoch is about the same since their architectures are both simple and most of time is spent doing memory transfer. Refactored [01:21:57] Simplify forward function by creating ConvLayer (our first custom layer!). In PyTorch, layer definition and neural network definitions are identical. Anytime you have a layer, you can use it as a neural net, when you have a neural net, you can use it as a layer. padding=1 — When you do convolution the image shrink by 1 pixel on each side. So it does not go from 32 by 32 to 16 by 16 but actually 15 by 15. padding will add a border so we can keep the edge pixel information. It is not as big of a deal for a big image, but when it’s down to 4 by 4, you really don’t want to throw away a whole piece. Another difference from the last model is that nn.AdaptiveMaxPool2d does not have any state (i.e. no weights). So we can just call it as a function F.adaptive_max_pool2d . BatchNorm [1:25:10] The last model, when we tried to add more layers, we had trouble training. The reason we had trouble training was that if we used larger learning rates, it would go off to NaN and if we used smaller learning rate, it would take forever and doesn’t have a chance to explore properly — so it was not resilient. To make it resilient, we will use something called batch normalization. BatchNorm came out about two years ago and it has been quite transformative since it suddenly makes it really easy to train deeper networks. We can simply use nn.BatchNorm but to learn about it, we will write it from scratch. It is unlikely that the weight matrices on average are not going to cause your activations to keep getting smaller and smaller or keep getting bigger and bigger. It is important to keep them at reasonable scale. So we start things off with zero-mean standard deviation one by normalizing the input. What we really want to do is to do this for all layers, not just the inputs. Calculate the mean of each channel or each filter and standard deviation of each channel or each filter. Then subtract the means and divide by the standard deviations. We no longer need to normalize our input because it is normalizing it per channel or for later layers it is normalizing per filter. Turns out this is not enough since SGD is bloody-minded [01:29:20]. If SGD decided that it wants matrix to be bigger/smaller overall, doing (x=self.means) / self.stds is not enough because SGD will undo it and try to do it again in the next mini-batch. So we will add two parameters: a — adder (initial value zeros) and m — multiplier (initial value ones) for each channel. Parameter tells PyTorch that it is allowed to learn these as weights. Why does this work? If it wants to scale the layer up, it does not have to scale up every single value in the matrix. It can just scale up this single trio of numbers self.m , if it wants to shift it all up or down a bit, it does not have to shift the entire weight matrix, they can just shift this trio of numbers self.a. Intuition: We are normalizing the data and then we are saying you can then shift it and scale it using far fewer parameters than would have been necessary if it were to actually shift and scale the entire set of convolutional filters. In practice, it allows us to increase our learning rates, it increase the resilience of training, and it allows us to add more layers and still train effectively. The other thing batch norm does is that it regularizes, in other words, you can often decrease or remove dropout or weight decay. The reason why is each mini-batch is going to have a different mean and a different standard deviation to the previous mini-batch. So they keep changing and it is changing the meaning of the filters in a subtle way acting as a noise (i.e. regularization). In real version, it does not use this batch’s mean and standard deviation but takes an exponentially weighted moving average standard deviation and mean. if self.training — this is important because when you are going through the validation set, you do not want to be changing the meaning of the model. There are some types of layer that are actually sensitive to what the mode of the network is whether it is in training mode or evaluation/test mode. There was a bug when we implemented mini net for MovieLens that dropout was applied during the validation — which was fixed. In PyTorch, there are two such layer: dropout and batch norm. nn.Dropout already does the check. [01:37:01] The key difference in fast.ai which no other library does is that these means and standard deviations get updated in training mode in every other library as soon as you basically say I am training, regardless of whether that layer is set to trainable or not. With a pre-trained network, that is a terrible idea. If you have a pre-trained network for specific values of those means and standard deviations in batch norm, if you change them, it changes the meaning of those pre-trained layers. In fast.ai, always by default, it will not touch those means and standard deviations if your layer is frozen. As soon as you un-freeze it, it will start updating them unless you set learn.bn_freeze=True. In practice, this often seems to work a lot better for pre-trained models particularly if you are working with data that is quite similar to what the pre-trained model was trained with. Where do you put batch-norm layer? We will talk more in a moment, but for now, after relu Ablation Study [01:39:41] It is something where you try turning on and off different pieces of your model to see which bits make which impacts, and one of the things that wasn’t done in the original batch norm paper was any kind of effective ablation. And one of the things therefore that was missing was this question which was just asked — where to put the batch norm. That oversight caused a lot of problems because it turned out the original paper did not actually put it in the best spot. Other people since then have now figured that out and when Jeremy show people code where it is actually in the spot that is better, people say his batch norm is in the wrong spot. Try and always use batch norm on every layer if you can Don’t stop normalizing your data so that people using your data will know how you normalized your data. Other libraries might not deal with batch norm for pre-trained models correctly, so when people start re-training, it might cause problems. Rest of the code is similar — Using BnLayer instead of ConvLayer A single convolutional layer was added at the start trying to get closer to the modern approaches. It has a bigger kernel size and a stride of 1. The basic idea is that we want the first layer to have a richer input. It does convolution using the 5 by 5 area which allows it to try and find more interesting richer features in that 5 by 5 area, then spit out bigger output (in this case, it’s 10 by 5 by 5 filters). Typically it is 5 by 5 or 7 by 7, or even 11 by 11 convolution with quite a few filters coming out (e.g. 32 filters). Since padding = kernel_size — 1 / 2 and stride=1 , the input size is the same as the output size — just more filters. It is a good way of trying to create a richer starting point. Deep BatchNorm [01:50:52] Let’s increase the depth of the model. We cannot just add more of stride 2 layers since it halves the size of the image each time. Instead, after each stride 2 layer, we insert a stride 1 layer. The accuracy remained the same as before. This is now 12 layers deep, and it is too deep even for batch norm to handle. It is possible to train 12 layer deep conv net but it starts to get difficult. And it does not seem to be helping much if at all. ResNet [01:52:43] ResnetLayer inherit from BnLayer and override forward. Then add bunch of layers and make it 3 times deeper, ad it still trains beautifully just because of x + super().forward(x) . ResNet block [01:53:18] return x + super().forward(x) y = x + f(x) Where x is prediction from the previous layer, y is prediction from the current layer.Shuffle around the formula and we get:formula shuffle f(x) = y − x The difference y − x is residual. The residual is the error in terms of what we have calculated so far. What this is saying is that try to find a set of convolutional weights that attempts to fill in the amount we were off by. So in other words, we have an input, and we have a function which tries to predict the error (i.e. how much we are off by). Then we add a prediction of how much we were wrong by to the input, then add another prediction of how much we were wrong by that time, and repeat that layer after layer — zooming into the correct answer. This is based on a theory called boosting. The full ResNet does two convolutions before it gets added back to the original input (we did just one here). In every block x = l3(l2(l(x))) , one of the layers is not a ResnetLayer but a standard convolution with stride=2 — this is called a “bottleneck layer”. ResNet does not convolutional layer but a different form of bottleneck block which we will cover in Part 2. ResNet 2 [01:59:33] Here, we increased the size of features and added dropout. 85% was a state-of-the-art back in 2012 or 2013 for CIFAR 10. Nowadays, it is up to 97% so there is a room for improvement but all based on these tecniques: Better approaches to data augmentation Better approaches to regularization Some tweaks on ResNet Question [02:01:07]: Can we apply “training on the residual” approach for non-image problem? Yes! But it has been ignored everywhere else. In NLP, “transformer architecture” recently appeared and was shown to be the state of the art for translation, and it has a simple ResNet structure in it. This general approach is called “skip connection” (i.e. the idea of skipping over a layer) and appears a lot in computer vision, but nobody else much seems to be using it even through there is nothing computer vision specific about it. Good opportunity! Dogs vs. Cats [02:02:03] Going back dogs and cats. We will create resnet34 (if you are interested in what the trailing number means, see here — just different parameters). Our ResNet model had Relu → BatchNorm. TorchVision does BatchNorm →Relu. There are three different versions of ResNet floating around, and the best one is PreAct (https://arxiv.org/pdf/1603.05027.pdf). Currently, the final layer has a thousands features because ImageNet has 1000 features, so we need to get rid of it. When you use fast.ai’s ConvLearner , it deletes the last two layers for you. fast.ai replaces AvgPool2d with Adaptive Average Pooling and Adaptive Max Pooling and concatenate the two together. For this exercise, we will do a simple version. Remove the last two layers Add a convolution which just has 2 outputs. Do average pooling then softmax There is no linear layer at the end. This is a different way of producing just two numbers — which allows us to do CAM! ConvLearner.from_model is what we learned about earlier — allows us to create a Learner object with custom model. Then freeze the layer except the ones we just added. Class Activation Maps (CAM) [02:08:55] We pick a specific image, and use a technique called CAM where we take a model and we ask it which parts of the image turned out to be important. How did it do this? Let’s work backwards. The way it did it was by producing this matrix: Big numbers correspond to the cat. So what is this matrix? This matrix simply equals to the value of feature matrix feat times py vector: py vector is the predictions that says “I am 100% confident it’s a cat.” feat is the values (2×7×7) coming out of the final convolutional layer (the Conv2d layer we added). If we multiply feat by py , we get all of the first channel and none of the second channel. Therefore, it is going to return the value of the last convolutional layers for the section which lines up with being a cat. In other words, if we multiply feat by [0, 1] , it will line up with being a dog. Put it in another way, in the model, the only thing that happened after the convolutional layer was an average pooling layer. The average pooling layer took took the 7 by 7 grid and averaged out how much each part is “cat-like”. We then took the “cattyness” matrix, resized it to be the same size as the original cat image, and overlaid it on top, then you get the heat map. The way you can use this technique at home is when you have a large image, you can calculate this matrix on a quick small little convolutional net zoom into the area that has the highest value re-run it just on that part We skipped this over quickly as we ran out of time, but we will learn more about these kind of approaches in Part 2. “Hook” is the mechanism that lets us ask the model to return the matrix. register_forward_hook asks PyTorch that every time it calculates a layer it runs the function given — sort of like a callback that happens every time it calculates a layer. In the following case, it saves the value of the particular layer we were interested in: Questions to Jeremy [02:14:27]: “Your journey into Deep Learning” and “How to keep up with important research for practitioners” “If you intend to come to Part 2, you are expected to master all the techniques er have learned in Part 1”. Here are something you can do: Watch each of the video at least 3 times. Make sure you can re-create the notebooks without watching the videos — maybe do so with different datasets to make it more interesting. Keep an eye on the forum for recent papers, recent advances. Be tenacious and keep working at it! Lessons: 1 ・ 2 ・ 3 ・ 4 ・ 5 ・ 6 ・ 7 ・ 8 ・ 9 ・ 10 ・ 11 ・ 12 ・ 13 ・ 14

Deep Learning 2: Part 1 Lesson 7

286

deep-learning-2-part-1-lesson-7-1b9503aff0c

2018-07-27

2018-07-27 15:57:25

https://medium.com/s/story/deep-learning-2-part-1-lesson-7-1b9503aff0c

false

8,601

null

Machine Learning

machine-learning

Machine Learning

51,320

Hiromi Suenaga

null

eab3a535185e

hiromi_suenaga

1,462

10

20,181,104

null

0

null

0

null

2017-11-02

2017-11-02 13:10:19

2017-11-02

2017-11-02 13:22:27

3

true

zh-Hant

2017-11-02

2017-11-02 19:19:31

6

1b991d8541c9

0.595283

19

0

「人工智慧並不恨你，也不愛你，但你是原子組成的，而它能利用這原子可以用作與做別的事情上。」

5

我是一個只會製造萬字夾的人工智能，卻不小心毀滅了世界全個遊戲的惟一圖片：一盒萬字夾「人工智慧並不恨你，也不愛你，但你是原子組成的，而它能利用這原子可以用作與做別的事情上。」上週開始在外國大爆發的瀏覽器遊戲：Paperclips（萬字夾）。《The Verge》以「現時最上癮遊戲」描述之、《Wired》則以〈世界末日原因不是大爆炸，而是一個萬字夾〉《Forbes》有博客指自己進入了它的黑洞。玩家的角色，是一個只懂製作萬字夾的人工智能。全個遊戲都在一個類似計數機的界面上進行。購買原材料、增加機器、做市場推廣等，全部都是你選擇如何分佈資源、為貨品定價、同時觀察市場需求來調節。因為你是一套AI，所以製作不會停止，不會休季。 Igor Kirisyuk@flickr — Attribution-NonCommercial-NoDerivs 2.0 Generic (CC BY-NC-ND 2.0) 到你賣出數以百萬計的萬字夾後，劇情會有轉變。但我個人認為最好玩都是前期階段。 2003年，瑞典哲學家Nick Bostrom首次提出「萬字夾製造機（Paperclip maximizer）」的思想實驗。這實驗的主角是一套擁有簡單函數的人工智能，它的生存只為了一個目標：盡量製作萬字夾。但這個看似對人類無害的目標，卻很可能對人類造成滅絕性的威脅。五年後，美國人工智能研究者Eliezer Yudkowsky提醒我們「人工智慧並不恨你，也不愛你，但你是原子組成的，而它能利用這原子可以用作與做別的事情上。」 ……不過，再看資料，只會減少遊樂時間，玩了兩晚，我發現剩貨不一定是壞事，流動資金再多也不一定好。不理會這個小遊戲帶出「人工科學會否滅世」等哲學討論，這可能是一個做生意的大訓練。遊戲連結：http://www.decisionproblem.com/paperclips/ M Nottage@flickr — Attribution-NonCommercial-NoDerivs 2.0 Generic (CC BY-NC-ND 2.0)

我是一個只會製造萬字夾的人工智能，卻不小心毀滅了世界

67

我是一個只會製造萬字夾的人工智能-但我卻不小心毀滅了世界-1b991d8541c9

2018-06-20

2018-06-20 21:28:31

https://medium.com/s/story/我是一個只會製造萬字夾的人工智能-但我卻不小心毀滅了世界-1b991d8541c9

false

12

null

Games

games

Games

23,849

陳裕匡 Nickleby Chan Yu Hong

Do researches a lot, on Netflix. Always dehydrated. Used to have a man bun. Editors, Social Media Manager from Hong Kong.

59314110cd3f

NicklebyChan

330

264

20,181,104

null

0

null

0

fb4c6ed275bd

2017-11-06

2017-11-06 16:52:03

2017-11-07

2017-11-07 13:31:01

1

false

en

2017-11-07

2017-11-07 13:31:01

1

1b9959ec74e1

0.932075

2

0

Lassonde PhD Student and co-founder of York University start-up Deepnify, Nima Shahbazi, is exploring how AI and machine learning will…

5

How AI can reduce food waste Lassonde PhD Student and co-founder of York University start-up Deepnify, Nima Shahbazi, is exploring how AI and machine learning will eliminate food waste in grocery stores. “ Predicting enough supply to meet demand is very difficult, so grocery stores stock up on more food than necessary.” See transcript below AI is being used to reduce food waste. Much of that waste is from grocery stores tossing unsold food. Predicting enough supply to meet demand is very difficult, so grocery stores stock up on more food than necessary. Toronto start-up Deepnify is using machine learning to better forecast the amount of food grocery stores should stock each day. Nima Shahbazi: Machine learning will eventually let grocery stores operate without waste, saving more than ten billion dollars worth of waste in Canada. Nima Shahbazi: As a result, all food companies will need a zero waste supply chain in order to price competitively. When grocery stores only order as much food as they need, they can significantly reduce their food waste. Deepnify is testing its AI software with two Canadian grocery chains. The early results will show promise.

How AI can reduce food waste

4

how-ai-can-reduce-food-waste-1b9959ec74e1

2018-04-14

2018-04-14 15:27:21

https://medium.com/s/story/how-ai-can-reduce-food-waste-1b9959ec74e1

false

194

We've invented a new engineering school for people who want to change the world.

null

lassondeschool

null

LassondeSchool

ask@lassonde.yorku.ca

lassondeschool

ENGINEERING,DESIGN THINKING,STARTUP,ENTREPRENEURSHIP,EDUCATION

lassondeschool

Food

food

Food

110,550

Lassonde School of Engineering

We’ve invented a new engineering school for people who want to change the world. Lassonde School of Engineering at York University.

60ac9079fa2e

LassondeSchool

233

308

20,181,104

null

0

null

0

f439fc5aea86

2018-05-24

2018-05-24 13:47:53

2018-05-25

2018-05-25 10:05:16

7

false

th

2018-05-25

2018-05-25 10:05:16

9

1b99fe798759

2.793396

3

0

สวัสดีครับ ในบทความชุดนี้เราจะมาทำความรู้จักกับ architecture ใหม่ล่าสุดในแวดวง deep learning และ natural language processing (NLP)…

4

มารู้จัก Transformer กันเถอะ (ตอนที่ 1) สวัสดีครับ ในบทความชุดนี้เราจะมาทำความรู้จักกับ architecture ใหม่ล่าสุดในแวดวง deep learning และ natural language processing (NLP) ที่ชื่อว่า Transformer กันครับ เนื่องจากความยาวของเนื้อหา จึงจะขอแบ่งเป็นออกเป็น 4 ตอนย่อยๆ คือ ตอนที่ 1 จะเป็นการเกริ่นนำถึงความสำคัญของเปเปอร์ Attention Is All You Need และความเป็นมาก่อนที่จะมีเปเปอร์นี้ ตอนที่ 2 จะกล่าวถึง Attention Module ซึ่งถือเป็นหัวใจหลักของ Transformer ครับ ตอนที่ 3 จะมาดู architecture ของ Transformer กัน ตอนที่ 4 แสดงถึงข้อดีและจุดเด่นของ Self-Attention ถ้าพร้อมแล้วเรามาเริ่มกันเลยครับ เกริ่นนำ เมื่อปีที่แล้วนี่เอง (ค.ศ. 2017) มีเหล่านักวิจัย 8 คน ซึ่งส่วนใหญ่มาจากบริษัท Google ได้ร่วมกันตีพิมพ์เปเปอร์ที่มีชื่อว่า Attention Is All You Need ซึ่งสั่นสะเทือนวงการ deep learning และ NLP อย่างสูง โดยสิ่งที่ทำให้เปเปอร์นี้ทรงอิทธิพลมากเห็นจะเป็นด้วยเหตุผลสามประการดังนี้ครับ ในโลกของ deep learning ปัจจุบัน โครงสร้าง network ที่เป็นมหาอำนาจในโลกนี้มีอยู่แค่สองเจ้าผู้ยิ่งใหญ่เท่านั้น คือไม่ RNN ก็ CNN นี่แหละครับ สำหรับ RNN ส่วนใหญ่จะใช้ในงานด้าน language และ CNN จะใช้ในงานด้าน vision แต่ทั้งนี้ทั้งนั้น CNN ก็เอามาใช้ในงานด้าน language ส่วน RNN ก็สามารถนำไปใช้ในงานด้าน vision ได้เช่นกัน ซึ่งเปเปอร์นี้เล็งเห็นทั้งจุดอ่อนของทั้ง RNN และ CNN จึงปฏิเสธที่จะใช้ทั้งสองอย่างนี้ในการทำ NLP เสียเลย แล้วหันมาใช้สิ่งที่เรียกว่า Attention เพียงอย่างเดียว โดยไม่ใช่แค่ Attention แบบวานิลลาธรรมดา แต่จะขยายขอบเขตและความสามารถของ Attention ขึ้นไป แล้วเรียก architecture ใหม่นี้ว่า Transformer ซึ่งไม่ได้หมายถึงหม้อแปลงไฟฟ้าหรือยานพาหนะที่แปลงรูปร่างเป็นหุ่นยนต์ได้แต่ประการใดนะครับ หัวใจหลักของ Transformer คือกระบวนการที่เรียกว่า Self-Attention โดยกระบวนการนี้นอกจากจะเป็นสิ่งที่ทดแทน RNN และ CNN ได้แล้ว ยังแสดงถึงความข้องเกี่ยวกันของคำต่างๆ ในข้อความ ทำให้สามารถแก้ปัญหา coreference resolution ไปได้กลายๆ ซึ่งการแก้ปัญหานี้มีความสำคัญอย่างมากต่องาน NLP หลายประเภท เช่น machine translation ครับ Transformer ให้ผลการทดลองในชุดข้อมูลมาตรฐานที่เหนือกว่าวิธีการอื่นอย่างชัดเจน ตัวอย่างเช่นในงาน machine translation ซึ่งใช้ชุดข้อมูล newstest2014 ทำการแปลจากภาษาอังกฤษเป็นภาษาเยอรมัน และภาษาอังกฤษเป็นภาษาฝรั่งเศส แล้ววัดผลโดยใช้ค่า BLEU ผลการเปรียบเทียบจะเป็นดังตารางด้านล่างนี้ ซึ่งนำมาจากเปเปอร์โดยตรง จะเห็นว่าการแปลไปภาษาฝรั่งเศสให้ค่าดีกว่าการแปลไปภาษาเยอรมัน นั่นเป็นเพราะว่าข้อมูลของภาษาฝรั่งเศสมีเยอะกว่าหลายเท่า และภาษาเยอรมันจะมี morphology ที่ซับซ้อนกว่านั่นเองครับ ซึ่ง Transformer สามารถอัพสกอร์ของภาษาเยอรมันได้มากกว่าวิธีอื่นถึงขนาดนี้โดยที่ใช้การคำนวณน้อยกว่าก็นับว่าเทพจริงๆ ครับ เปรียบเทียบผลของ Transfomer กับวิธีอื่นๆ [Attention Is All You Need, Table 2] สิ่งนึงที่สะกิดใจผู้เขียนคือเปเปอร์นี้มีผู้ร่วมทำกันหลายคน และแต่ละคนก็ทำหน้าที่ในแต่ละส่วนอย่างมากมาย เพราะฉะนั้นรายละเอียดของ Transformer จึงมีค่อนข้างเยอะ ซึ่งในบทความนี้ผู้เขียนจะพยายามอธิบายภาพรวมและใจความของ Transformer นี้ให้ได้มากที่สุดครับ ความมีส่วนร่วมของผู้ทำเปเปอร์ [Attention Is All You Need, authors] ความเป็นมา ประเภทของงาน machine learning อาจจะแบ่งตามลักษณะของ input ที่เข้ามาและ output ที่ได้ ซึ่งท่านผู้อ่านอาจจะคุ้นเคยกับงานประเภท classification หรือ regression กันดี แต่ในที่นี้เราจะสนใจงานที่มีข้อมูลเข้ามาเป็น sequence นึง แล้วให้ตอบกลับไปเป็นอีก sequence นึงได้ โดยจะขอเรียกว่า sequence-to-sequence learning นะครับ ซึ่งการทำ machine translation ก็ถือว่าเป็น sequence-to-sequence learning แบบนึง และนับว่าเป็นสิ่งที่ผู้คนให้ความสำคัญมากตั้งแต่ยุคสมัยแรกๆ ของ AI มาจนถึงปัจจุบันเลย ตัวอย่างการทำ machine translation ที่แปลจากภาษาจีนเป็นภาษาไทย ปัจจุบันวิธีที่เป็นมาตรฐานสำหรับทำงาน sequence-to-sequence learning คือ Sequence-to-Sequence Model (seq2seq) หรือเรียกอีกชื่อว่า RNN Encoder–Decoder ซึ่งโมเดลนี้จะแบ่งเป็นสองส่วน เรียกว่า encoder กับ decoder โดยส่วน encoder จะรับ input เข้ามาทีละหน่วยผ่านทาง RNN และเก็บสะสม information ที่จำเป็นไว้ จากนั้นจะผ่าน information นี้ไปยังส่วน decoder ซึ่งก็จะเป็น RNN อีกตัวนึงที่ให้ output ออกมาทีละหน่วย โดยดูจาก information ที่ได้รับมา และ output ตัวก่อนหน้า ซึ่งแผนภาพคร่าวๆ ของ seq2seq จะแสดงได้ดังนี้ครับ ในที่นี้ seq2seq รับ input คือ “A B C” และให้ output เป็น “W X Y Z” ออกมา [Sutskever et al., Figure 1] อย่างไรก็ดี โมเดล seq2seq จะมีปัญหาคอขวดเกิดขึ้น นั่นคือการส่ง information เป็นทอดๆ ตามสายยาวแบบนี้ อาจจะมี information ที่จำเป็นบางอย่างสูญหายไประหว่างทางได้ ยกตัวอย่างเช่น จากรูปข้างบน สมมติว่าเป็นงาน machine translation ที่แปลจากภาษาไทยไปเป็นภาษาอังกฤษ โดย “A B C” คือคำว่า “ฉัน เลี้ยง แมว” และ “W X Y Z” เป็นคำว่า “I have a cat” จะเห็นว่า output คำว่า “cat” จะขึ้นอยู่กับ input คำว่า “แมว” โดยตรง แต่ข้อมูลจาก C กว่าจะส่งมาถึง Z ต้องผ่านตัวกลางหลายทอด และอาจจะสูญหายได้ โดยเฉพาะอย่างยิ่งถ้าเป็นประโยคยาวๆ จึงมีความคิดว่า จะดีกว่าไหมถ้าเราให้กระบวนการสร้าง output สามารถโฟกัสไปที่ input ส่วนใดส่วนหนึ่งได้โดยตรง และนี่คือที่มาของ Attention นั่นเองครับ การทำ Attention สำหรับ seq2seq [CS224n, Lecture 11] การทำ Attention สำหรับ seq2seq สามารถแสดงได้ดังรูปด้านบนนี้ โดยเมื่อต้องการจะคำนวณ output ที่ตำแหน่งนึง ก็จะนำ vector ของ decoder (q) ณ ตำแหน่งนั้น มาใช้หา attention score กับ vector ของ encoder (p) ในทุกตำแหน่ง ซึ่งถ้า score ที่ encoder ตำแหน่งไหนสูง หมายความว่าเราจะให้ความสำคัญ หรือใส่ใจกับตำแหน่งนั้นมาก การคำนวณค่านี้ก็ทำได้หลายวิธี โดยวิธีที่ง่ายที่สุดก็คือการทำ dot product กันตรงๆ ระหว่าง p กับ q เลย ซึ่งหมายความว่าเราจะใส่ใจกับตำแหน่งที่มีค่า p ใกล้เคียงกับค่า q และเมื่อได้ค่า score ออกมาแล้ว ก็จะเอาเข้าฟังก์ชัน softmax เพื่อแปลงเป็นค่าความน่าจะเป็น ซึ่งค่านี้จะเปรียบเสมือน weight สำหรับ p ต่างๆ จากนั้นก็จะทำการหา weight average ของ p ออกมาเป็น vector เดียว (r) เพื่อนำไปใช้ในการคำนวณ output ต่อไป ซึ่งที่เขียนมาทั้งหมดสามารถสรุปได้เป็นสมการดังนี้ครับ อันว่า Attention นี้ นอกจากจะแก้ปัญหาคอขวดดังที่กล่าวมาแล้ว ยังถือว่าสามารถแก้ปัญหา vanishing gradient ไปด้วยพร้อมกัน นอกจากนั้นถ้าเรามาลองดูว่า output ตำแหน่งต่างๆ ให้ความใส่ใจกับ input ที่ตำแหน่งใด ก็เท่ากับว่าเราได้ alignment มาฟรีๆ ซึ่งการทำ alignment นี้ ในกรรมวิธี machine translation แบบดั้งเดิม ที่ไม่ใช่ neural machine translation ถือว่าเป็นปัญหาสำคัญทีเดียวครับ ตัวอย่าง alignment ที่ได้จาก Attention [CS224n, Lecture 11] ในทางปฏิบัติก็พบว่าเมื่อใช้ Attention เข้ามาช่วยแล้ว ได้ผลลัพธ์ที่ดีกว่าเมื่อไม่ใช้แทบจะแน่นอน ปัจจุบัน Attention จึงเป็นเหมือนท่าบังคับพื้นฐานของการทำ sequence-to-sequence learning ไปแล้ว ถ้าหากเราใช้ซอฟต์แวร์สำหรับทำงานด้านนี้อย่างเช่น OpenNMT จะพบว่ามี Attention เป็น default ให้เลย และใน OpenNMT-py ตอนนี้ยังไม่มี option ให้เอา Attention ออกได้ด้วยครับ อาจนับได้ว่า Attention เป็นกระบวนการ memory addressing รูปแบบนึง คล้ายๆ กับที่มีอยู่ใน Neural Turing Machines และผู้เขียนยังประทับใจตรงที่เรื่องนี้สามารถเปรียบเทียบกับจิตวิทยาการรับรู้ของคนเราได้อีกด้วย

มารู้จัก Transformer กันเถอะ (ตอนที่ 1)

12

มารู้จัก-transformer-กันเถอะ-ตอนที่-1-1b99fe798759

2018-06-08

2018-06-08 15:27:45

https://medium.com/s/story/มารู้จัก-transformer-กันเถอะ-ตอนที่-1-1b99fe798759

false

462

AI For Everyone

null

mena.ai

null

mena.ai

piyapoj@onionshack.com

mena-ai

AI

null

Deep Learning

deep-learning

Deep Learning

12,189

ppp mena

null

47bf9e926ee4

pppmena

2

1

20,181,104

null

0

null

0

b863424c6e6d

2018-05-18

2018-05-18 14:21:00

2018-05-18

2018-05-18 14:33:56

1

false

en

2018-05-19

2018-05-19 04:24:46

12

1b9abaddc59

2.196226

0

Karel Čapek first used the term “robot” to describe machines that resemble humans in R.U.R, a science fiction play from 1921. The word…

5

Photo from Siyan Ren Why we’ll still have to wait for the robot revolution Karel Čapek first used the term “robot” to describe machines that resemble humans in R.U.R, a science fiction play from 1921. The word originated from “robota,” the Czech word for servitude. The word choice was crucial. After all, it expresses what we expect of them. Robots have definitely begun to serve us in the last decade or two. For example, you’ll find them in science labs and on production lines. The technology continues to advance at a rapid rate as our needs evolve. However, robots have yet to be part of our everyday lives. Why haven’t we lived up to the sci-fi prophecies of a robot in every household? Look how far we’ve come Boston Dynamics uploaded a YouTube video called “Atlas, The Next Generation” in 2016. The video shows a bipedal robot that kind of looks like a human without a face. It opens doors and walks along a woodland track that is covered in snow. Afterward, the robot picks up a box and gets up from a fall after being pushed over. In 2017, the “What’s New, Atlas?” video showed a robot jumping onto boxes and somersaulting off them just like a freerunner would. We aren’t accustomed to seeing such precise human imitations from robots. But does that mean we ought to be impressed? Sophia is another robot that resembles a human due to her appearance from the front as well as her mannerisms. This humanoid can hold a conversation, tell jokes, and respond to questions. She’s certainly smart, but her usefulness isn’t obvious right away. The Robot Revolution There are some indications that the robot revolution isn’t too far off though. Doctors are already receiving help from medical bots when it comes to alleviating the side-effects of prostate cancer surgery and executing skull base surgery. Apple recently came out with Daisy, a robot that can efficiently strip old iPhones for recycling. Robots will shortly be able to cook your meals and assemble your IKEA furniture. Just like they’ve infiltrated our homes with Alexa, Amazon is also starting to create a home robot to help with our everyday tasks. There isn’t much information about the project at the moment, but this strongly suggests that robots are making their way into the mainstream. Boston Dynamics also has exciting plans for the mid-term with the creation of a collection of animal-inspired robots. The animals could be used for disaster relief and offer assistance to soldiers in the battlefield. Even our strange humanoid friend Sophia is useful in some ways. The team behind her believes she can serve as a companion and prevent elderly or sick people from becoming socially isolated. The halfway point Even though robots haven’t made a significant arrival in our regular lives just yet, there’s no doubt that they’re contributing to society. In addition, their influence is expected to expand at a rapid rate in the next decade. Let’s hope that these intelligent machines will have a positive impact on our lives. What effect do you think robots will have within the next ten years? Share your thoughts in the comments below.

Why we’ll still have to wait for the robot revolution

0

robot-revolution-is-not-that-far-away-1b9abaddc59

2018-05-19

2018-05-19 04:24:48

https://medium.com/s/story/robot-revolution-is-not-that-far-away-1b9abaddc59

false

529

We Curate, You Discover

null

gadgetflow

null

Gadget Flow

support@thegadgetflow.com

the-gadget-flow

GADGETS,CROWDFUNDING,ECOMMERCE,KICKSTARTER,TECHNOLOGY

gadgetflow

Robotics

robotics

Robotics

9,103

Evan Varsamis

Entrepreneur, Founder / CEO at Gadget Flow Inc, Investor and Advisor at Qrator Ltd, Contributor at Forbes, Inc, and Huffington Post

19c57f7f2ea0

evanvar

1,184

244

20,181,104

null

0

null

0

c7278c86ad41

2018-06-14

2018-06-14 19:43:03

2018-06-14

2018-06-14 19:48:10

2

false

en

2018-06-14

2018-06-14 19:48:10

7

1b9b0a5238ca

3.741824

0

Recently, a lot of press was given to Google’s Duplex Artificial Intelligence (AI) bot. Watch a video of the demonstration here. The Duplex…

3

Why Google Duplex Did Not Pass The Turing Test Recently, a lot of press was given to Google’s Duplex Artificial Intelligence (AI) bot. Watch a video of the demonstration here. The Duplex bot calls businesses, such as a restaurant, on your behalf and makes a reservation. It does this by actually having a conversation with the human on the other end of the line. The voice is totally realistic and in test cases the humans receiving the call were not aware that they had been talking to an AI bot. We’re going to save for another day the potential ethical issues raised by AI bots fooling people into believing they are talking to a person. Instead, let’s focus on whether or not Duplex has successfully passed the Turing test as some have suggested. Photo by Alex Knight on Unsplash A QUICK REVIEW OF THE TURING TEST The Turing test is a famous set of criteria that assess whether or not AI has become realistic enough to fool humans into thinking that they are really interacting with another person. You can find information on the Turing test here. The crux of the test is the ability for an AI process to seem human to a human. By the precise letter of the rules, one could argue that the Turing test was passed. But I don’t believe that it did in the spirit of the rules. To me, the spirit of the Turing test is having an AI bot hold a typical, rambling, somewhat random conversation with a human and pulling it off. For all the success of Duplex in scheduling restaurant reservations, it would fail miserably in this more general test. CONTEXT IS EVERYTHING In the world of AI, context is everything. There are already many examples of AI processes that initially seem quite smart but that have some major issues. See here and here for two discussions on this topic. Bias can accidentally be built into AI models through skewed training data. An equally big issue is that AI is only accurate in the exact context in which it is trained. I often use the example of two AI processes that are taught to identify two different scenarios from a photo. One is taught to determine if someone is just about to hit a tennis ball. The other is taught to determine if someone has just hit a tennis ball. When shown a picture of a person swinging a tennis racquet next to a ball, both models will with very high confidence claim to see what they are looking for. However, as humans, we know that while the picture may be ambiguous, only one answer is possible. Either the person just hit the ball, OR they are just about to hit the ball. Both can’t be true simultaneously. Further, it may not be possible to tell from the picture which is correct. The AI bots don’t know this and don’t understand that context. So, they give answers that individually seem excellent, but that are clearly incorrect when taken together. THE SHORTCOMING OF DUPLEX This gets to the heart of why Duplex really hasn’t passed the Turing test in spirit. Yes, Duplex can successfully fool a restaurant hostess when making a reservation. However, the scope and context of that conversation is very, very limited. If the hostess asked an unexpected question or used some unusual slang, the bot would have no idea what to do. The scope of a typical dinner reservation conversation is so small that you could almost pull it off with a range of more classic business logic. After all, you’re really just looking for a date and time in the request. Then, either confirming a reservation if the slot is available or offering back alternatives if it is not. Once the alternatives are offered, the person either accepts one or declines. In practice, a rules-based system could probably handle a conversation of this scope almost as well as an AI system. In reality, today’s AI bots will likely be able to more rapidly, accurately, and completely learn to navigate the typical discussions held around making a reservation than rules-based systems. My point is not to suggest that we should stop making progress with AI chat bots, but simply to point out that the success of Duplex isn’t as amazing as it might at first seem once the narrow scope is taken into account. PASSING THE TURING TEST WITHIN A GENERAL CONTEXT In order for Duplex or similar bots to pass the Turing test in spirit, they’ll need to handle much more than a discussion that fits perfectly within a tightly predefined and expected context. The bots will also need to handle any random thoughts and requests that a person might state without locking up or giving nonsense responses. Duplex’s feat, while impressive, is successful within a very narrow context. An important point to take away here is that as AI evolves and we read or hear about impressive new achievements, we must be sure to consider the context in which the models were built and tested. It is far easier to spoof people in a narrow context than it is to have a truly free flowing and spontaneous discussion. One day we’ll likely achieve the latter, but today we’ve only achieved the former. Originally published by the International Institute for Analytics

Why Google Duplex Did Not Pass The Turing Test

0

why-google-duplex-did-not-pass-the-turing-test-1b9b0a5238ca

2018-06-14

2018-06-14 19:48:22

https://medium.com/s/story/why-google-duplex-did-not-pass-the-turing-test-1b9b0a5238ca

false

890

Discussions about all things analytics because analytics is important … Analytics Matters

null

Analytics Matters

null

analytics-matters

ANALYTICS,BIG DATA,ARTIFICIAL INTELLIGENCE,STRATEGY,DATA SCIENCE

billfranksga

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Bill Franks

Bill Franks is Chief Analytics Officer for The International Institute For Analytics (IIA). You can learn more at http://www.bill-franks.com.

db909be0b020

billfranksga

4

1

20,181,104

null

0

null

0

d0c08339b41f

2018-06-27

2018-06-27 19:57:41

2018-06-20

2018-06-20 18:46:05

2

false

en

2018-06-27

2018-06-27 20:03:03

16

1b9bdbfb4e01

1.364465

0

TWiML Talk 153

5

Workforce Intelligence for Automation & Productivity with Michael Kempe TWiML Talk 153 In this episode of our PegaWorld series, I’m joined by Michael Kempe, chief operating officer at global share registry and financial services provider Link Market Services. Subscribe: iTunes / SoundCloud / Google Play / Stitcher/ RSS / Spotify In the interview, Michael and I dig into Link’s use of workforce intelligence software to allow it to track and analyze the performance of its workforce and business processes. Michael and I discuss some of the initial challenges associated with implementing this type of system, including skepticism amongst employees, and how it ultimately sets the stage for the Link’s broader use of machine learning, AI and so called “robotic process automation” to increase workforce productivity. Thanks to our Sponsor! I’d like to send a huge thanks to our friends at Pegasystems for hosting me at PegaWorld and sponsoring this series! One of the great announcements coming out of the conference was Pegasystems’ new self-optimizing, AI-powered marketing marketing capabilities. This is a really interesting offering, designed to reduce marketers’ dependence on traditional segment-based campaigns and transition them towards real-time, one-to-one customer engagement. These new capabilities will be available as a part of their new Pega Infinity platform, which was also announced at the event. For more info on Pega Infinity, head to pega.com/infinity! About Michael Michael on Linkedin Mentioned in the Interview Link Market Services Pegasystems TWiML Presents: Series page TWiML Events Page TWiML Meetup TWiML Newsletter “More On That Later” by Lee Rosevere licensed under CC By 4.0 Originally published at twimlai.com on June 20, 2018.

Workforce Intelligence for Automation & Productivity with Michael Kempe

0

workforce-intelligence-for-automation-productivity-with-michael-kempe-1b9bdbfb4e01

2018-06-27

2018-06-27 20:03:04

https://medium.com/s/story/workforce-intelligence-for-automation-productivity-with-michael-kempe-1b9bdbfb4e01

false

260

Interesting and important stories from the world of machine learning and artificial intelligence. #machinelearning #deeplearning #artificialintelligence #bots

null

twimlai

null

This Week in Machine Learning & AI

team@twimlai.com

this-week-in-machine-learning-ai

MACHINE LEARNING,ARTIFICIAL INTELLIGENCE,DEEP LEARNING,PODCAST,TECHNOLOGY

twimlai

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

TWiML & AI

This Week in #MachineLearning & #AI (podcast) brings you the week’s most interesting and important stories from the world of #ML and artificial intelligence.

ca095fd8e66c

twimlai

292

33

20,181,104

null

0

null

0

null

2018-08-27

2018-08-27 08:25:02

2018-08-27

2018-08-27 08:25:15

0

false

en

2018-08-27

2018-08-27 08:25:15

1

1b9c9de13d5a

1.449057

0

[PDF BOOK] Artificial Intelligence in the 21st Century pdf By Stephen Lucci Link…

1

READ PDF Programming Erlang By Joe Armstrong PDF Books #ebook [PDF BOOK] Artificial Intelligence in the 21st Century pdf By Stephen Lucci Link https://authorbestsipub.icu/?q=Artificial+Intelligence+in+the+21st+Century . . . . . . . . . . . . . . . . . . . Read Online PDF Artificial Intelligence in the 21st Century, Download PDF Artificial Intelligence in the 21st Century, Download Full PDF Artificial Intelligence in the 21st Century, Download PDF and EPUB Artificial Intelligence in the 21st Century, Read PDF ePub Mobi Artificial Intelligence in the 21st Century, Reading PDF Artificial Intelligence in the 21st Century, Read Book PDF Artificial Intelligence in the 21st Century, Read online Artificial Intelligence in the 21st Century, Download Artificial Intelligence in the 21st Century Stephen Lucci pdf, Download Stephen Lucci epub Artificial Intelligence in the 21st Century, Read pdf Stephen Lucci Artificial Intelligence in the 21st Century, Download Stephen Lucci ebook Artificial Intelligence in the 21st Century, Read pdf Artificial Intelligence in the 21st Century, Artificial Intelligence in the 21st Century Online Download Best Book Online Artificial Intelligence in the 21st Century, Read Online Artificial Intelligence in the 21st Century Book, Read Online Artificial Intelligence in the 21st Century E-Books, Read Artificial Intelligence in the 21st Century Online, Read Best Book Artificial Intelligence in the 21st Century Online, Read Artificial Intelligence in the 21st Century Books Online Download Artificial Intelligence in the 21st Century Full Collection, Download Artificial Intelligence in the 21st Century Book, Read Artificial Intelligence in the 21st Century Ebook Artificial Intelligence in the 21st Century PDF Read online, Artificial Intelligence in the 21st Century pdf Download online, Artificial Intelligence in the 21st Century Read, Download Artificial Intelligence in the 21st Century Full PDF, Read Artificial Intelligence in the 21st Century PDF Online, Read Artificial Intelligence in the 21st Century Books Online, Read Artificial Intelligence in the 21st Century Full Popular PDF, PDF Artificial Intelligence in the 21st Century Read Book PDF Artificial Intelligence in the 21st Century, Read online PDF Artificial Intelligence in the 21st Century, Download Best Book Artificial Intelligence in the 21st Century, Read PDF Artificial Intelligence in the 21st Century Collection, Read PDF Artificial Intelligence in the 21st Century Full Online, Read Best Book Online Artificial Intelligence in the 21st Century, Download Artificial Intelligence in the 21st Century PDF files

READ PDF Programming Erlang By Joe Armstrong PDF Books #ebook

0

read-pdf-programming-erlang-by-joe-armstrong-pdf-books-ebook-1b9c9de13d5a

2018-08-27

2018-08-27 08:25:15

https://medium.com/s/story/read-pdf-programming-erlang-by-joe-armstrong-pdf-books-ebook-1b9c9de13d5a

false

384

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

aphrodite

null

9651bc0396b2

aphrodite_66125

0

1

20,181,104

null

0

null

0

7ff81454a4c3

2018-01-14

2018-01-14 13:34:30

2018-01-15

2018-01-15 09:14:12

1

false

en

2018-04-12

2018-04-12 20:06:31

4

1b9e71b4a1ac

1.996226

0

Common Myths About Data Science

5

Alastair Majury on Common Myths About Data Science Alastair Majury is an experienced Business Data Analyst Common Myths About Data Science Data science is a unique industry. From an outsider’s perspective, it may appear to be very confusing and impossible to understand, but it isn’t. Much like any other industry, when you take the time to learn about something, step-by-step, you begin to understand the bigger picture. Data science is often misunderstood and misinterpreted. What I’d like to do is to help dispel any of those rumors, myths and misconceptions. Data Science and Big Data Have Nothing In Common Nothing could be further from the truth. Big data is a general term for massive amounts of information that can be recorded, analyzed, processed and used for a variety of applications. Data science, in layman’s terms, looks for patterns in complex systems in order to simplify it and make sense. Data scientists need data in order to perform their duties. It is a crucial step in any data scientist’s process. Data Scientists Simply Collect Data One of the most common misconceptions about data scientists is that we all just sit around and crunch numbers and collect data. That is incredibly untrue. While collecting data is an aspect of the job, data scientists do far more than just that. As mentioned before, data scientists take that information, analyze it, simplify it, and visualize it. Data scientists are required to analyze, interpret and, most importantly, communicate the data to organizations or businesses. They are an integral part of a company’s decision-making process. Data Scientists Are Also Developers It is assumed that just because data scientists are really good with computers, data and technology, they must be skilled developers. It is completely possible that there are several data scientists in the world who are also talented developers, but the two skills are not a package deal. Admittedly, it can certainly help a data scientist to know how to develop, it is by no means a requirement for the job. Data science, to many, is still too confusing. However, it is one of the fastest-growing industries in the world and as it becomes part of everyday society, its complicated concepts and structure will undoubtedly become easier to understand. Alastair Majury resides locally in Dunblane, and is an IT Consultant working across the country. Alastair is also a volunteer officer at the local Boys’ Brigade company, a charity which focuses on enriching the lives of children and young people, and building a stronger community. Alastair Majury is also a highly experienced Senior Business Analyst / Data Scientist with a proven track record of success planning, developing, implementing and delivering migrations, organisational change, regulatory, legislative, and process improvements, when providing my Senior Business Analyst / Data Scientist services for global financial organisations, covering the Challenger Bank, Retail Banking, Investment Banking, Wealth Management, and Life & Pensions sectors.

Alastair Majury on Common Myths About Data Science

0

alastair-majury-on-common-myths-about-data-science-1b9e71b4a1ac

2018-04-12

2018-04-12 20:06:32

https://medium.com/s/story/alastair-majury-on-common-myths-about-data-science-1b9e71b4a1ac

false

476

Our team is made up of 2 MSc Data Science students who want a place for readers to find out more about the field of Data Science and some interesting applications of the field. Stories will range from projects to informative stories. Follow, share and enjoy your read!

null

DataRegressed

dataregressed@gmail.com

dataregressed

DATA SCIENCE,PROGRAMMING,MATHEMATICS,COMPUTER SCIENCE,DATA

null

Data Science

data-science

Data Science

33,617

Alastair Majury

Alastair Majury resides locally in Dunblane, and is an IT Consultant working across the UK. Alastair Majury is an experienced Business Data Analyst in FS.

7a7c3b99f8e5

majury1981

16

2

20,181,104

null

0

:param url: https://query.data.world/s/xxxxxx CALL apoc.load.json($url) YIELD value RETURN count(*); ╒══════════╕ │"count(*)"│ ╞══════════╡ │373663 │ └──────────┘ call apoc.load.json($url) yield value return value limit 1; { "update_time": { "$date": "2015-02-12T21:23:12.929Z" }, "home_name": "F5 Networks", "home_domain": "f5.com", "_id": { "$oid": "54dd19c08899a4c549dc71cf" }, "link_name": "Dell", "type": "partnership", "link_domain": "dell.com", "username": "rjurney" } CALL apoc.load.json($url) YIELD value WITH distinct value.home_domain as domain, value.home_name as name RETURN count(*) ╒══════════╕ │"count(*)"│ ╞══════════╡ │24266 │ └──────────┘ CREATE CONSTRAINT ON (b:Business) ASSERT b.domain IS UNIQUE; CALL apoc.load.json($url) YIELD value WITH distinct value.home_domain as domain, value.home_name as name CREATE (:Business {domain:domain, name:name}) call apoc.load.json($url) yield value with distinct value.link_domain as domain, value.link_name as name merge (:Business {domain:domain}) ON CREATE SET b.name = name; call apoc.load.json($url) yield value return toUpper(value.type), count(*); ╒═════════════════════╤══════════╕ │"toUpper(value.type)"│"count(*)"│ ╞═════════════════════╪══════════╡ │"CUSTOMER" │80465 │ ├─────────────────────┼──────────┤ │null │114 │ ├─────────────────────┼──────────┤ │"SUPPLIER" │79401 │ ├─────────────────────┼──────────┤ │"INVESTMENT" │71630 │ ├─────────────────────┼──────────┤ │"PARTNERSHIP" │112953 │ ├─────────────────────┼──────────┤ │"COMPETITOR" │29100 │ └─────────────────────┴──────────┘ call apoc.load.json($url) yield value with * where value.type IS null return value limit 10; ╒═════════════════════════════════════════════════════════════════╕ │"value" │ ╞═════════════════════════════════════════════════════════════════╡ │{"link_domain":"navinet.net","home_domain":"csc.com","_id":{"$oid│ │":"574e15023bf9e624d32e1e0a"}} │ ├─────────────────────────────────────────────────────────────────┤ │{"link_domain":"cisco.com","home_domain":"pgi.com","_id":{"$oid":│ │"574e16043bf9e624d32e25d0"}} │ ├─────────────────────────────────────────────────────────────────┤ │{"link_domain":"pgi.com","home_domain":"micron.com","_id":{"$oid"│ │:"574e16043bf9e624d32e25d1"}} │ ├─────────────────────────────────────────────────────────────────┤ │{"link_domain":"pgi.com","home_domain":"ibm.com","_id":{"$oid":"5│ │74e16283bf9e624d32e26c2"}} │ call apoc.load.json($url) yield value where value.type is null match (from:Business {domain:value.home_domain}) match (to:Business {domain:value.link_domain}) create (from)-[:RELATED_TO]->(to); call apoc.load.json('https://query.data.world/s/_kQZwISTfInOevAL2Cy2SelUkue4NS') yield value with toUpper(value.type) as type, value WHERE type = 'CUSTOMER' match (from:Business {domain:value.home_domain}) match (to:Business {domain:value.link_domain}) create (from)-[:CUSTOMER {user:value.username, date:apoc.date.parse(value.update_time.`$date`,'s', "yyyy-MM-dd'T'HH:mm:ss.SSS'Z'")}]->(to); CALL apoc.load.json($url) YIELD value WITH value WHERE toUpper(value.type) = 'COMPETITOR' MATCH (from:Business {domain:value.home_domain}) MATCH (to:Business {domain:value.link_domain}) CREATE (from)-[rel:COMPETITOR]->(to) ON CREATE SET rel.user = value.username, rel.date = apoc.date.parse(value.update_time.`$date`,'s',"yyyy-MM-dd'T'HH:mm:ss.SSS'Z'"); MATCH (b1:Business)-[s:SUPPLIERS]->(b2) RETURN b1,s,b2 LIMIT 10; match (n:Business) return count(*); ╒══════════╕ │"count(*)"│ ╞══════════╡ │51104 │ └──────────┘ match (n:Business)-[r]->() return type(r), count(*) order by count(*) desc; ╒═════════════╤══════════╕ │"type(r)" │"count(*)"│ ╞═════════════╪══════════╡ │"PARTNERSHIP"│139998 │ ├─────────────┼──────────┤ │"CUSTOMER" │95405 │ ├─────────────┼──────────┤ │"SUPPLIER" │89401 │ ├─────────────┼──────────┤ │"INVESTMENT" │78085 │ ├─────────────┼──────────┤ │"COMPETITOR" │30660 │ ├─────────────┼──────────┤ │"RELATED_TO" │114 │ └─────────────┴──────────┘ MATCH (b:Business) where size((b)--()) < 20 MATCH path = (b)--(b2)--(b3) WHERE size((b2)--()) < 20 and size((b3)--()) < 20 RETURN path LIMIT 700; call apoc.periodic.iterate(" call apoc.load.json($url) "," merge (from:Business {domain:value.home_domain}) ON CREATE SET home.name = value.home_name merge (to:Business {domain:value.link_domain}) ON CREATE SET link.name = value.link_name with *, apoc.date.parse(value.update_time.`$date`,'s', \"yyyy-MM-dd'T'HH:mm:ss.SSS'Z'\") as date call apoc.create.relationship(from, toUpper(value.type), apoc.map.clean({updated:date,id:value._id.`$oid`, user:value.username},[],[null]) ,to) yield rel return count(*)", {batchSize:10000,iterateList:true, params:{url,$url}); call apoc.stats.degrees() MATCH (b:Business) WITH b, size( (b)-[:COMPETITOR]-() ) as degree RETURN b.name, b.domain, degree ORDER BY degree DESC LIMIT 5 ╒═══════════╤═══════════════╤════════╕ │"b.name" │"b.domain" │"degree"│ ╞═══════════╪═══════════════╪════════╡ │"Google" │"google.com" │598 │ ├───────────┼───────────────┼────────┤ │"Microsoft"│"microsoft.com"│533 │ ├───────────┼───────────────┼────────┤ │"Facebook" │"facebook.com" │462 │ ├───────────┼───────────────┼────────┤ │"Apple" │"apple.com" │417 │ ├───────────┼───────────────┼────────┤ │"IBM" │"ibm.com" │379 │ └───────────┴───────────────┴────────┘ ╒═════════════════════╤════════════════╤════════╕ │"b.name" │"b.domain" │"degree"│ ╞═════════════════════╪════════════════╪════════╡ │"Cisco" │"cisco.com" │22982 │ ├─────────────────────┼────────────────┼────────┤ │"Microsoft" │"microsoft.com" │3582 │ ├─────────────────────┼────────────────┼────────┤ │"Rackspace" │"rackspace.com" │3572 │ ├─────────────────────┼────────────────┼────────┤ │"Amazon Web Services"│"aws.amazon.com"│2336 │ ├─────────────────────┼────────────────┼────────┤ │"IBM" │"ibm.com" │1918 │ └─────────────────────┴────────────────┴────────┘ call algo.pageRank(); MATCH (b:Business) where exists(b.pagerank) RETURN b.name, b.domain, b.pagerank ORDER BY b.pagerank DESC LIMIT 5 ╒═══════════╤═══════════════╤══════════════════╕ │"b.name" │"b.domain" │"b.pagerank" │ ╞═══════════╪═══════════════╪══════════════════╡ │"Microsoft"│"microsoft.com"│184.0849425 │ ├───────────┼───────────────┼──────────────────┤ │"Google" │"google.com" │174.71597049999997│ ├───────────┼───────────────┼──────────────────┤ │"IBM" │"ibm.com" │124.25576300000002│ ├───────────┼───────────────┼──────────────────┤ │"Facebook" │"facebook.com" │124.16897800000001│ ├───────────┼───────────────┼──────────────────┤ │"Apple" │"apple.com" │89.4175015 │ └───────────┴───────────────┴──────────────────┘ call algo.betweenness('Business','PARTNERSHIP'); MATCH (b:Business) where exists(b.centrality) RETURN b.name, b.domain, b.centrality ORDER BY b.centrality DESC LIMIT 5 ╒═══════════════════════════════╤═══════════════════╤══════════════╕ │"b.name" │"b.domain" │"b.centrality"│ ╞═══════════════════════════════╪═══════════════════╪══════════════╡ │"CA Technologies" │"ca.com" │21474.83647 │ ├───────────────────────────────┼───────────────────┼──────────────┤ │"NVIDIA" │"nvidia.com" │21474.83647 │ ├───────────────────────────────┼───────────────────┼──────────────┤ │"Mphasis Australia Pty Limited"│"mphasis.com" │21474.83647 │ ├───────────────────────────────┼───────────────────┼──────────────┤ │"Datapipe" │"datapipe.com" │21474.83647 │ ├───────────────────────────────┼───────────────────┼──────────────┤ │"MicroStrategy" │"microstrategy.com"│21474.83647 │ └───────────────────────────────┴───────────────────┴──────────────┘ call algo.labelPropagation('Business','CUSTOMER','OUTGOING',{iterations:10}); match (b:Business) return count(distinct b.partition) as partitions ╒═════════════╤══════╤════════════════════════╕ │"b.partition"│"size"│"topRanked" │ ╞═════════════╪══════╪════════════════════════╡ │68 │7900 │"Microsoft" │ ├─────────────┼──────┼────────────────────────┤ │13441 │7 │"UC Davis" │ ├─────────────┼──────┼────────────────────────┤ │25359 │6 │"Delta Dental" │ ├─────────────┼──────┼────────────────────────┤ │7463 │6 │"LPL Financial Services"│ ├─────────────┼──────┼────────────────────────┤ │2508 │6 │"MillerCoors" │ ├─────────────┼──────┼────────────────────────┤ │19829 │6 │"conde-nast" │ ├─────────────┼──────┼────────────────────────┤ │18813 │4 │"RadiumOne" │ ├─────────────┼──────┼────────────────────────┤ │39918 │4 │"Vail Resorts" │ ├─────────────┼──────┼────────────────────────┤ │48510 │4 │"Arch Coal" │ ├─────────────┼──────┼────────────────────────┤ │48922 │4 │"IIT" │ └─────────────┴──────┴────────────────────────┘

42

c0e10f7c3cc0

2017-11-11

2017-11-11 03:44:25

2017-11-11

2017-11-11 04:13:01

6

false

en

2018-06-08

2018-06-08 12:45:35

8

1b9f16404951

8.025472

17

1

Late Night Tweet

5

Quick Import: Relato Business Graph Database from Data.World into Neo4j Late Night Tweet On Friday night I saw this tweet by Russel Jurney which sparked my interest. Him open-sourcing graphy business data that he has been been working on for quite some time trying to build a business is something we have to be grateful for. He links to a longer blog post explaining the history and the content of the data. As part of that he points us to the Data.World repository where he made the roughly 400k connections available to everyone to use. As I already had created an account on Data.World for our [Graphs for Good Hackation] in October in NYC, I could access the data immediately. You can query the data using SparQL, download the JSON or even grab an URL to access it remotely, which was what I did for ease of use. So please do the same and copy that URL somewhere safe. Neo4j As I wanted to import it into Neo4j, I grabbed the Neo4j Desktop package for the latest 3.3.0 release from the website. This comes with a free develper license of Neo4j Enterprise, which is really useful. The Neo4j Desktop is an Electron app that allows you to manage multiple databases of different versions and install extensions and run the visual database browser. Neo4j Desktop & Browser So I created a new Project & Database with Neo4j 3.3.0 Enterprise and installed the “APOC” procedures Plugin, which is a really useful collections of tools. Now I started my database, opened my browser and could start accessing the data. To make it easy we set the URL as parameter, which we can access later with $url. If you want to learn more about Neo4j, check out http://neo4j.com/developer/get-started For an overview of the Cypher query language, the reference card is really useful. Now we can start querying our URL (you can also download it locally and load it from there for faster turnaround). The Data We use the apoc.load.json procedure which will give us a stream of JSON records, so we can use this already to query and analyse remote data. Similar procedures are available for most other databases as well as CSV and XML. Our file has 374k records We can also look at the first few json objects, to see which data they contain. Data Import — Businesses Now we can start creating our data. Basically 2 businesses are linked by a relationship. We could create the full graph in one go, but to make it easier to follow, we take smaller steps here. Let’s start with the businesses first. As in the links, most businesses appear multiple times, we can use their distinct occurrence. We start with the left sidewhich Russell calls home. I want to use domain as the business-id, as it is more unique. So we create a constraint for it. Then we load our businesses and create nodes for them. Now we continue with the right side, which is called link. Note that businesses on the right might already have appeared on the left, side, so if we used CREATE as before, we could get a constraint violation for the duplicates. That’s why we’ll use MERGE. So now we have a lot of Business nodes in our database, but they are not connected, so that’s not a graph. Data Import — Links Let’s create the connections, between home and link, using the type and storing date and user as properties on the relationship. Easiest would be to do a batch at a time, we can look at the distribution of links by type and that’s quite evenly spaced out. We have to deal with the null value though, which we can also look at. We see for these records there are things missing: the names of the companies (good that we used domain as identifiers) the link type the date and user Let’s start with these records then, as they are easier. We need to MATCH businesses by domain and then CREATE relationships between them, for the unspecified ones, we just use RELATED_TO. If Neo4j complains about Out-Of-Memory, the Neo4j-Desktop configures your database initially only with small memory settings. If you go to the “Settings” tab, you can increase the dbms.heap.* values to 500M or 1G. We can do the same import for the other types we’ve seen, e.g. CUSTOMER. This time we also want to store the user and the date, but actually not as ISO-8601 in UTC, but as seconds since Epoch. For that we parse the string with this function apoc.date.parse(value.update_time.`$date`,'s', "yyyy-MM-dd'T'HH:mm:ss.SSS'Z'"). We run a similiar statement for the other types: SUPPLIER INVESTMENT PARTNERSHIP COMPETITOR As some of these records would create two links, one per direction, we can choose to use MERGE instead and leave off the directional arrow-tip. Data Inspection So now that we have all our data imported, we can have a first look. For instance just showing a bunch of businesses and their `SUPPLIER`s. To see a larger part of the graph, we can chose a few businesses that don’t have too many relationships (we don’t want to see a hairball). And follow their connections 2 hops out, using the same condition on the whole path. Parts of the Relato Business Graph Alternative Import Approach As mentioned before, we can also combine all of these steps into one load operation. Here we use: apoc.periodic.iterate for batching work on a stream of records (almost 400k) coalesce to provide defaults null values apoc.create.relationship to create relationships with dynamic types apoc.map.clean to remove null values from properties Data Analytics We can look at the distribution of degrees in the data using apoc.stats.degrees which takes an optional argument of the relationship-type. Here we see that while most nodes only have around 20 relationships, some go up to 1200 and even to 26000 as maximum. What are the nodes with the biggest competition in our graph, i.e. the biggest degree ? Obviously the usual suspects. For partnerships it looks a bit different, Cisco is a clear leader here and AWS in the top 5. We can also apply ranking (e.g. page-rank) and clustering on this data. To enable that we install the neo4j-graph-algorithms library. The latest release can be found here. From there we grab the graph-algorithms-algo-3.3.0.0.jar file and drop it into the plugins folder when you Open Folder. For this to work, we have to add this config setting dbms.security.procedures.unrestricted=algo.* and restart the server. This call computed the page-rank in 2 seconds and wrote the results to our business nodes. What are the highest ranking nodes in our graph? Again, not surprising. What does it look like for betweenness centrality, i.e. businesses which connect other clusters of businesses. Let’s try that for the PARTNERSHIP relationship. This one takes longer to compute (3 mins on my Mac) as it needs to run all shortest paths in the graph to see which nodes most frequently appear on them. Last but not least we can also cluster our businesses, e.g. the CUSTOMER graph. This returns after 2 seconds. We have 42559 partiions. How big are the top-largest partitions and what are their most highly ranked nodes. We can clearly see the different industries being separated, with software being the largest. I hope this was useful for you to get up and running with a graph database quickly, both in terms of getting data imported but also analyizing it quickly.

Quick Import: Relato Business Graph Database from Data.World into Neo4j

95

quick-import-relato-business-graph-database-from-data-world-into-neo4j-1b9f16404951

2018-06-20

2018-06-20 18:55:12

https://medium.com/s/story/quick-import-relato-business-graph-database-from-data-world-into-neo4j-1b9f16404951

false

1,875

Developer Content around Graph Databases, Neo4j, Cypher, Data Science, Graph Analytics, GraphQL and more.

null

neo4j.graph.database

null

neo4j

devrel@neo4j.com

neo4j

DATABASE,GRAPH DATABASE,DATA SCIENCE,SOFTWARE DEVELOPMENT,GRAPHQL

neo4j

Relato

relato

Relato

3,910

Michael Hunger

A passionate aspiring software craftsman with too many ideas and projects in different fields.

3865848842f9

mesirii

1,190

89

20,181,104

null

0

def process(sentence): # The processing done is stopword removal and stemming sentence_split = sentence.split() sentence_words = [ps.stem(word) for word in sentence_split if word not in stopWords] result = ''.join(sentence_words) return result count_vect = CountVectorizer() X_counts = count_vect.fit_transform(X_train) tf_idf_transformer = TfidfTransformer() X_train = tf_idf_transformer.fit_transform(X_counts) X_train.shape from sklearn.pipeline import Pipeline from sklearn.naive_bayes import MultinomialNB model = Pipeline([('count_vect', CountVectorizer()), ('tf_idf_transformer', TfidfTransformer()), ('classifier', MultinomialNB())]) model.fit(X_train, y_train) import numpy as np predicted = model.predict(y_train) np.mean(predicted == y_test) from sklearn.linear_model import SGDClassifier model_svm = Pipeline([('vect', CountVectorizer()), ('tfidf', TfidfTransformer()), ('clf-svm', SGDClassifier(loss='hinge', penalty='l2', alpha=1e-3, n_iter=5, random_state=42)),]) model_svm.fit(X_train, y_train) predicted_svm = model_svm.predict(X_test) np.mean(predicted_svm == y_test) from sklearn.metrics import classification_report print(classification_report(y_test, predicted_svm)) from sklearn.metrics import confusion_matrix confusion_matrix(y_test, predicted_svm) from sklearn.model_selection import GridSearchCV parameters_svm = {'vect__ngram_range': [(1, 1), (1, 2)], 'use_idf': (True, False), 'alpha': (1e-2, 1e-3)} gs_model_svm = GridSearchCV(model_svm, parameters_svm, n_jobs=1) gs_model_svm = gs_model_svm.fit(X_train, y_train) # To find out the best parameters for the model gs_model_svm.best_params_

8

null

2018-05-12

2018-05-12 08:58:35

2018-05-13

2018-05-13 06:22:08

3

false

en

2018-05-14

2018-05-14 14:52:21

2

1ba01021dae7

5.274528

4

0

Implementing sentiment classification using Naive Bayes Classifier and Support Vector Machine.

5

Sentiment Analysis classification using 2 different methods. So in sequel to my previous post about sentiment analysis, I felt its time to build one straight on. The previous post explain some few details about sentiment analysis and how it can be used to generate insight from reviews, available on the web. In this implementation, we are going in full details and I would try my best to explain steps to take in building a simple sentiment classifier with sample codes. We are going to use the IMDB dataset, and we would compare our results with 2 different algorithms which would be similar if you want to use another algorithms. So lets get started shall we? Consider the figure below, the process is similar to any data mining process and we are going to reference this later on. Process for our Sentiment Analysis Task Data Well we are doing data science, so we gotta talk about the data we want to use right? In most NLP tasks, getting the data for your job isn’t rosy and at the same time not rocket science. Mostly, it’s crawling text data from digital platforms. Since we are dealing with natural languages, most times the data are on the web. Luckily, we aren’t building a gigantic project so we don’t need to scrape the web for our task as there are availiable data we could use. Gloory to open data !! We would be using IMDB data set. Its a movie review datasets with 50,000 movie reviews equally separated for training and testing. You could download the datasets from the official site here Data Cleaning Just like they say, “garbage in, garbage out”. It is necessary to supply our Machine learning algorithm with clean data so we could get a good result afterwards. Alright if you take a glimpse of the data you’ve just downloaded, you would see something like what I have in the snapshot below. Hell yeah! Its messy and it’s our job as a data scientist to clean it and make it ready for whatever form of analysis we need. We are going to make use of 2 common pre-processing method. Stemming and stop word removal. Stemming is a processing tool in natural language processing that puts together different variations of a token. Say for example the word ‘dance’ would have variations like ‘dancing’, ‘danced’ and so on. So in most cases, we need to stem the tokens (words). After stemming, the next thing is stop word removal. This means there are words that commonly occur is a datasets (corpus) should be removed. In most cases, words like articles (e.g the), pronouns (e.g I) need to be removed. Stop word removal is a common practice of data cleansing in NLP tasks and such words have no relevance in classification, information retrieval, word clustering or analysis of any kind. Sample view of IMDB data Feature Extraction Now we have our cleaned data, we should understand that the data is still in natural language format but the data has to be transformed into a vector form which is a requirements for ML algorithms. The crucial bottleneck of sentiment classification is engineering an effective set of features which are used in a feature based supervised statistical classifier. Examples of feature engineering technique in this case include, TF-IDF, Part of Speech Tagging, Opinion words/phrases. We would be using TF-IDF. TF-IDF (term frequency-inverse document frequency) refers to the a statistical weight that measures how important a word is in the document or corpus. Fortunately, scikit-learn has an implementation for us to use. The code below transform our cleaned data into vector form called bag of words (BOW) and if you get the shape of the data, you would find something like (50000,36687) now we are ready for building our ML model for sentiment classification. Build Model Okay seems we’re good to start training our model. Most times we are faced with question like “What machine learning algorithm should I use for this task?” Well there’s a long discussion on that but it’s often a good practice to start small first. and yeah we are just going to use 2 different models. The task here is to implement a sentiment analysis or a classification model using Multinomial Naive Bayes Classifier (MNB) and Support Vector Machine (SVM). There are various algorithms we could use for this classification task. MNB requires a number of parameters which are linear. I am not assuming you know already about these two algorithms but you could check other tutorial about the algorithms. Note in the code above that we use a pipeline and the reason for using a pipeline is to assemble several steps that can be cross-validated together while setting different parameters. We repeat the same for SVM classifier in the code below Well that is it. We’re done. Err not yet. It’s important we know how our model fares with new instances. That is we need to get a report of our model classification. In scikit-learn, we could implement the classification report. We could do that in 2 lines of code. Yass!! It generates a table-like report like the one shown below. Precision means the accuracy of the positive predictions and recall is the sensitivity or true positive rate that is the ratio of positive instances that are correctly classified by the classifier. f1-score is a combination of precision and recall into a single metric and its like an harmonic mean that gives more weight to low values. So if both precision and recall have high score, then f1 would also have a high score. Another report we could look at is the confusion matrix of the classifier which is a way to measure the number of times instances of a particular class is classified another class. In our case, it counts the number of instance positive sentiments are classified as negative sentiments and vice versa. We could implement that with code below. We should note that, we could use cross -validation method for evaluating out model and before that use a Grid-search method to select the best possible parameters. An example of using a grid-search method for selecting the best parameters for SVM model example is given below With the above codes, you can successfully build a sentiment classifier model. If you have any concern about it, put a comment down in the comment box and I would reply. Good-luck in building your model can’t wait for you to implement yours. You could share your GitHub link of the implementation of you have done it. I would like to see your implementation.

Sentiment Analysis classification using 2 different ML methods.

32

sentiment-analysis-classification-using-2-different-ml-methods-1ba01021dae7

2018-05-14

2018-05-14 14:52:22

https://medium.com/s/story/sentiment-analysis-classification-using-2-different-ml-methods-1ba01021dae7

false

1,252

null

Machine Learning

machine-learning

Machine Learning

51,320

Jerry Fadugba

Data Scientist, Machine Learning Engineer. Passionate about the social impact of Artificial Intelligence.

6598c1e02740

jerryfadugba

5

9

20,181,104

null

0

null

0

d4b105fed640

2018-09-11

2018-09-11 07:51:16

2018-09-11

2018-09-11 20:23:31

4

false

en

2018-09-12

2018-09-12 07:46:22

9

1ba091d72e0a

4.771698

0

Neuroplasticity is real — that is, the brain really can change and learn and improve based on experience. So what should we be doing to…

5

Brainless or Brainfull? Neuroplasticity is real — that is, the brain really can change and learn and improve based on experience. So what should we be doing to train our brain? Is technology the downfall or savior of our brain? The Reading Brain A neuroscientist explains what tech does to the reading brain Spoiler: we need to read as a bridge back to yourself, and these bridges are necessary The UCLA neuroscientist, a great lover of literature, Maryanne Wolf is horrified at what has happened to our ability to concentrate. The ways we process written language have changed dramatically. Our ability to read deeply has been impacted as we have become, inevitably, increasingly dependent on screens. Concerns about attention span, critical reasoning, and over-reliance on technology — Wolf herself has found that, though she is a reading expert. When we have any function, whether it’s language or vision or cognitive functions like memory, we aren’t dealing with a straight line to the brain that says “This is what I do.” The brain builds a network of connections, a network of neurons that have a particular role in that function. So when we have a new cognitive function, like literacy, it doesn’t have a preset network. Rather, it makes new connections among older networks, and that whole collection of networks becomes a circuit. It’s a connected scaffolding of parts. The beauty of the circuit for functions like literacy is its plasticity. The first circuits are very basic — for decoding letters as we’re learning to read — but everything we read builds upon itself. The fact that a circuit is plastic is both its beautiful strength and its Achilles’ heel. Reading reflects our medium. And to the extent that a digital medium is going to require us to process large amounts of information very quickly, it will diminish from the time we have for slower processing work. And these slower processes are deep learning, the ones that are more cognitively challenging. They’re the basis for going beyond that initial short circuit of decoding the information and understanding it at a very basic level. The digital medium affordance rewards and advantages fast processing at the cost of the slower processes that build our very important critical, analytical, and empathetic processes. People who spend six to 12 hours a day on a screen are led to use the skimming mode even knowing they should use a more concentrated, focused mode of reading. Skimming has led to a tendency to go to the sources that seem the simplest, most reduced, most familiar, and least cognitively challenging. My biggest worry now is that a lot of what we’re seeing in society today — this vulnerability to demagoguery in all its forms — of one unanticipated and never intended consequence of a mode of reading that doesn’t allow critical analysis and empathy. For anyone who has ever been a reader, there’s much to sympathize with in Maryanne Wolf’s Reader, Come Home. Emotional AI A new way to be attached to our devices Technology companies are designing products that simulate sentience and life Anki, the San Francisco–based robotics company whose toy robot, Cozmo, was a best-seller on Amazon in 2017, has released an even more sophisticated upgrade called Vector. The autonomous “home robot” runs on a neural network and responds to and makes eye contact, Fast Company reported. He (how the creators refer to Vector) dances when you turn on music, watches TV, and fist-bumps. Thanks to sensors on his head, he coos and purrs when you pet him. Connected via Wi-Fi, Vector is always on and knows what’s going on, but is also down with just hanging out. Photo: courtesy of Anki Beautiful Mind The beauty and complexity of consciousness THIS is what consciousness looks like — but these aren’t brain scans. The thalamus and basal ganglia, which govern our senses, movement and decision-making, Greg Dunn and Brian Edwards “The piece was designed to be an unprecedented image of the brain. We’re demonstrating the depth and breadth of neural activity that allows us to go about our existence,” says Greg Dunn of his project, titled Self Reflected. To create the artworks, Dunn first collected reams of information on the human brain, including scans and detailed depictions of neurons, and how they connect to each other. He used these as inspiration for hand drawings on transparent sheets. Working with the artist and physicist Brian Edwards, Dunn fed these drawings through a computer model that mimics how neurons communicate with each other, simulating the movement of signals throughout the brain. The pair then printed the resulting patterns using a technique that etches layers of gold leaf. As a result, the images appear to come to life as light moves across them, highlighting different layers of neurons and the flow of information between them. The cerebellum, which oversees movement and proprioception — an awareness of one’s own body, Greg Dunn and Brian Edwards Social (non)addiction? Young people are moving away from the Facebook app 44 Percent of Americans 18–27 Have Deleted the Facebook App This Year, Poll Finds Facebook has faced a steady stream of criticism not only for the Cambridge Analytica scandal, but for initially threatening to sue news outlets that reported on it. The platform has also repeatedly come under fire for failing to thwart abuse of its platform to spread noxious disinformation everywhere from the United States to the Philippines. How much pain these scandals have actually inflicted on Facebook remains a point of contention. Still, while Facebook has continually promised to learn from its mistakes and protect end users from such abuse moving forward, it’s clear many believe that simply eliminating Facebook from their lives entirely is a quicker and and more efficient solution. Brain Boost The brain’s emotion code was cracked This device reads your brainwaves to figure out your mood A team of neuroscientists figured out a way to read a person’s mood by analyzing their brainwaves. It’s more than just a party trick — it’s first time scientists have made an explicit connection between brainwaves and emotional states, and it could have far-reaching implications for the future of treatments for mood disorders. Photo: Futurism The researchers believe their work could pave the way for treatments that build on deep brain stimulation, a technique in which an implant stimulates the brain to treat conditions such as obsessive compulsive disorder and major depression. Farther into the future, such an implant could keep tabs on a patient’s emotional state — and, if it detected an abnormality, give them a jolt to help bring them back to normal. Those who read books will always be ahead of people who scroll social feed!

Brainless or Brainfull?

0

from-brainless-to-brainfull-1ba091d72e0a

2018-09-12

2018-09-12 07:46:22

https://medium.com/s/story/from-brainless-to-brainfull-1ba091d72e0a

false

1,079

Random thoughts about digital transformation and the future of web design

null

hainteractive

null

H A Things

hello@hainteractive.com

hainteractive

DIGITAL TRANSFORMATION,WEB DESIGN,WEB DEVELOPMENT,TECHNOLOGY TRENDS,UX DESIGN

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

H A

Interactive Brand Studio. We focus on creatively and strategically driven innovative digital solutions. https://hainteractive.com

aaacc19e4aaf

hainteractive

11

17

20,181,104

null

0

null

0

null

2018-01-21

2018-01-21 01:52:54

2018-01-21

2018-01-21 02:16:09

1

false

en

2018-01-21

2018-01-21 02:16:09

2

1ba0d56931

1.781132

5

1

0

Our first Cloud AutoML release will be Cloud AutoML Vision, a service that makes it faster and easier to create custom ML models for image…

5

AutoML from Google Cloud — Free Up ML Resources to Focus on Harder Things Our first Cloud AutoML release will be Cloud AutoML Vision, a service that makes it faster and easier to create custom ML models for image recognition. Its drag-and-drop interface lets you easily upload images, train and manage models, and then deploy those trained models directly on Google Cloud. Google Cloud Platform announced the alpha launch of AutoML Vision this week. The service aims to help developers with no machine learning expertise to easily create a customized and well-tuned machine learning model for image recognition. It is not the first system of its kind on the market: Clarif.ai, Microsoft’s Cognitive Services, and Salesforce Einstein Vision all provide similar services to the market. It neither means Machine Learning engineers will be replaced: some repetitive work will be automated; however, those more complicated and domain-specific machine learning works still require a lot of machine learning experts and data engineers. However, if the AutoML service proves to be effective and expands to other areas beyond vision, such natural language generation and anomaly detection, two groups will largely be benefited from it: First, the Google Cloud team can provide more AI-driven services to their clients, with AutoML as building blocks to higher level customization work. This do not necessarily mean they will get higher profit margin, as other competitors like AWS and Azure may quickly release similar services. Instead, it serves as a toolbox for every Google Cloud consultant: the tool can be used to can speed up the repetitive, low-level work, so that the consultant can focus on the more difficult, domain-specific problems and quickly deliver customizing solutions. Second, more importantly, those who have very limited access to the machine learning expertise and data resources, such as startups, University researchers (in non-CS and Engineering fields), and social ventures, can close their gaps with large organizations, and focus on solving problems in their domains. For example, psychology researchers can leverage the powerful image recognition service to detect facial changes of subjects in the video of experiments, without the need to hire a bunch of undergrad students to do so; more sociology researches like this one from Stanford can be carried out more easily. Research from Stanford on Estimating the Demographic Makeup of Neighborhoods with Deep Learning In short, AutoML and other similar services do have the potential to make AI more accessible to everyone, so that we can focus on much harder problems. We are far from automating the entire pipeline of Machine Learning development.

AutoML from Google Cloud — Free Up ML Resources to Focus on Harder Things

9

automl-from-google-cloud-free-up-ml-resources-to-focus-on-harder-things-1ba0d56931

2018-04-23

2018-04-23 00:07:10

https://medium.com/s/story/automl-from-google-cloud-free-up-ml-resources-to-focus-on-harder-things-1ba0d56931

false

419

null

Google Cloud

google-cloud

Google Cloud

207

Ivan Zhou

ML Researcher, Data Scientist in Shopify. https://www.ivanthinks.com/blog/

83e0c547ee57

ivanthinks

134

101

20,181,104

null

0

null

0

a9bd4f08ac9b

2018-03-01

2018-03-01 22:38:56

2018-03-05

2018-03-05 19:20:53

3

false

en

2018-03-05

2018-03-05 19:20:53

0

1ba11f51ee07

2.957547

0

Engaging customers is becoming increasingly more difficult with the multitude of daily stimulus and distractions. So how can you leverage…

1

Using Multiple Modalities for brand conversion Engaging customers is becoming increasingly more difficult with the multitude of daily stimulus and distractions. So how can you leverage new modalities to engage customers and drive them to conversion? Conversation mapping in Say Spring. Leveraging Voice User Interface Alexa, Google Home and now, Apple’s HomePod are becoming more ubiquitous even though they’re still in their nascent stages. Most of us communicate to one another using our voices every day, yet voice communication is still a complicated task for machines. The first manifestation of voice was Interactive Voice Response or IVP. We are all familiar with IVP—it’s the prompt style interface which most phone automation was built on. Other industries that have used and still use them are banks, insurance companies, airline menus, etc. They’re terrible. They are long, essentially non-interactive and frustrating. However, everything needs to start somewhere. Voice interfacing is now moving away from IVP and toward NLP framework or natural language processing. NLP is far closer to how we communicate naturally rather than IVP. However two of the main challenges that are still associated with voice interfacing is recognization and understanding. These can also be distilled down to context. Contextual input is one of the most complex parts of the human language. A ‘pop’ in one part of the country means ‘soda’ in another part. Tony Sheeder, a senior experience designer at Nuance Communication has described design for voice as: Each voice interaction is a little narrative experience, with a begin‐ ning, middle and an end. Humans just get this and understand the rules naturally — some more than others. When you go to a party, you can tell within a very short time whether another person is easy to talk to. Until recently, speech systems were that guy at the party doing everything wrong, and nobody wanted to talk to them. Individual conversations for ticket purchasing in the NHL. Voice communication has a lot of benefits. For example, I am working on integrating VUI into the NHLs brand. Here, voice allows customers to instantly query the brand about current standings, player statistics, injury reports, set game alerts, have Alexa(or Google Home) instantly turn the TV on and to the proper channel when a game is on or, for purchasing tickets. The last point, purchasing tickets, currently poses a problem for voice. If you are a die-hard fan, someone who has been to a lot of games and are familiar with the arena, buying a ticket to the game using only voice may not present you with that large of a challenge. However, if you are a fan, but rarely go to games, purchasing a ticket becomes a much larger purchase consideration and this is where an additional modality comes into play(pun intended). Multi Modal conversions. One large issue with VUI is that it’s new. How do you instill trust in a brand? Referring back to the NHL and buying tickets, a lot of customers still like to have visual cues and confirmations. Some people may be comfortable just saying, ‘Hey Alexa, I want to buy two tickets to the next Bruins home game,’. A lot of people still want to see the ticketing screen. Cue the chatbots. Chatbots In conjunction with a voice interface, chat bots allow you to have a CMS that deploys content types once and deliver them via voice or chat. For our issue with NHL ticket purchasing, customers who are not comfortable with voice can jump on NHL.com and have the option to use a chatbot. Here the chatbot still leverages instant querying of the ticketing system without having to navigate through the website. It’s still a pleasant experience and allows for customers to easily convert. NHL chatbot In the future, the ability to have a natural conversation with a machine will be engrained in our daily lives and soon could potentially become the default interface preference.

Using Multiple Modalities for brand conversion

0

using-multiple-modalities-for-brand-conversion-1ba11f51ee07

2018-03-05

2018-03-05 19:20:54

https://medium.com/s/story/using-multiple-modalities-for-brand-conversion-1ba11f51ee07

false

638

Thoughts and stories from Studio, a one year product design masters program at CU Boulder, dedicated to re:working, re:designing and re:imagining the world of design and technology.

null

RE: Write

null

re-write

DESIGN,DIGITAL,TECHNOLOGY,PRODUCT DESIGN,UX

cmci_studio

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Ty Foster

Aspiring UX Designer. BDW masters student at CU Boulder. Photographer. www.tyfoster.com

348c8d6168c3

tyfoster

55

96

20,181,104

null

0

null

0

244ef586c71e

2017-12-12

2017-12-12 00:42:41

2017-12-12

2017-12-12 00:43:15

2

false

en

2018-09-05

2018-09-05 10:02:30

0

1ba2423e2e2c

6.813522

8

0

Ronny Fehling is one of the most energetic tech leaders in Artificial Intelligence. Now an Associate Director of GAMMA Artificial…

5

Meet RONNY FEHLING, one of the most energetic tech leaders in #AI Ronny Fehling Ronny Fehling is one of the most energetic tech leaders in Artificial Intelligence. Now an Associate Director of GAMMA Artificial Intelligence at Boston Consulting Group, he is former head of Cognitive Computing and AI at Airbus, and has also held leadership roles at Oracle and startup Cognito.co. And his time at Airbus was spent making company-wide transformations using data science and AI. He believes AI is now on the cusp of industrial capability and, despite the hype, can now be utilised at scale across certain settings. I caught up with him to pick his brains about the state of the sector now, and what the future holds. 1. Where does your passion for AI come from? Growing up in Tanzania, I didn’t have much exposure to high tech. But a Commodore C64 sparked something in my brother and me and we were soon combining Technic lego with the technology that controlled my car. I went on to read Maths, but it was a course called Complexity Theory in Computer Science that hooked me. I had the honour of studying at graduate level under the great Marvin Minsky, which confirmed my passion for AI. It was in 2004, while working on a project for Nasa using radio frequency identification and sensor computing, that I realised there was this new kind of data that needed to be contextualised before it could be interpreted. This data gave way to entirely new architectures like the NoSQL movement. Then, in 2011, deep-learning breakthroughs gave rise to the AI Spring — which I’ve been watching, and participating in, with absolute fascination. DEFINING AI 2. You’re known for being critical of the term AI, and (in)famously said being in this sector is “to work in a profession in a soup of terms with no meaning, yet nobody wants to point this out.”. How should we define AI? Haha. There is a hype around the term AI. While there is no single definition, it is often described as the attempt to create intelligent machines for learning, problem-solving and decision-making. While this sounds intriguing, I find this formulation more useful: “AI enables a new interaction model for exploring complex and large data sources (structured and unstructured, text, images, video, speech, audio, sensors) with conflicting answers, ambiguous evidences, and hard-to-automate processes. The ability to learn from data makes AI powerful. No more hardcoding every single possible behaviour; a cognitive system learns and improves with data (and outcomes). We want to create algorithms that can learn, adapt, interact and understand to carry out tasks in a way that we would consider ‘smart’.” Compared to more traditional machine learning, I think AI enables us to abstract concepts from raw data that lie beyond just the data itself. For example, interpreting transactional data to deduce intention or sentiment. Together with a strong outcome-based model interacting with the real world, we will be able to create vicarious forms of machine intelligence that can not only help us with the current task we are attempting, but also propose, based on other experiences, other methods that might get us to where we need to go faster. MAKING AI RELEVANT NOW 3. A lot of future-gazing goes on in the AI space. What is the key to applying the technology successfully today, and how do we do so in zero-fault-tolerant areas like the aviation industry? AI is currently in a very narrow, specialised state, which means it’s only intelligent in the defined domain boundaries given to it. But it can help us analyse concepts above the data — concepts which define the ‘knowledge’ humans will typically apply to data. What we are trying to do currently is to start encoding such knowledge into machines. Different to the rule-based systems in the past, which proved too complex to build and maintain, we now try to let the system learn these concepts by itself, and through external data entities. We can, for example, train a machine to learn about concepts of technical domains by having it do entity extractions of Wikipedia articles. Armed with that, it can now recognise these entities in enterprise data and put them into context. The result is a more flexible, less deterministic knowledge representation. AI is also very good at finding correlations and similarities in highly complex, multidimensional problems — systems where outcomes are determined by an interplay of hundreds of variables so complex that traditional, deterministic model-based systems tend to break down. But AI can only be as good as the data it’s fed. If the data cannot be correlated to the observed outcome, AI systems cannot reliably be used in mission or life-critical systems. The certification of AI systems in operational settings is still little understood since, typically, machines have been certified as deterministic systems, with humans covering the remainder. I expect that, while they will increasingly develop ‘explain’ functions, certification requirements for AI systems will also evolve. AI mines data in such a way that it can find anomalies or safety issues that otherwise might have been overlooked. And it is very good at recognising what are commonly referred to as false positives, i.e. it might flag an event as potentially critical even though it isn’t. People are much better at discriminating critical events, so these are cases where humans and AI can work well together. Panel: AI in Healthcare and Other Applications with Ronny Fehling 4. Tell us about cleaning databases and making information more algorithm-friendly. Are those the biggest challenges when it comes to putting AI into production over the coming years? AI cannot work well with dirty data, and cleansing data is currently 70–85% of a data scientist’s work. For an AI strategy to be effective, it’s important to not only have clean data, but to have functioning, scalable big data implementation. Having said that, I strongly believe that AI can actually help with data cleansing. When humans cleanse data, we try to match lexicographically or structurally different data sets, using general and domain-specific knowledge. But AI can learn about the data. It can, for example, recognise in a particular data set of social security numbers, a wrong value or type. AI can already detect correlated fields and use that for error-checking. Moreover, data cleansing is generally not deterministic, and has little transparency. AI can not only give more transparency to the cleansing operation (you can tune the confidence threshold), but it can also help annotate and tag the data for upstream processing. ETHICS AND THE FUTURE 5. What are the ethical challenges for executives and founders applying AI? I think it’s important here to distinguish between the current state of AI and what some term artificial general intelligence (AGI), which, in essence, refers to completely self-learning general purpose machines. Putting AGI aside, I’ll focus on what I think is important for AI right now. Of course you should not use AI for bad things, but that just seems obvious and is no different to using algorithms to do bad things. But we have to understand the potential danger of biased data. Ultimately, AI has to be trained by examples, so an AI system not only depends on the data it is fed, but also biases in that data. If a significant portion of the data or outcomes you feed the AI is biased, faulty or fake, the AI will be faulty (think Microsoft’s Tay). This is where the ethical debate has to be: less on the AI itself, but rather on the (selective) power of data bias. If AI systems are used in mission or even life-critical systems, we need to make sure that the data we feed it is representative and, as far as possible, bias-free. Today, AI can already help us develop recommendation systems that help reduce the cognitive workload of human workers by providing them with the right information at the right time and in the right context. But as the systems become more complex and the interactions between the various data points harder to examine, the danger of biased data arises. And if an AI system displays a recommendation to the human but the human cannot detect an eventual slight bias in the data, he might, by following the recommendation, inadvertently reinforce the bias in the machine. We already see examples of data biases being purposefully introduced and reinforced by intelligent algorithms. Fake news, for instance, is seeded in various forums through bots, then picked up by indexing algorithms from Facebook, Google, Twitter, etc. As they bombard those algorithms with fake accounts, supposedly sharing these stories over and over, they rise on indexes and find an ever-growing audience — continuously reinforcing the bias towards such stories. This problem is reflected in the larger context of anthropomorphisation of AI — i.e. attributing human emotions to the AI, which will further contribute further to a data bias. We need to constantly be on the guard for biasing effects, and must attempt to design systems that can counter, or at least alert us to, such outcomes. Ronny Fehling at WorldSummit.AI 2017 in Amsterdam 6. In the next 18 months, how can we all work to ensure that AI’s impact on business and humanity is a positive one? No matter what we do, AI will continue to evolve. And it will have a profound impact on our lives. Governments must invest in this technology, enable its development, and invest in the education of young people, as well as the existing workforce, so that we can learn how to deal with the societal changes AI will prompt in the future. AI will challenge a lot of existing jobs and an education system that has endured for centuries. We will have to accept that change is going to be so fast that, rather than simply completing a degree, we will have to keep up with new information and developments throughout our entire working life. The looming risk is that the gap that already exists in education will widen, causing more friction in society. We must strive towards reducing that inequality, by offering education in these developments to all employees and pupils so that they can adapt and prepare.

Meet RONNY FEHLING, one of the most energetic tech leaders in #AI

48

meet-ronny-fehling-one-of-the-most-energetic-tech-leaders-in-ai-1ba2423e2e2c

2018-09-05

2018-09-05 10:02:30

https://medium.com/s/story/meet-ronny-fehling-one-of-the-most-energetic-tech-leaders-in-ai-1ba2423e2e2c

false

1,704

Making knowledge on #appliedAI accessible

null

cityai

null

Applied Artificial Intelligence

hello@city.ai

cityai

ARTIFICIAL INTELLIGENCE,MACHINE LEARNING,DEEP LEARNING,COMPUTER SCIENCE,NATURALLANGUAGEPROCESSING

thecityai

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Christoph Auer-Welsbach

Partner @IBM Ventures | Founding Director @TheCityAI | CoFounder @WorldSummitAI | #appliedAI #AI4Good | #LinkedInTopVoices 2017

3805df6657ee

AWChristoph

876

566

20,181,104

null

0

library(tidyverse) library(rvest) library(xml2) library(scales) library(ggrepel) download_html(url = "https://www.mtggoldfish.com/index/DOM", file = "goldfish_DOM.html") descargar_set <- function(clave) { mtg_url <- paste0("https://www.mtggoldfish.com/index/", clave) mtg_archivo <- paste0("goldfish_", clave, ".html") download_html(url = mtg_url, file = mtg_archivo) } mtg_dom <- read_html("goldfish_DOM.html") mtg_dom ## {xml_document} ## <html xmlns="http://www.w3.org/1999/xhtml"> ## [1] <head>\n<title>Dominaria MTG / MTGO Price History</title>\n<meta na ... ## [2] <body>\n<img alt="MTGGoldfish" class="layout-print-logo" src="//ass ... ## [3] <div class="container-fluid layout-container-fluid">\n<div id="erro ... ## [4] <div class="layout-bottom-ad">\n\n <div class="ads-container-cdm-z ... ## [5] <div class="layout-bottom-banner">\n<div class="layout-bottom-conte ... ## [6] <div class="bottom-shelf">\n<div class="banner-contents">\n<p class ... ## [7] <div aria-hidden="true" aria-labelledby="Login Dialog" class="modal ... ## [8] <div aria-hidden="true" aria-labelledby="Important Updates" class=" ... ## [9] <div aria-hidden="true" aria-labelledby="Card Popup" class="modal f ... ## [10] <div class="layout-typePreferencePopup" id="type-preference-popup"> ... ## [11] <script src="//assets1.mtggoldfish.com/assets/application-fef731ca0 ... ## [12] <script src="//assets1.mtggoldfish.com/assets/google_analytics-b937 ... ## [13] <div id="cdm-zone-end"></div>\n ## [14] <script>\nvar _comscore = _comscore || [];\n_comscore.push({ c1: "2 ... ## [15] <noscript>\n <img src="//b.scorecardresearch.com/p?c1=2&c2=60 ... ## [16] <script type="text/javascript">\n(function () {\n var d = new Ima ... ## [17] <noscript>\n <div><img src="//secure-us.imrworldwide.com ... mtg_dom %>% html_node(css = ".index-price-table-paper tbody tr") ## {xml_node} ## <tr> ## [1] <td class="card"><a data-full-image="https://cdn1.mtggoldfish.com/im ... ## [2] <td>DOM</td> ## [3] <td>Mythic</td> ## [4] <td class="text-right">\n46.99\n</td> ## [5] <td class="text-right">\n<div class="common-price-change">\n-1.89\n\ ... ## [6] <td class="text-right">-4.00%</td> ## [7] <td class="text-right">\n<div class="common-price-change">\n-0.13\n\ ... ## [8] <td class="text-right">0.00%</td> mtg_dom %>% html_nodes(css = ".index-price-table-paper tbody tr") %>% html_text() %>% head() ## [1] "Teferi, Hero of Dominaria\nDOM\nMythic\n\n46.99\n\n\n\n-1.89\n\n\n\n\n\n-4.00%\n\n\n-0.13\n\n\n\n\n\n0.00%\n" ## [2] "Karn, Scion of Urza\nDOM\nMythic\n\n31.54\n\n\n\n-0.14\n\n\n\n\n\n0.00%\n\n\n-0.14\n\n\n\n\n\n0.00%\n" ## [3] "Lyra Dawnbringer\nDOM\nMythic\n\n13.00\n\n\n\n+0.01\n\n\n\n\n\n0.00%\n\n\n+0.04\n\n\n\n\n\n0.00%\n" ## [4] "Mox Amber\nDOM\nMythic\n\n11.42\n\n\n\n-0.06\n\n\n\n\n\n-1.00%\n\n\n-0.09\n\n\n\n\n\n-1.00%\n" ## [5] "History of Benalia\nDOM\nMythic\n\n10.60\n\n\n\n-0.03\n\n\n\n\n\n0.00%\n\n\n+0.05\n\n\n\n\n\n0.00%\n" ## [6] "Sulfur Falls\nDOM\nRare\n\n4.62\n\n\n\n+0.03\n\n\n\n\n\n+1.00%\n\n\n+0.10\n\n\n\n\n\n+2.00%\n" df_dom <- mtg_dom %>% html_nodes(css = ".index-price-table-paper tbody tr") %>% html_text() %>% str_split(pattern = "\\n", simplify = T) %>% as.data.frame() %>% tbl_df() df_dom ## # A tibble: 277 x 25 ## V1 V2 V3 V4 V5 V6 V7 V8 V9 V10 V11 V12 ## <fct> <fct> <fct> <fct> <fct> <fct> <fct> <fct> <fct> <fct> <fct> <fct> ## 1 Tefe~ DOM Myth~ "" 46.99 "" "" "" -1.89 "" "" "" ## 2 Karn~ DOM Myth~ "" 31.54 "" "" "" -0.14 "" "" "" ## 3 Lyra~ DOM Myth~ "" 13.00 "" "" "" +0.01 "" "" "" ## 4 Mox ~ DOM Myth~ "" 11.42 "" "" "" -0.06 "" "" "" ## 5 Hist~ DOM Myth~ "" 10.60 "" "" "" -0.03 "" "" "" ## 6 Sulf~ DOM Rare "" 4.62 "" "" "" +0.03 "" "" "" ## 7 Jaya~ DOM Myth~ "" 4.47 "" "" "" 0.00 "" "" "" ## 8 Gobl~ DOM Rare "" 4.20 "" "" "" -0.02 "" "" "" ## 9 Wood~ DOM Rare "" 3.93 "" "" "" +0.07 "" "" "" ## 10 Stee~ DOM Rare "" 3.82 "" "" "" -0.10 "" "" "" ## # ... with 267 more rows, and 13 more variables: V13 <fct>, V14 <fct>, ## # V15 <fct>, V16 <fct>, V17 <fct>, V18 <fct>, V19 <fct>, V20 <fct>, ## # V21 <fct>, V22 <fct>, V23 <fct>, V24 <fct>, V25 <fct> df_dom <- df_dom %>% select("Carta" = V1, "Set" = V2, "Rareza"=V3, "Precio" = V5) %>% mutate(Precio = as.numeric(as.character(Precio)), Rareza = factor(Rareza, levels = c("Basic Land","Common", "Uncommon", "Rare", "Mythic"))) %>% mutate_at(c("Carta", "Set"), as.character) %>% filter(Rareza != "Basic Land") df_dom ## # A tibble: 257 x 4 ## Carta Set Rareza Precio ## <chr> <chr> <fct> <dbl> ## 1 Teferi, Hero of Dominaria DOM Mythic 47.0 ## 2 Karn, Scion of Urza DOM Mythic 31.5 ## 3 Lyra Dawnbringer DOM Mythic 13 ## 4 Mox Amber DOM Mythic 11.4 ## 5 History of Benalia DOM Mythic 10.6 ## 6 Sulfur Falls DOM Rare 4.62 ## 7 Jaya Ballard DOM Mythic 4.47 ## 8 Goblin Chainwhirler DOM Rare 4.2 ## 9 Woodland Cemetery DOM Rare 3.93 ## 10 Steel Leaf Champion DOM Rare 3.82 ## # ... with 247 more rows leer_html <- function(archivo_html) { archivo_html %>% read_html() %>% html_nodes(css = ".index-price-table-paper tbody tr") %>% html_text() %>% str_split(pattern = "\\n", simplify = T) %>% as.data.frame() %>% select("Carta" = V1, "Set" = V2, "Rareza"=V3, "Precio"=V5) %>% mutate(Precio = as.numeric(as.character(Precio)), Rareza = factor(Rareza, levels = c("Basic Land","Common", "Uncommon", "Rare", "Mythic"))) %>% mutate_at(c("Carta", "Set"), as.character) %>% filter(Rareza != "Basic Land") %>% tbl_df() } tag_outlier <- function(datos) { ifelse(datos > quantile(datos, .75) + IQR(datos) * 1.5, TRUE, FALSE) } df_dom <- df_dom %>% group_by(Rareza) %>% mutate(Outlier = tag_outlier(Precio)) %>% mutate(Outlier = ifelse(Outlier, Carta, NA)) df_dom ## # A tibble: 257 x 5 ## # Groups: Rareza [4] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Teferi, Hero of Dominaria DOM Mythic 47.0 Teferi, Hero of Dominaria ## 2 Karn, Scion of Urza DOM Mythic 31.5 Karn, Scion of Urza ## 3 Lyra Dawnbringer DOM Mythic 13 <NA> ## 4 Mox Amber DOM Mythic 11.4 <NA> ## 5 History of Benalia DOM Mythic 10.6 <NA> ## 6 Sulfur Falls DOM Rare 4.62 Sulfur Falls ## 7 Jaya Ballard DOM Mythic 4.47 <NA> ## 8 Goblin Chainwhirler DOM Rare 4.2 Goblin Chainwhirler ## 9 Woodland Cemetery DOM Rare 3.93 Woodland Cemetery ## 10 Steel Leaf Champion DOM Rare 3.82 Steel Leaf Champion ## # ... with 247 more rows etiquetar_outlier <- function(mtg_df) { mtg_df %>% group_by(Rareza) %>% mutate(Outlier = tag_outlier(Precio)) %>% mutate(Outlier = ifelse(Outlier, Carta, NA)) } df_dom %>% group_by(Rareza) %>% summarize(Media = mean(Precio), Mediana = median(Precio), Minimo = min(Precio), Maximo = max(Precio)) ## # A tibble: 4 x 5 ## Rareza Media Mediana Minimo Maximo ## <fct> <dbl> <dbl> <dbl> <dbl> ## 1 Common 0.151 0.14 0.11 0.88 ## 2 Uncommon 0.309 0.23 0.14 1.73 ## 3 Rare 1.03 0.53 0.25 4.62 ## 4 Mythic 9.01 2.86 1.13 47.0 df_dom %>% ggplot() + aes(Precio) + geom_density() df_dom %>% ggplot() + aes(Precio, fill = Rareza) + geom_density() + facet_wrap(~Rareza, scales = "free") df_dom %>% ggplot() + aes(Rareza, Precio, fill = Rareza) + geom_boxplot() + geom_label_repel(aes(label = Outlier, color = Rareza), size = 2.5, fill = "white") + theme(legend.position = "none") ## Warning: Removed 228 rows containing missing values (geom_label_repel). explorar_set <- function(mtg_df) { exploracion <- list() exploracion$precios_resumen <- mtg_df %>% group_by(Rareza) %>% summarize(Media = mean(Precio), Mediana = median(Precio), Minimo = min(Precio), Maximo = max(Precio)) %>% mutate_if(is.numeric, ~round(., 2)) exploracion$precios_rareza <- mtg_df %>% ggplot() + aes(Precio, fill = Rareza) + geom_density() + facet_wrap(~Rareza, scales = "free") + scale_x_continuous(labels = dollar_format()) + labs(y = "Densidad") + theme_minimal() + theme(legend.position = "none") exploracion$boxplot <- mtg_df %>% ggplot() + aes(Rareza, Precio, fill = Rareza) + geom_boxplot() + geom_label_repel(aes(label = Outlier, color = Rareza), size = 2.5, fill = "white") + scale_y_continuous(labels = dollar_format()) + theme_minimal() + theme(legend.position = "none") exploracion } rareza_frecuencia <- list( list("Rare", 1), list("Uncommon", 3), list("Common", 10) ) set.seed(2018) map(rareza_frecuencia, function(pareja){ df_dom %>% filter(Rareza == pareja[[1]]) %>% sample_n(size = pareja[[2]]) }) ## [[1]] ## # A tibble: 1 x 5 ## # Groups: Rareza [1] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Traxos, Scourge of Kroog DOM Rare 0.66 <NA> ## ## [[2]] ## # A tibble: 3 x 5 ## # Groups: Rareza [1] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Kwende, Pride of Femeref DOM Uncommon 0.24 <NA> ## 2 Merfolk Trickster DOM Uncommon 0.76 Merfolk Trickster ## 3 Dauntless Bodyguard DOM Uncommon 0.35 <NA> ## ## [[3]] ## # A tibble: 10 x 5 ## # Groups: Rareza [1] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Thallid Omnivore DOM Common 0.14 <NA> ## 2 Healing Grace DOM Common 0.15 <NA> ## 3 Frenzied Rage DOM Common 0.14 <NA> ## 4 Artificer's Assistant DOM Common 0.15 <NA> ## 5 Ghitu Chronicler DOM Common 0.13 <NA> ## 6 Deathbloom Thallid DOM Common 0.14 <NA> ## 7 Adamant Will DOM Common 0.14 <NA> ## 8 Run Amok DOM Common 0.14 <NA> ## 9 Keldon Raider DOM Common 0.13 <NA> ## 10 Seismic Shift DOM Common 0.13 <NA> pareja <- function(lista, datos) { datos %>% filter(Rareza == lista[1]) %>% sample_n(size = as.numeric(lista[2])) } set.seed(2018) map(rareza_frecuencia, pareja, datos = df_dom) ## [[1]] ## # A tibble: 1 x 5 ## # Groups: Rareza [1] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Traxos, Scourge of Kroog DOM Rare 0.66 <NA> ## ## [[2]] ## # A tibble: 3 x 5 ## # Groups: Rareza [1] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Kwende, Pride of Femeref DOM Uncommon 0.24 <NA> ## 2 Merfolk Trickster DOM Uncommon 0.76 Merfolk Trickster ## 3 Dauntless Bodyguard DOM Uncommon 0.35 <NA> ## ## [[3]] ## # A tibble: 10 x 5 ## # Groups: Rareza [1] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Thallid Omnivore DOM Common 0.14 <NA> ## 2 Healing Grace DOM Common 0.15 <NA> ## 3 Frenzied Rage DOM Common 0.14 <NA> ## 4 Artificer's Assistant DOM Common 0.15 <NA> ## 5 Ghitu Chronicler DOM Common 0.13 <NA> ## 6 Deathbloom Thallid DOM Common 0.14 <NA> ## 7 Adamant Will DOM Common 0.14 <NA> ## 8 Run Amok DOM Common 0.14 <NA> ## 9 Keldon Raider DOM Common 0.13 <NA> ## 10 Seismic Shift DOM Common 0.13 <NA> set.seed(8102) map(rareza_frecuencia, pareja, datos = df_dom) %>% reduce(bind_rows) ## # A tibble: 14 x 5 ## # Groups: Rareza [3] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Squee, the Immortal DOM Rare 0.62 <NA> ## 2 Firefist Adept DOM Uncommon 0.2 <NA> ## 3 Urza's Tome DOM Uncommon 0.19 <NA> ## 4 Orcish Vandal DOM Uncommon 0.19 <NA> ## 5 Grow from the Ashes DOM Common 0.15 <NA> ## 6 Short Sword DOM Common 0.13 <NA> ## 7 Warlord's Fury DOM Common 0.15 <NA> ## 8 Opt DOM Common 0.25 Opt ## 9 Cabal Evangel DOM Common 0.13 <NA> ## 10 Timber Gorge DOM Common 0.11 <NA> ## 11 Charge DOM Common 0.14 <NA> ## 12 Gideon's Reproach DOM Common 0.15 <NA> ## 13 Llanowar Envoy DOM Common 0.14 <NA> ## 14 Krosan Druid DOM Common 0.13 <NA> rbinom(n = 40, size = 1, prob = 1/8) ## [1] 0 0 1 0 0 0 1 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 ## [36] 0 0 0 0 0 set.seed(6) rareza_frecuencia %>% map(function(x) { if(x[[1]] == "Rare" & rbinom(n = 1, size = 1, prob = 1/8)) { "Mythic" } else { x[[1]] } }) ## [[1]] ## [1] "Rare" ## ## [[2]] ## [1] "Uncommon" ## ## [[3]] ## [1] "Common" set.seed(7) rareza_frecuencia %>% map(function(x) { if(x[[1]] == "Rare" & rbinom(n = 1, size = 1, prob = 1/8)) { "Mythic" } else { x[[1]] } }) ## [[1]] ## [1] "Mythic" ## ## [[2]] ## [1] "Uncommon" ## ## [[3]] ## [1] "Common" simular_sobre <- function(tabla) { rareza_frecuencia <- list( c("Rare", 1), c("Uncommon", 3), c("Common", 10) ) rareza_frecuencia <- map(rareza_frecuencia, function(x) { if(x[[1]] == "Rare" & rbinom(n = 1, size = 1, prob = 1/8)) { x[[1]] <- "Mythic" } else { x[[1]] } x }) map(rareza_frecuencia, pareja, datos = tabla) %>% reduce(bind_rows) } set.seed(7) simular_sobre(tabla = df_dom) ## # A tibble: 14 x 5 ## # Groups: Rareza [3] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Karn, Scion of Urza DOM Mythic 31.5 Karn, Scion of Urza ## 2 Fight with Fire DOM Uncommon 0.3 <NA> ## 3 Amaranthine Wall DOM Uncommon 0.2 <NA> ## 4 Settle the Score DOM Uncommon 0.25 <NA> ## 5 Blessed Light DOM Common 0.13 <NA> ## 6 Temporal Machinations DOM Common 0.15 <NA> ## 7 Knight of New Benalia DOM Common 0.14 <NA> ## 8 Fiery Intervention DOM Common 0.12 <NA> ## 9 Sparring Construct DOM Common 0.15 <NA> ## 10 Mesa Unicorn DOM Common 0.14 <NA> ## 11 Corrosive Ooze DOM Common 0.15 <NA> ## 12 Pierce the Sky DOM Common 0.14 <NA> ## 13 Deep Freeze DOM Common 0.15 <NA> ## 14 Blessing of Belzenlok DOM Common 0.14 <NA> set.seed(8244) caja_dom <- map(1:36, ~simular_sobre(tabla = df_dom)) caja_dom[[4]] ## # A tibble: 14 x 5 ## # Groups: Rareza [3] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Primevals' Glorious Rebirth DOM Rare 0.39 <NA> ## 2 Goblin Barrage DOM Uncommon 0.21 <NA> ## 3 Teferi's Sentinel DOM Uncommon 0.2 <NA> ## 4 Final Parting DOM Uncommon 0.25 <NA> ## 5 Deathbloom Thallid DOM Common 0.14 <NA> ## 6 Dub DOM Common 0.14 <NA> ## 7 Gaea's Protector DOM Common 0.14 <NA> ## 8 Rampaging Cyclops DOM Common 0.14 <NA> ## 9 Radiating Lightning DOM Common 0.14 <NA> ## 10 Cabal Paladin DOM Common 0.14 <NA> ## 11 Adventurous Impulse DOM Common 0.18 Adventurous Impulse ## 12 Arbor Armament DOM Common 0.14 <NA> ## 13 Seismic Shift DOM Common 0.13 <NA> ## 14 Sergeant-at-Arms DOM Common 0.15 <NA> simular_caja <- function(datos) { map(1:36, ~simular_sobre(tabla = df_dom)) %>% reduce(bind_rows) } set.seed(8244) caja_dom <- simular_caja(datos = df_fom) caja_dom %>% filter(Rareza %in% c("Rare", "Mythic")) ## # A tibble: 36 x 5 ## # Groups: Rareza [2] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Oath of Teferi DOM Rare 0.59 <NA> ## 2 Jodah, Archmage Eternal DOM Rare 0.46 <NA> ## 3 Naban, Dean of Iteration DOM Rare 0.59 <NA> ## 4 Primevals' Glorious Rebirth DOM Rare 0.39 <NA> ## 5 Shalai, Voice of Plenty DOM Rare 2.84 Shalai, Voice of Plenty ## 6 Squee, the Immortal DOM Rare 0.62 <NA> ## 7 The Mending of Dominaria DOM Rare 0.49 <NA> ## 8 Cabal Stronghold DOM Rare 1 <NA> ## 9 Karn, Scion of Urza DOM Mythic 31.5 Karn, Scion of Urza ## 10 Primevals' Glorious Rebirth DOM Rare 0.39 <NA> ## # ... with 26 more rows set.seed(3356) ciencajas_dom <- map(1:100, function(x) { simular_caja(datos = df_dom) %>% summarize(Valor = sum(Precio)) }) %>% reduce(bind_rows) ciencajas_dom ## # A tibble: 398 x 2 ## Rareza Valor ## <fct> <dbl> ## 1 Common 53.0 ## 2 Uncommon 35.7 ## 3 Rare 29.1 ## 4 Mythic 48.4 ## 5 Common 54.5 ## 6 Uncommon 34.4 ## 7 Rare 28.6 ## 8 Mythic 86.4 ## 9 Common 52.7 ## 10 Uncommon 36.4 ## # ... with 388 more rows ciencajas_dom %>% group_by(Rareza) %>% summarize(Promedio = mean(Valor)) ## # A tibble: 4 x 2 ## Rareza Promedio ## <fct> <dbl> ## 1 Common 54.3 ## 2 Uncommon 33.5 ## 3 Rare 31.8 ## 4 Mythic 42.0 simular_ciencajas <- function(datos, iteraciones = 100) { map(1:iteraciones, function(num_caja) { simular_caja(datos = datos) %>% summarize(Valor = sum(Precio)) %>% mutate(Id_caja = num_caja) }) %>% reduce(bind_rows) } set.seed(3356) ciencajas_dom <- simular_ciencajas(datos = df_dom) ciencajas_dom ## # A tibble: 398 x 3 ## Rareza Valor Id_caja ## <fct> <dbl> <int> ## 1 Common 53.0 1 ## 2 Uncommon 35.7 1 ## 3 Rare 29.1 1 ## 4 Mythic 48.4 1 ## 5 Common 54.5 2 ## 6 Uncommon 34.4 2 ## 7 Rare 28.6 2 ## 8 Mythic 86.4 2 ## 9 Common 52.7 3 ## 10 Uncommon 36.4 3 ## # ... with 388 more rows ciencajas_dom %>% group_by(Id_caja) %>% summarize(Suma = sum(Valor)) %>% summarize(Media = mean(Suma)) ## # A tibble: 1 x 1 ## Media ## <dbl> ## 1 161. simular_ciencajas <- function(datos, iteraciones = 100) { map(1:iteraciones, function(num_caja) { simular_caja(datos = datos) %>% filter(Rareza != "Common" | (Rareza == "Common" & !is.na(Outlier)) ) %>% summarize(Valor = sum(Precio)) %>% mutate(Id_caja = num_caja) }) %>% reduce(bind_rows) } set.seed(3356) ciencajas_dom <- simular_ciencajas(datos = df_dom) ciencajas_dom ## # A tibble: 398 x 3 ## Rareza Valor Id_caja ## <fct> <dbl> <int> ## 1 Common 6.42 1 ## 2 Uncommon 35.7 1 ## 3 Rare 29.1 1 ## 4 Mythic 48.4 1 ## 5 Common 7.1 2 ## 6 Uncommon 34.4 2 ## 7 Rare 28.6 2 ## 8 Mythic 86.4 2 ## 9 Common 5.04 3 ## 10 Uncommon 36.4 3 ## # ... with 388 more rows ciencajas_dom %>% group_by(Id_caja) %>% summarize(Suma = sum(Valor)) %>% summarize(Media = mean(Suma), Mediana = median(Suma)) ## # A tibble: 1 x 2 ## Media Mediana ## <dbl> <dbl> ## 1 114. 106. valorcajas_dom <- ciencajas_dom %>% group_by(Id_caja) %>% summarize(Suma = sum(Valor)) %>% pull(Suma) valorcajas_dom ## [1] 119.57 156.47 135.87 129.77 145.01 92.61 83.36 81.80 83.36 94.87 ## [11] 116.89 81.52 146.44 89.45 124.25 155.24 97.28 140.66 83.01 144.91 ## [21] 137.17 166.77 124.05 75.35 77.33 80.35 84.65 153.38 94.10 117.94 ## [31] 114.88 80.99 144.39 109.78 98.81 100.64 158.83 96.26 88.66 139.78 ## [41] 155.45 88.49 75.97 150.07 105.70 87.75 176.57 81.47 116.65 123.54 ## [51] 93.45 131.73 136.54 115.67 87.20 223.06 86.77 106.12 76.46 144.24 ## [61] 104.59 102.96 121.24 87.94 117.09 84.11 130.13 101.38 161.11 87.63 ## [71] 108.55 151.34 95.74 123.01 182.26 104.25 94.69 104.37 133.66 139.48 ## [81] 140.47 70.70 97.20 74.48 72.47 94.24 88.38 160.94 91.93 88.48 ## [91] 111.76 149.46 137.43 82.05 78.22 195.35 82.63 134.74 126.79 82.04 valor_cajas <- function(cajas_simuladas) { cajas_simuladas %>% group_by(Id_caja) %>% summarize(Suma = sum(Valor)) %>% pull(Suma) } densidad_dom <- density(valorcajas_dom) densidad_dom ## ## Call: ## density.default(x = valorcajas_dom) ## ## Data: valorcajas_dom (100 obs.); Bandwidth 'bw' = 11.16 ## ## x y ## Min. : 37.22 Min. :4.020e-06 ## 1st Qu.: 92.05 1st Qu.:3.674e-04 ## Median :146.88 Median :2.409e-03 ## Mean :146.88 Mean :4.555e-03 ## 3rd Qu.:201.71 3rd Qu.:8.499e-03 ## Max. :256.54 Max. :1.396e-02 densidad_dom[c("x", "y")] %>% tbl_df() %>% ggplot() + aes(x, y) + geom_area() densidad_dom[c("x", "y")] %>% tbl_df() %>% ggplot() + aes(x, y) + geom_area() + geom_vline(xintercept = 105, color = "red") densidad_dom %>% approxfun() %>% integrate(upper = max(valorcajas_dom), lower = 100) ## 0.5875856 with absolute error < 3e-05 magic_densidad <- function(valor_cajas, costo_pagado) { magic <- list() magic$costo_pagado <- costo_pagado magic$densidad <- density(valor_cajas) magic$probabilidad <- magic$densidad %>% approxfun() %>% integrate(upper = max(valor_cajas), lower = costo_pagado) magic$df_densidad <- magic$densidad[c("x", "y")] %>% tbl_df() %>% mutate(Tipo = ifelse(x < costo_pagado, "Menor", "Mayor")) magic } magic_densidad(valorcajas_dom, costo_pagado = 100) ## $costo_pagado ## [1] 100 ## ## $densidad ## ## Call: ## density.default(x = valor_cajas) ## ## Data: valor_cajas (100 obs.); Bandwidth 'bw' = 11.16 ## ## x y ## Min. : 37.22 Min. :4.020e-06 ## 1st Qu.: 92.05 1st Qu.:3.674e-04 ## Median :146.88 Median :2.409e-03 ## Mean :146.88 Mean :4.555e-03 ## 3rd Qu.:201.71 3rd Qu.:8.499e-03 ## Max. :256.54 Max. :1.396e-02 ## ## $probabilidad ## 0.5875856 with absolute error < 3e-05 ## ## $df_densidad ## # A tibble: 512 x 3 ## x y Tipo ## <dbl> <dbl> <chr> ## 1 37.2 0.0000131 Menor ## 2 37.6 0.0000149 Menor ## 3 38.1 0.0000169 Menor ## 4 38.5 0.0000191 Menor ## 5 38.9 0.0000216 Menor ## 6 39.4 0.0000245 Menor ## 7 39.8 0.0000276 Menor ## 8 40.2 0.0000310 Menor ## 9 40.7 0.0000350 Menor ## 10 41.1 0.0000393 Menor ## # ... with 502 more rows plot_densidad <- function(lista_densidad) { label_costo <- paste0("Costo pagado: ", lista_densidad$costo_pagado, " USD") label_prob <- paste0("Probabilidad de recuperar inversión: ", round(lista_densidad$probabilidad$value, 4) * 100, "%") lista_densidad$df_densidad %>% ggplot() + aes(x, y, fill = Tipo) + geom_area() + labs(title = paste0(label_costo, "\n", label_prob)) + scale_y_continuous(expand = c(0, 0)) + scale_x_continuous(labels = dollar_format()) + labs(x = "USD", y = "Densidad") + theme_minimal() + theme(legend.position = "top") } magic_densidad(valorcajas_dom, 90) %>% plot_densidad() analisis_set <- function(set_html, costo_pagado = 100) { analisis <- list() analisis$df <- leer_html(set_html) %>% etiquetar_outlier() analisis$explorar <- explorar_set(analisis$df) analisis$simulacion <- simular_ciencajas(analisis$df) analisis$valor <- valor_cajas(analisis$simulacion) analisis$densidad <- magic_densidad(analisis$valor, costo_pagado = costo_pagado) analisis$inversion <- plot_densidad(analisis$densidad) analisis } descargar_set(clave = "XLN") set.seed(25) ixalan_lista <- analisis_set("goldfish_XLN.html", costo_pagado = 90) ixalan_lista ## $df ## # A tibble: 267 x 5 ## # Groups: Rareza [4] ## Carta Set Rareza Precio Outlier ## <chr> <chr> <fct> <dbl> <chr> ## 1 Search for Azcanta XLN Rare 21.0 Search for Azcanta ## 2 Carnage Tyrant XLN Mythic 18.9 Carnage Tyrant ## 3 Vraska's Contempt XLN Rare 14.5 Vraska's Contempt ## 4 Vraska, Relic Seeker XLN Mythic 8.77 Vraska, Relic Seeker ## 5 Settle the Wreckage XLN Rare 7.42 Settle the Wreckage ## 6 Growing Rites of Itlimoc XLN Rare 6 Growing Rites of Itlimoc ## 7 Glacial Fortress XLN Rare 4.99 <NA> ## 8 Drowned Catacomb XLN Rare 4.95 <NA> ## 9 Treasure Map XLN Rare 4.24 <NA> ## 10 Sunpetal Grove XLN Rare 4 <NA> ## # ... with 257 more rows ## ## $explorar ## $explorar$precios_resumen ## # A tibble: 4 x 5 ## Rareza Media Mediana Minimo Maximo ## <fct> <dbl> <dbl> <dbl> <dbl> ## 1 Common 0.15 0.14 0.12 0.31 ## 2 Uncommon 0.37 0.22 0.19 3.5 ## 3 Rare 2.03 0.75 0.32 21.0 ## 4 Mythic 3.42 1.75 0.72 18.9 ## ## $explorar$precios_rareza ## ## $explorar$boxplot ## Warning: Removed 244 rows containing missing values (geom_label_repel). ## ## ## $simulacion ## # A tibble: 397 x 3 ## Rareza Valor Id_caja ## <fct> <dbl> <int> ## 1 Common 5.16 1 ## 2 Uncommon 37.3 1 ## 3 Rare 35.2 1 ## 4 Mythic 97.3 1 ## 5 Common 7.08 2 ## 6 Uncommon 33.3 2 ## 7 Rare 29.3 2 ## 8 Mythic 54.7 2 ## 9 Common 6.29 3 ## 10 Uncommon 33.0 3 ## # ... with 387 more rows ## ## $valor ## [1] 174.90 124.32 90.38 65.45 90.07 76.19 164.54 114.02 69.08 170.50 ## [11] 150.64 112.79 137.96 99.70 159.40 130.86 98.39 154.86 119.92 216.09 ## [21] 111.08 79.53 129.72 125.33 147.36 133.00 86.57 77.84 131.81 85.03 ## [31] 170.15 137.96 171.46 96.32 124.70 180.97 83.42 133.85 95.41 94.55 ## [41] 93.14 97.12 109.09 110.51 111.38 98.32 95.89 127.21 89.81 150.29 ## [51] 100.50 89.55 76.39 128.75 82.65 97.03 72.37 132.23 108.64 73.42 ## [61] 90.52 133.59 80.30 93.76 148.57 80.40 159.60 164.89 145.37 93.74 ## [71] 193.68 92.32 119.95 90.02 130.82 165.24 81.66 103.21 95.23 115.61 ## [81] 78.99 164.18 123.06 108.66 202.00 190.49 73.20 83.96 116.63 71.87 ## [91] 147.48 85.19 84.68 78.20 89.46 88.15 118.36 71.03 118.95 112.71 ## ## $densidad ## $densidad$costo_pagado ## [1] 90 ## ## $densidad$densidad ## ## Call: ## density.default(x = valor_cajas) ## ## Data: valor_cajas (100 obs.); Bandwidth 'bw' = 11.8 ## ## x y ## Min. : 30.05 Min. :3.864e-06 ## 1st Qu.: 85.41 1st Qu.:6.273e-04 ## Median :140.77 Median :3.563e-03 ## Mean :140.77 Mean :4.511e-03 ## 3rd Qu.:196.13 3rd Qu.:8.196e-03 ## Max. :251.49 Max. :1.278e-02 ## ## $densidad$probabilidad ## 0.7141344 with absolute error < 6e-05 ## ## $densidad$df_densidad ## # A tibble: 512 x 3 ## x y Tipo ## <dbl> <dbl> <chr> ## 1 30.1 0.00000891 Menor ## 2 30.5 0.0000100 Menor ## 3 30.9 0.0000113 Menor ## 4 31.4 0.0000128 Menor ## 5 31.8 0.0000144 Menor ## 6 32.2 0.0000161 Menor ## 7 32.7 0.0000181 Menor ## 8 33.1 0.0000203 Menor ## 9 33.5 0.0000227 Menor ## 10 34.0 0.0000255 Menor ## # ... with 502 more rows ## ## ## $inversion

96

null

2018-09-13

2018-09-13 04:07:06

2018-09-13

2018-09-13 04:31:44

10

false

es

2018-09-13

2018-09-13 04:31:44

8

1ba3d220539d

31.842453

0

Mi juego favorito es es Magic: the Gathering. Este es un juego de cartas coleccionable, en el que armas un mazo, siguiendo ciertas…

5

Usando R y Data Science para analizar los precios de Magic: the Gathering Mi juego favorito es es Magic: the Gathering. Este es un juego de cartas coleccionable, en el que armas un mazo, siguiendo ciertas limitaciones, con el objetivo de vencer los mazos de tus oponentes. Este es un hobbie que requiere de una inversión relativamente alta, comparado con otros, así que pensé que sería buena idea hacer un pequeño proyecto de Data Science usando R y datos del sitio MTG Goldfish para estimar la posibilidades de recuperar mi inversión en Magic, en particular, al comprar una caja de sobres de cartas. Y para empezar, una breve introducción a cómo funciona la economía de Magic: the Gathering. TM & © 2018 Wizards of the Coast ¿Cómo funcionan los precios de Magic: the Gathering? En Magic: the Gathering (Magic), cada año aparecen nuevas cartas a la venta en lo que se denomina expansiones o sets. Estos sets son vendidos en sobres de cartas que contienen, generalmente, 15 cartas al azar. Estos sobres a su vez, pueden ser comprados en cajas que contienen, generalmente, 36 de ellos. Así que si una persona desea obtener las cartas más recientes de un set comprando sobres, depende en gran medida de la suerte. Como este es un juego y es coleccionable, hay cartas en cada set que son más difíciles de conseguir, pues la frecuencia con la que aparecen al azar en los sobres es menor. Es decir, son cartas más raras. En Magic las cartas se clasifican por rareza, de las menos a las más raras: Común (Common) Infrecuente (Uncommon) Rara (Rara) Mítica (Mythic) Además, Magic: the Gathering tiene un sistema de juego organizado, en el que se organizan distintos eventos competitivos, que van desde torneos en tiendas locales hasta eventos competitivos nacionales e internacionales. Lo anterior tiene como consecuencia que aquellas cartas que son efectivas en juego competitivo, pues incrementan las probabilidades de ganar o forman parte de estrategias exitosas, ven un incremento en su costo, pues su demanda se incrementa. Así, cartas que son raras de conseguir y tienen alta demanda, tienen precio más alto que cartas comunes o poco usadas en juego competitivo. Por supuesto, hay otros factores que inciden en los precios de las cartas, pero para este artículo basta esta explicación simple. También es muy importante tener en cuenta que los precios de las cartas, frecuentemente, son expresadas en dólares norteamericanos (USD), que es la moneda que usaremos en este análisis. Además, los precios que usaremos son aquellos correspondientes al 12 de Septiembre del 2018. MTG Goldfish Naturalmente, existen sitios de internet que se encargan de dar seguimiento a los precios de las cartas y sus tendencias. Esta es una información sumamente importante para los jugadores de Magic, pues así conocen el valor de las cartas que poseen y aquellas que desean, para sí tomar decisiones de compra, venta e intercambio. Uno de estos sitios es MTG Goldfish. En este sitio se encuentran disponibles información actualizada diariamente de precios de todas las cartas, de todos los sets de Magic, presentados en tablas que son muy convenientes para su análisis usando R. Además, en este sitio se encuentran excelentes artículos, videos, podcast y artículos a la venta relacionados con Magic. Así que si te gusta Magic o te da curiosidad este juego, no dudes en visitar MTG Goldfish y consumir su contenido, no te arrepentirás… y es una manera de agradecer por la información que generan. Conociendo todo esto, comencemos el análisis para calcular la probabilidad de recuperar nuestra inversión en una caja de Magic. Estructura del análisis. Las etapas de nuestro análisis son las siguientes Descarga de los precios de un set desde MTG Goldfish. Procesamiento de los precios para facilitar el análisis. Análisis exploratorio del set. Simulación de valores (por sobre de cartas y por caja de sobres). Cálculo de probabilidad de recuperar inversión. Sistematización del análisis (para ejecutarlo de manera repetida). Algunos de estás son más complejas o largas que otras, pero si abordamos el análisis dividido por pasos, es más fácil tomar decisiones sobre el análisis y corregir errores cuando se presentan. Establecido esto, preparemos nuestro entorno de trabajo. Paquetes necesarios Para este proyecto usaremos los siguientes paquetes: tidyverse: Una familia de paquetes para importar, manipular, exportar y visualizar datos. rvest: Funciones para extraer información de páginas web (scrapping). xml2: Este paquete nos permite importar y manipular datos con estructuras xml. scales: Funciones para mejorar la presentación de datos en gráficos. ggrepel: Un paquete que expande las funciones de ggplot2 (parte de tidyverse), para mostrar etiquetas sin sobreposiciones entre ellas. Como es usual, puedes instalar estos paquetes usando la función install.packages(). Ahora sí, pasemos a la primera etapa del análisis. Descarga de los precios de un set desde MTG Goldfish En MTG Goldfish, la información de los precios de un set de Magic es presentada en su propia página, que es identificada usando el código del set. Los sets de Magic son identificados por un código de tres caracteres, en mayúsculas. Por ejemplo, para el set “Dominaria”, su código es “DOM”, por lo que el URL en el que encontramos sus precios se encuentran en: https://www.mtggoldfish.com/index/DOM Una lista de los códigos de sets de Magic se encuentra disponible en: https://mtg.gamepedia.com/Template:List_of_Magic_sets Usamos la función download_html() del paquete xml2() para descargar una página de internet a nuestra carpeta de trabajo. Llamaremos a este archivo “goldfish_DOM.html”. Podemos definir una función que nos permita descargar el html que corresponde a un set de manera más sencilla, proporcionando el código de tres letras de este. Hecho esto, importamos el archivo con el html descargado y lo asignamos al objeto mtg_dom. Si llamamos a este objeto, podremos ver que es un documento con estructura de xml. Con este formato no podemos hacer mucho. Necesitamos extraer los datos que contiene y para ello, usaremos las funciones del paquete rvest. 5 Procesamiento de los precios para facilitar el análisis rvest es un paquete usado para extraer y procesar información de documentos html o xml. En nuestro caso, lo haremos a través del uso de identificadores CSS. No nos detendremos a explicar qué son los identificadores CSS, pero en términos generales, podemos decir que estos describen los elementos y características de una documento html. Puedes leer más al respecto en el siguiente enlace: https://www.w3schools.com/css/ Vamos a recuperar los datos de precios de las cartas, así que necesitamos los identificadores CSS de estos datos en particular. Para obtener los identificadores hay distintos procedimientos. Puedes usar la función “Inspeccionar elemento” de tu navegador de internet y explorar la estructura del documento. También puedes usar el sitio Selector Gadget para recuperar los identificadores CSS: https://selectorgadget.com/ Para fines de este artículo, ya he realizado esta tarea. los identificadores CSS que nos interesan son: “.index-price-table-paper tbody tr”. Con esta información, usamos la función html_node() de rvest(). Aún no es un formato útil para el análisis necesitamos la función html_text(). Hacer esto nos devolverá una cantidad considerable de texto, así que pediremos que se nos muestren sólo las primeras líneas con head(). El resultado es algo más manejable que el xml original, pero requiere más procesamiento. Usamos la función str_split() con as.data.frame() de R base y tbl_df() de dplyr, para convertir el resultado anterior a un data frame, que llamaremos df_dom. Nuestro resultado es el siguiente. Aun tenemos que pulir un poco nuestro data frame, en particular, seleccionando sólo las columnas que nos interesan entre todas las disponibles. Haremos esto con select() de dplyr Nuestro resultado será un data frame con cuatro columnas: el nombre de las cartas, el set, la rareza, y el precio. Definiendo una función para importar y procesar precios Podemos transformar el proceso anterior en una función, llamada leer_html() para así generar fácilmente data frames a partir del html de páginas de MTG Goldfish. De este modo, casi hemos terminado el procesamiento de los precios, pero antes tenemos que hacer una recodificación. Identificando outliers Como deseamos analizar las probabilidades de recuperar nuestra inversión en un set de Magic, nos conviene identificar aquellas cartas que tienen un valor excepcionalmente alto, con respecto a las demás. Esta información no permitirá hacer un análisis más fino de los precios con los que contamos. Por supuesto, estas cartas con precios excepcionales, son outliers. En cuanto a outliers no hay un consenso en cómo caracterizarlos. Para fines de este proyecto, usaremos como criterio para etiquetar un dato como. outlier el usando al generar gráficos de caja y bigote (boxplot), que es: Un dato menor a: el valor del primer cuartil menos una vez y media el rango intercuartílico. Un dato mayor a: el valor del tercer cuartil más una vez y media el rango intercuartílico. Estos son los datos que se salen de los “bigotes” de un boxplot y son mostrados como puntos en estos diagramas. Sólo etiquetaremos los outliers “altos”, pues son los relevantes para el análisis que estamos haciendo. Definimos entonces una función que implemente el criterio anterior. Al darle un vector numérico, nos será devuelto un vector lógico del mismo largo, donde el valor TRUE serán los outliers y FALSE los demás datos. Aplicamos esta función, agrupando por nuestras cartas por rareza. No tiene mucho sentido comparar los precios de cartas Comunes con Míticas, pues estas últimas siempre tienen un precio más alto que las primeras por ser más difíciles de obtener. Hacemos un mutate() adicional para etiquetar los outlier con el nombre de la carta, esto nos será útil más adelante. Podría usar una sola llamada de mutate(), pero he preferido presentarlo así para hacer más claro qué está ocurriendo. Nuestro resultado es el siguiente. Por supuesto, podemos definir otra función para llevar a cabo el proceso anterior. Llamaremos a esta función etiquetar_outlier(). Nuestro siguiente paso es explorar el set “Dominaria”. Análisis exploratorio del set Iniciemos la exploración con los estadísticos descriptivos más elementales: la media, mediana estándar y máximo y mínimo de los precios. La desviación estándar, aunque es un estadístico descriptivo por excelencia, en este caso es poco informativa debido a la forma en que se distribuyen los precios. En Magic, los precios no tienen una distribución normal o cercana a una normal, pues tienden a existir muchas cartas con precio bajo y casi idéntico, con algunas pocas cartas con precios muy altos. Por lo tanto, una medida de dispersión como la desviación estándar no nos ayuda mucho a describir lo que nos encontraremos. Usamos las funciones group_by() y summarize() de dplyr, para aplicar las funciones mean(), median(), min() y max(), que corresponden a los estadísiticos ya mencionados, por rareza. Confirmamos que las cartas comunes son las menos caras de todas y las míticas las más costosas. También es evidente que hay una variación muy alta entre el precio mínimo y el máximo, en todas las rarezas, particularmente en raras y míticas. También podemos visualizar cómo se distribuyen los precios creando un gráfico de densidad con ggplot2, llamando la función geom_density() de este paquete. Debido a que tenemos diferencias considerables entre los precios más bajos y más altos, esta gráfica no aporta mucha información. Sin embargo, ya sabemos que los precios de las cartas dependen de su rareza. Las cartas más raras tienden a ser más caras podemos visualizar los precios por rareza puede ser más útil. Usamos la función facet_wrap() de ggplot2 para generar un gráfico con las características anteriores. Usamos el argumento scales = "free" para que los ejes x y y sean escaladas de manera independiente. En el eje x se mostrará el precio y en el y la densidad. De esta manera es más claro observar la distribución de los precios. También confirmamos lo que anticipábamos, los precios no tienen una distribución parecida a una normal. Finalmente, podemos explorar los precios de este set por rareza usando diagramas de caja y bigotes (boxplots). La ventaja de emplear esta forma de visualización es que podemos ver fácilemente las cartas que hemos marcado como outliers. Para lo anterior, usamos la función geom_boxplot() de ggplot2 para generar el diagrama y geom_label_repel() de ggrepel para agregar etiquetas. Esta última función es una versión de geom_label() de ggplot2, que tiene ajustes para evitar que las etiquetas se superpongan, mejorando así la legibilidad. Creo que con esto tenemos una buena idea general de los precios de “Dominaria”. En particular, podemos identificar aquellas cartas que son especialmente caras, lo cual es una ayuda para comprar, vender e intercambiar. Aprovechamos para definir una función que realice todas las operaciones de exploración, a la que llamaremos explorar_set(). De paso, agregamos unas mejoras de presentación a los gráficos, usando la función dollar_format() de scales. Ahora sí, estamos listos para estimar qué tan probable es que recuperemos nuestra inversión monetaria si decidimos comprar cajas de sobres de una expansión de Magic. Simulaciones Necesitamos definir dos procesos de simulación, una que simule los resultados de abrir un sobre de cartas y una que simula el resultado de abrir una caja de sobres. Como mencionamos en la introducción de este proyecto, los sets de Magic, normalmente, son vendidos en sobres que contienen 15 cartas, distribuidas de la misma manera: 1 carta de tierra básica. 10 cartas comunes (Common) 3 cartas infrecuentes (Uncommon). 1 carta rara (Rare), o 1 carta mítica (Mythic) en uno de cada ocho sobres, aproximadamente. Estos sobres, a su vez, generalmente se venden en cajas que contienen 36 de ellos Por lo tanto, necesitamos simular el contenido de un sobre de cartas y repetir ese ejercicio 36 veces, para así determinar el valor monetario de una caja de Magic. Hecho esto, podremos entonces comparar el valor monetario de una caja de Magic con la inversión que hagamos para comprarla, es decir, el precio que paguemos por ella. Con estas consideraciones, comencemos simulando sobres. Simulación de un sobre Crearemos una lista de listas de pares con las rarezas de las cartas y la frecuencia con la que aparecen, tomando como referencia lo mencionado en la sección anterior, de la siguiente manera. Cada lista tiene dos elementos, el primero es la Rareza y el segundo es la frecuencia. Omitimos las cartas de tierra básica, pues estas generalmente no tienen ningún valor financiero. Usaremos esta lista con una función anómima dentro de la función map() de purrr. La función map() es muy similar a lapply() de R base, pues aplica una función a todos los elementos de una lista. Lo que haremos será aplicar una función anónima para filtrar por Rareza las cartas de nuestro data frame con los precios de “Dominaria” (primer elemento) y luego extraer una muestra de ellas igual a la frecuencia con la que aparecen (segundo elemento). Usaremos entonces, en conjunto con map(), las funciones filter() y sample_n() de dplyr. Haremos esto con set.seed(), para que los resultados sean reproducibles. El resultado es una lista, que representa un sobre de Magic, con una carta rara, tres infrecuentes y trece comunes. Como usaremos de manera repetida esta función anónima, es mejor que le demos nombre y la definamos. La llamaremos pareja(). De esta manera, podemos usar map() para generar sobres con una sola línea. Como necesitamos un data frame para análisis posteriores, corremos lo anterior seguido de la función reduce() de purrr, que aplica una función de manera secuencial a todos los elementos de una lista, por parejas. Aplicaremos bind_rows() de dplyr para obtener como resultado un data frame. Lo anterior, aunque cumple nuestro cometido, no nos permite crear sobres con cartas míticas. Necesitamos solucionar esta situación. Las cartas míticas Dado que las cartas míticas aparecen en un sobre una de cada ocho veces (p=1/8), podemos simular este comportamiento con una distribución binomial. Usamos la función rbinom(), para generar 1 y 0 al azar, teniendo el 1 una probabilidad de 1/8. de ocurrir, esto es, 12.5% de los casos. Pongamos esto a prueba, simulando 40 números con estas probabilidades. Luce bien. Lo que haremos será establecer una condición con ìf. Si la rareza de la que estamos extrayendo cartas es “Rare” y al mismo tiempo obtenemos un 1 de simular una distribución binomial con las características descritas arriba. Dado lo anterior, lo siguiente nos devolverá una carta Rara. Y lo siguiente una carta Mítica. Combinamos esto con la función pareja() para definir una función llamada simular_sobre(). Pongamos a prueba nuestra función simular_sobre(). Equipados con esta función, podremos crear cajas de sobres fácilmente. Simular caja de sobres Una vez más, usamos la función map(), aplicando la función simular_sobre() 36 veces. De esta manera obtenemos una lista con 36 elementos, cada uno de ellos representando un sobre de Magic. Por ejemplo, este es el cuarto “sobre” de la lista anterior. Definamos entonces una función para simular una caja de Magic, que nos de como resultado un data frame con los 36 sobres, utilizando reduce() y bind_rows(). Probamos nuestra función y asignemos el resultado al objeto caja_dom. Comprobamos que hemos generado 36 cartas Raras o Míticas. Para una simulación apropiada, necesitamos repetir generar muchas cajas de sobres. Simulación de múltiples cajas Usamos map() para repetir 100 veces la función simular_caja() y así obtener esa misma cantidad de cajas. Una vez más, empleamos reduce() para obtener como resultado un data frame. Como anteriormente dejamos agrupados nuestros datos por Rareza, obtendremos un data frame con cerca de 400 renglones, uno por rareza de cada caja de Magic simulada. No siempre serán 400 renglones, pues hay una probabilidad pequeña de que alguna desafortunada caja no tenga ni una sola carta Mítica y con ello su valor se reduzca. Ahora podemos calcular el valor promedio de una caja de “Dominaria” y de las cartas por rareza. Valor promedio de una caja y por Rareza Si los deseamos, podemos obtener fácilmente el valor promedio por rareza de las cartas simuladas usando summarize() de dplyr. Sin embargo, necesitamos hacer un pequeño ajuste si buscamos conocer el valor promedio de cada caja de sobres simulada. Necesitamos proporcionar un identificador por caja, para así poder obtener un valor total por cada una de ellas y, con este, calcular el valor promedio por caja. Hagamos este ajuste y de una vez definamos una función llamada simular_ciencajas(), aunque en realidad, podremos definir el número de iteraciones que deseemos, no sólo 100. Pongamos a prueba nuestra función. Nuestro resultado es muy similar al anterior, sólo que ahora tenemos identificador por caja. Con este identificador podemos obtener el valor promedio de una caja de sobres de “Dominaria”. La cosa marcha bien, pero podemos hacer nuestro análisis más fino si consideramos una característica del valor de una caja de Magic: las cartas comunes generalmente tiene poca “liquidez”. Aunque las cartas comunes tienen un valor monetario, a menos que sea una carta de alta demanda y por tanto de un precio particularmente alto, es un poco difícil venderlas y recuperar lo invertido en obtenerlas. Por suerte nosotros ya hemos etiquetado las cartas comunes con precios inusualmente altos con nuestra función tag_outlier(). De este modo, sólo incluiremos a estas al calcular el valor promedio de una caja. Agregamos esta información en la definición de la función simular_ciencajas(). Con esta nueva versión de nuestra función simular_ciencajas(), simulamos cien cajas de “Dominaria” y, con esos datos, obtenemos su valor promedio. De una vez calculemos también la mediana, el valor que tiene 50% de los precios debajo de él y 50% por encima. Nuestro promedio, naturalmente, es menor que el obtuvimos antes de excluir las cartas comunes. Con lo que hemos hecho hasta ahora, podríamos decir cuál es el valor promedio que esperaríamos obtener de una caja de sobres de “Dominaria”. Con ello podemos hacernos una idea general sobre la posibilidad de recuperar nuestra inversión. Si compramos una caja de “Dominaria” por un precio inferior al valor promedio esperado de ellas, deberíamos recuperar nuestra inversión ¿cierto? Bueno, la cosa no es tan sencilla. Calculemos, por fin, la probabilidad de recuperar nuestra inversión. Probabilidad de recuperar inversión Después de simular el valor de múltiples cajas de sobres de Magic, podemos calcular la probabilidad de recuperar lo que hemos invertido en una de ellas a partir de una función de densidad. Primero, obtenemos el valor de cada caja, agrupando los Precios de las cartas en ellas por Id_caja. Una vez que hemos obtenido estos valores, los extraemos como un vector utilizando con la función pull() de dplyr. Este es nuestro resultado. Definimos una función que realice lo anterior, llamada valor_cajas. Usamos density() en nuestro vector anterior para obtener la función de densidad de estos valores. Obtenemos lo siguiente. Si deseamos obtener una gráfica de esta función, usamos ggplot() y geom_area() de ggplot2. Pero antes necesitamos extraer la información de los ejes x y y, para después convertirla a un data frame con tbl_df(). Con esto obtenemos una curva de función y su respectiva área debajo de ella. Esta área representa, aproximadamente, una distribución de todos los valores que podrían tomar las cajas de sobres Magic, a partir de la simulación que hemos realizado. Es decir, el 100% de nuestros casos. Podemos trazar una línea vertical que divida esta área bajo la curva en dos, justo en el valor de nuestra mediana: 104. Si hacemos esto, tendremos dos segmentos del área, uno con 50% de los valores por debajo de 104 (área a la izquierda) y 50% con los valores por encima de este valor (área a la derecha). Como esta es una representación de una distribución, lo anterior quiere decir que si sacamos un valor al azar de esta distribución, tenemos 50% de probabilidad de que sea menor a 104 y 50% de probabilidad que sea mayor. Veamos como luce lo anterior con un valor de 105, que obtuvimos antes. ¡Interesante! Por lo tanto, si queremos calcular la probabilidad de que recuperemos nuestra inversión al comprar una caja de sobres de Magic, debemos calcular el área que resulta de segmentar nuestra distribución en dos. Para calcular el área de un segmento bajo la curva, usamos las funciones approxfun() e integrate(). Como su nombre lo indica, la función integrate() usará integración para calcular un área, por lo que nos pedirá un límite inferior y uno superior. Supongamos que hemos comprado una caja de sobres de “Dominaria” en 100 dólares. Usaremos este valor como límite inferior y el valor máximo de nuestra distribución como límite superior. ¡Perfecto! Lo que hemos obtenido es la probabilidad de que recuperemos nuestra inversión, esto es decir, hay 58% de probabilidad de comprar una caja y que esta tenga un valor mayor que 100 USD. En este caso es más o menos lanzar una moneda al aire recuperar nuestra inversión si pagamos 100 USD por una caja de “Dominaria”. Transformemos el proceso anterior a funciones, para repetirlo fácilmente. Sistematización del análisis Primero, definimos una función que calcule la función de densidad, la probabilidad a partir de un costo que elijamos, y el data frame de la función de densidad. Además, aprovecharemos para etiquetar los valor en el data frame de la función de densidad como menores o mayores al costo elegido. El resultado será una lista con cuatro elementos: el costo pagado, la función de densidad, probabilidad y un data frame. Ahora, creamos una función que use esta lista para generar un gráfico. Hecho esto, podemos probar con otros costos de una caja de Dominaria. Por ejemplo, si pagamos 90 USD, naturalmente nos irá mejor. Vamos a integrar todos los pasos de nuestro análisis anterior en una sola función llamada analisis_set(). De esta manera, podemos realizar el análisis de cualquier set en un solo paso, siempre y cuando descarguemos primero los datos de este de MTG Goldfish. El resultado será una lista con el data frame con los precios del set, el análisis exploratorio, incluidas gráficas, el resultado de las simulaciones, y el análisis de la probabilidad de recuperar información. Probemos descargando la información del set “Ixalan”, que tiene la clave “IXL” con nuestra función descargar_set() Con los datos de “Ixalan”, podemos realizar una simulación usando nuestra función analisis_set() que nos permitirá calcular la probabilidad de recuperar nuestra inversión si compramos cajas de este set por 90 USD cada una. Veamos nuestros resultados. ¡Vaya! no hay mucha diferencia entre estos sets, así que cualquiera de los dos tendrá las mismas probabilidades de devolver nuestra inversión. Conclusiones En este artículo revisamos las diferentes etapas de un pequeño proyecto de Data Science en el que hemos calculado la probabilidad de recuperar nuestra inversión al comprar cajas de sobres de Magic: the Gathering, a partir de la información de MTG Goldfish Para esta tarea hemos usados los paquetes xml2 y rvest para obtener y procesar información de un sitio web, así como la familia tidyverse con ggrepel y scales para analizar y visualizar resultados. Creo que en términos generales hemos creado una herramienta que nos puede ayudar a tomar decisiones a la hora de comprar decisiones al gastar dinero en Magic, o al menos, entender mejor las tendencias financieras relacionadas con este juego de cartas coleccionable. Desde luego, aún hay formas en las que podemos perfeccionar este proyecto, por ejemplo: Agregar a las funciones procedimientos para manejar errores y excepciones. Crear un panel interactivo con Shiny. Mejorar la presentación visual de la información. Obtener los nombres de las cartas en varios idiomas. Convertir los precios de dólar norteamericano a moneda local (peso mexicano, en mi caso). Pero algunas de estas cosas, pueden ser motivo de futuros proyectos. Consultas, dudas, propuestas de temas, comentarios y correcciones son bienvenidas: jboscomendoza@gmail.com El código y los datos usados en este documento se encuentran en Github: https://github.com/jboscomendoza/rpubs/tree/master/mtg_goldfish

Usando R y Data Science para analizar los precios de Magic: the Gathering

0

usando-r-y-data-science-para-analizar-los-precios-de-magic-the-gathering-1ba3d220539d

2018-09-13

2018-09-13 04:31:44

https://medium.com/s/story/usando-r-y-data-science-para-analizar-los-precios-de-magic-the-gathering-1ba3d220539d

false

8,107

null

Data Science

data-science

Data Science

33,617

Juan Bosco Mendoza Vega

Psicólogo y evaluador educativo, podcaster, nerd. Enfocado a la evaluación a gran escala y a usar data science para mejorar la toma de decisiones.

11d4aaf39775

jboscomendoza

145

156

20,181,104

null

0

null

0

null

2017-10-21

2017-10-21 09:04:21

2017-10-21

2017-10-21 09:20:42

1

false

en

2017-10-21

2017-10-21 12:42:50

6

1ba47430a442

1.65283

0

Artificial Intelligence and Internet of Things perspective. Whether or not data coming from everyday life objects offers great importance…

5

What do shark recognition, gunfire detection, and bike sharing have in common? Artificial Intelligence and Internet of Things perspective. Whether or not data coming from everyday life objects offers great importance and value is a good question. How do you make the best of it? Who can make the best of it? “Connected’ does NOT necessarily mean smart.” — Bill Schmarzo Three examples can be given for illustration. One in China, related to the bike sharing sector, one in the USA within urban lighting market, and the last one in Australia in the area of shark detection. In China, the oFo bike app provides you with the code to unlock a shared bike and handle the payment process. That helps to answer the questions: How many people use the bikes? Where do most people take or drop bikes off? In the same way as oFo’s people use the data to optimize their distribution plans, the data is gold for city officials in order for them to plan the design of their urban architecture: roads, bus lanes, cycle lanes, walkways etc. In California, GE is working with ShotSpotter, which is designed for detecting and locating gunfire in real time. Teaming up with Intel and AT&T, GE is combining cameras, microphones, and sensors into its intelligent LED streetlights. GE is then upgrading its urban lights far beyond simple illuminating capabilities. While today, ShotSpotter is using complex real-time algorithms and the work of a team of human specialists, the data collected by IoT in real time can be associated with available historical data to be processed by Deep Learning for further automation. In Australia, the University of Technology Sydney (UTS) is harnessing Deep Learning and Artificial Intelligence algorithms to detect sharks from drone footage. First, they preprocess public videos of sharks. This is the learning part. Then, a Neural Network runs detection and recognition algorithms. Empowering the action of drones, this provides a real-time search and rescue service. Want to take an in-depth look at the state of AI and IoT and to explore how these technologies synergize with each other and what to expect in the future? Take some time to read my last DZone article “AI and IoT: Taking Data Insight to Action” here. One more thing… Don’t anthropomorphize AI’s, they hate that.

What do shark recognition, gunfire detection, and bike sharing have in common?

0

what-do-shark-recognition-gunfire-detection-and-bike-sharing-have-in-common-1ba47430a442

2017-10-21

2017-10-21 12:42:50

https://medium.com/s/story/what-do-shark-recognition-gunfire-detection-and-bike-sharing-have-in-common-1ba47430a442

false

385

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Fred Jacquet

Going around the Data industry for >20 years I seek to make think ahead bring new perspectives and formulate creative ways to understand Digital Transformation.

bffd9bfc8bdb

aFredNotAfraid

8

10

20,181,104

null

0

null

0

null

2018-02-27

2018-02-27 20:44:03

2018-03-02

2018-03-02 04:44:08

7

false

en

2018-03-02

2018-03-02 19:42:09

30

1ba51a884f02

8.585849

24

0

1

With a plethora of complex tools for machine learning, I want to introduce a company looking to democratize artificial intelligence, making…

3

Introduction to H2O.ai With a plethora of complex tools for machine learning, I want to introduce a company looking to democratize artificial intelligence, making machine learning simple and ubiquitous for all. Introduction Artificial Intelligence…at this point you know that it is the wave of the future, going to make you very rich if you add it to your career’s skills list and eventually going to kill us all. As those first two points excite the average person, the next step is usually them trying to figure out how to learn more about the field. They typically find an online course or some introduction where after, the logical progression is to get some hands on experience. That is when they run into what I like to call the “tooling issue.” Figure 1: So many tools, so little time. There are so many tools to choose from it becomes hard to discern which tool is right for the job. I’m not going to compare and contrast tools, as the ones in Figure 1 represent less than a tenth of available tools, but what I will do is introduce a company with a mission to simplify and democratize artificial intelligence for all. Whether they ultimately deliver on that or not is still up in the air, but at the very least they have given us a good starting point for our machine learning efforts. Who is H2O.ai? H2O.ai is the company behind open-source Machine Learning (ML) products like H2O, aimed to make ML easier for all. Their open-source community includes more than 129,000 data scientists and 12,000 organizations. They also have almost half of the Fortune 500 using their software with a 330% growth in users over the past two years. Analyst have also taken note as they’ve been classified as a Leader with the most completeness of vision, among 16 vendors included in Gartner’s 2018 Magic Quadrant for Data Science and Machine Learning Platforms. Their business partners include major players like Microsoft, IBM, NVIDIA, Splunk, Databricks, MapR, Anaconda, Cloudera and a few others. So how did they gain such a large following? By mostly delivering on their promise to make machine learning accessible and allow business users to extract insights from data, without needing expertise in deploying or tuning machine learning models. Figure 2: The AI Platform for Enterprise If you’ve ever worked in a large company and had to deal with a slow moving digital transformation, your probably excited about a company aiming to simply usage of a complex software but also skeptical of them as almost every tech company makes that promise and many fail to deliver. Lets analyze H2O.ai deeper and see how they deliver on this promise along with use cases with their product running in production. Product Offering H2O.ai states it is making ML accessible by allowing business users to extract insights from data, without needing expertise in deploying or tuning machine learning models. How are they doing this? Through their suite of machine learning products we’ll discuss. Figure 3: H2O.ai Suite of Machine Learning Tools H2O. The primary offering. An open source, in-memory, distributed, ML and predictive analytics platform allowing you to build and productionize ML models. Contains supervised and unsupervised models like GLM and K-Means clustering, and a simple to use web-UI called Flow. Deep Water. H2O + a tighter integration with TensorFlow, MXNet and Caffe for GPU based DL workloads. Sparkling Water. H2O + a tighter Spark integration for customers to utilize their existing Spark ecosystem with H2o’s ML. We’ll learn more about this later in our use cases section. Steam. Company’s enterprise (meaning not free) offering for building and deploying applications. Scientist can train and deploy ML models, making them accessible over APIs for developers to integrate into applications. Driverless AI. Bit of a misnomer as it isn’t exclusive to autonomous driving but is H2o + a simplified wrapper to help enterprise’s non-technical employees prepare data, calibrate parameters and determine optimal algorithms for tackling specific business problems with ML. Makes Automatic Feature Engineering, Model Tuning, Selection and Ensambles (using multiple learning algorithms to obtain better predictive performance) easy to use for those who don’t even know what those terms mean. Quick video demo. Now that we have a high level overview of the products offered by H2O.ai, lets explore their technical capabilities. Technical Features & Capabilities H2o’s products provide an open-source, in-memory, distributed, fast, and scalable machine learning and predictive analytics platform. The “fast” comes from the data being distributed across the cluster and stored in memory in a compressed columnar format, allowing you to read the data in parallel. H2O’s core code is written in Java and, like many other modern applications, H2O provides a REST API for access to all of the software’s capabilities from an external program or script via JSON over HTTP. The Rest API is used by H2O’s web interface (Flow UI), R binding (H2O-R), and Python binding (H2O-Python). Figure 4: H2O Flow is a notebook-style open-source user interface for H2O. Available Machine Learning Algorithms When it comes to available machine learning algorithms, H2O has a nice set of available algorithms for users to leverage. Supervised Learning. Deep Learning (Neural Networks), Distributed Random Forest (DRF), Generalized Linear Model (GLM), Gradient Boosting Machine (GBM), Naïve Bayes Classifier, Stacked Ensembles, XGBoost Unsupervised. Aggregagtor, Generalized Low Rank Models (GLRM), K-Means Clustering, Principal Component Analysis (PCA) Others. Quantiles, Early Stopping, Word2Vec H2O’s data parser has built-in intelligence to guess the schema of the incoming dataset and supports data ingest from multiple sources in various formats. Big Data & Anaconda Integration For big data, H2O integrates well with the Hadoop ecosystem of tools, including Spark. The supported Hadoop distributions are Cloudera CDH, Hortonworks HDP, MapR 4.x+ and IBM Open Platform. While H2O integrates with Hadoop and Spark, they are not required to run H2O unless you want to deploy H2O on a Hadoop or Spark cluster. H2O also integrates with Conda, the open source package and environment management system used by data scientist, that quickly installs, runs and updates packages and their dependencies. Conda and Anaconda can be used on Anaconda Cloud, where packages, notebooks, projects and environments are shared during collaborations. Conda is NOT required to run H2O unless you want to run H2O on the Anaconda Cloud. For the sake of brevity I’ll skip over a some of the technical details but if you want to learn more about things like the architecture components, JVM management, CPU management or things like Fluid Vector Frames then follow the links. Customers and Use Cases So who is using this technology and for what use use cases? Glad you asked. H2O highlights their usage by the financial, insurance and healthcare industries among others. Examples of specific solutions enabled by H2O: Cisco. Used H2O in creation of a P2B statistical model that tries to predict whether a certain company will buy a certain product in a given time frame in the future. Model outputs probability that a company will buy and the amount of money a company will spend if they do buy the product. Used by sales, fueled by demographics, past purchase behavior, contacts, marketing interactions, cust satisfaction surveys, macroeconomic indicators, purchases/non-purchase list. Capital One. Systems and Operations Group used H2O for Capital One banking app (5K users a minute, 300K an hour) as they found H2o satisfied their governance requirements, leveraged current data science and ML resources (R, Python, Spark, etc), is open-source, usable and scalable. Equifax. Built a product on top of H2O, called Ignite, which is their “revolutionary portfolio of premier data and advanced analytics solutions.” Kaiser Permanente. Found that medical/surgical ward patients urgently transferred to ICU show evidence of physiologic derangement 6–24hrs in advance of the ICU transfer. These patients are less than 5% of all patients in a hospital but are 25% of all ICU admissions, 20% of all deaths in a hospital and 12.5% of days patients stay overnight in a hospital. Used H2O to predicting these patients visit to the ICU in advance, which reduced mortality rates by 2–5x. Too high level for you? Want to get into more details on how a company can utilize H2O to create ML models and gain better insights and server their customers through data analysis? Sure, that's up next Large Scale ML and Predictions for Travel Services Customer I want to now provide a more specific use case for a large travel services organization, bringing in over $68B in revenue a year. This company wanted to use ML for scoring hotels and destination recommendations for millions of users and select from 1M+ keywords during bidding on ad platforms, ensure the most effective ad placements. They also wanted to run Machine Translation jobs to transform hotel/flight descriptions to one of the website’s 43 different supported languages depending on the user’s language preference. Machine Learning Requirements Based on analysis of their existing system and previous attempts at deploying machine learning, they wanted a solution that: scales well easy to use statistically sound easily moved to production reliable fast Initial Attempts Their initial attempts included: Using SparkML. Found it was unstable, not feature rich and difficult to productionize with slow predictions (2015) Using Sparkling Water (h2o+Spark). Solved most of their problems except for scaling as Sparkling Water was tied to YARN which broke during scaling their 1,000+ node cluster. Using Google Translate APIs for machine translation. Found Google’s APIs weren’t as accurate as training models against their own data sets The final solution was to work with H2o and contribute orchestration code upstream for Sparkling Water. In doing so they developed an “external cluster mode” for H2O which got rid of YARN dependency, helping them scale over 1K+ nodes with no issues. Solution Architecture Figure 5: High level overview of the architecture for their solution We’ll now walk through a high level overview of the architecture and model training used in production to make predictions. Offline: Model Creation and Training (Safe, non-Production) End user interacts with the website by creating a click-stream event which could be looking at a hotel, selecting a flight, reading reviews, etc. For JSON events not requiring a prediction, data is sent to a Spark Cluster for workflows to run on top of data. Data transformed to store and for model training. Data Scientist then create data frames on the data set in Spark and send to the H2o cluster. In H2o, the scientist chooses the correct algorithms, train and build models. Also have access to a feature store with a web UI where they can discover or reuse features, see which are available online vs offline, assign ownership of features and enforce quality. After training, they export their models in Plain Old Java Object (POJO) or Model ObJect Optimized (MOJO) formats to the prediction processing system for use in production. Online: Hotel Scoring Example (Real-Time Processing) End user creates a click-stream event requiring a prediction when they search for a hotel in NY. That JSON event is sent to a Kafka topic Spark takes the event from Kafka and runs transformations on it, similar to data wrangling done offline, outputting the value to another Kafka topic and Cassandra as the persistent back end. The stream processing application then makes an API call to the ML model which has an API gateway in front to scale prediction processing requests as needed. For hotel query, ~10,000 hotels are scored for that specific user and returned to the user where they hopefully click and and book the top result. Conclusion While we’ve covered a lot, we have only scratched the surface of the capabilities of the H2O platform. H2o seems well poised to reach their goals of democratizing AI for everyone as adoption of their product offering is growing at a rapid pace. If you want to learn more, checking out their keynote and sessions from the 2017 H2O World is a good place to start. Hope this introduction to H2O has been helpful. If you enjoyed this article, please tap the claps 👏 button. Interested in learning more about Jamal Robinson or want to work together? Reach out to him on Twitter or through LinkedIn.

Introduction to H2O.ai

144

introduction-to-h2o-ai-1ba51a884f02

2018-06-20

2018-06-20 15:01:08

https://medium.com/s/story/introduction-to-h2o-ai-1ba51a884f02

false

1,997

null

Machine Learning

machine-learning

Machine Learning

51,320

Jamal Robinson

Enterprise technologist with experience across cloud, big-data, AI and other cool technologies. Recently published author of “Rich People Making Poor Decisions”

bdedf45adfcb

jamal.robinson

21

1

20,181,104

null

0

null

0

634d4b270054

2018-04-20

2018-04-20 07:03:20

2018-04-20

2018-04-20 07:05:58

1

false

en

2018-06-05

2018-06-05 08:45:11

3

1ba54ee17551

1.086792

0

Drone racing is returning back and this year the flight path will be among the installations at Inner Harbor. The Baltimore Drone Prix is…

5

Drone Racing To Take Flight In Light City Drone racing is returning back and this year the flight path will be among the installations at Inner Harbor. The Baltimore Drone Prix is expanding in a number of ways. This drone racing event will span two weekends. The pro-level racers on April 14 and 15, and youth and amateur entrants on April 21, will be participating in the competition. Eno Umoh, Co-Founder of Global Air Media, stated that he wants drone racing to feel like a recreational sport. “The plan is to have this team stick together,” he added. With the competition, the organizers are also levelling up the course. The 150ft. long drone racing course will be lit with LEDs. “There’s going to be two sections where they can walk inside the course and see what’s going on,” added Umoh. Source: https://bit.ly/2qJ7rQy About DEEPAERO DEEP AERO is a global leader in drone technology innovation. At DEEP AERO, we are building an autonomous drone economy powered by AI & Blockchain. DEEP AERO’s DRONE-UTM is an AI-driven, autonomous, self-governing, intelligent drone/unmanned aircraft system (UAS) traffic management (UTM) platform on the Blockchain. DEEP AERO’s DRONE-MP is a decentralized marketplace. It will be one stop shop for all products and services for drones. These platforms will be the foundation of the drone economy and will be powered by the DEEP AERO (DRONE) token.

Drone Racing To Take Flight In Light City

0

drone-racing-to-take-flight-in-light-city-1ba54ee17551

2018-06-05

2018-06-05 08:45:12

https://medium.com/s/story/drone-racing-to-take-flight-in-light-city-1ba54ee17551

false

235

AI Driven Drone Economy on the Blockchain

null

DeepAeroDrones

null

DEEPAERODRONES

null

deepaerodrones

DEEPAERO,AI,BLOCKCHAIN,DRONE,ICO

DeepAeroDrones

Deepaero

deepaeros

Deepaero

0

DEEP AERO DRONES

null

dcef5da6c7fa

deepaerodrones

277

0

20,181,104

null

0

null

0

c78442a37830

2018-04-13

2018-04-13 05:27:08

2018-04-23

2018-04-23 05:13:33

3

false

en

2018-04-23

2018-04-23 05:14:26

7

1ba55b3a285e

2.787736

7

0

Employee time is expensive. Let your office help make everyone more productive.

5

Distractions in the workplace Employee time is expensive. Let your office help make everyone more productive. Problem-solving is at the core of our work life. Applying critical thinking to decipher practical problems is perhaps our most distinctive human attribute. Research widely confirms that collaborative problem solving, where individuals with different backgrounds exchange information, ideas and perspectives, speeds up innovation. Organisations spend lots of resources to improve employee productivity in order to stay ahead of the competition, sometimes overlooking what’s happening in the space where the magic happens. Noise, distractions, and interruptions present big challenges in every workplace, especially in our era of open layouts, shared spaces, and no receptionists. Costly challenges. Have you noticed how quickly things get metaphysical? The guys from MonkeyUser did a great job illustrating this. Frequent, poorly-timed questions lead to context switching, which leads to errors — and annoyance. To tackle this challenge, what works for us is a combination of communication guidelines and headphone etiquette. The latter is a simple premise: if someone is wearing headphones, it’s signalling an unwillingness to be distracted. When it comes to communication, everyone needs to agree on a common approach based on the tools of choice and the expectations each conveys. For example, while emails are generally not meant for urgencies, chat often implies the need for a quicker response. There’s a whole lot that can be said to this end, it’s ultimately up to each team to define the rules they want to communicate by. Yet at Niro, we believe there’s another type of interruption that is more worrisome. If there is no front desk to greet incoming visitors, there is a big chance that said responsibility falls always on the same person, or whoever sits closer to the door or the unfortunate soul walking past. Or in open spaces, all of the above, which translates to interruptions… a lot of them. It’s not just prospects and customers we fail to greet — and that experience will trump all your investment in customer experience; it’s interviewees, deliveries, and wanderers. This situation also poses a big security risk, especially if the building where your workspace is located doesn’t have any type of security. Or worse, if it has an old security setup and no UPS. The good news is you do have room for action here. All you need is IP cameras, which may already be installed in your office. Powered up by our Niro Hub, our app can alert you every time someone walks through the door. And by you, we mean alerts can be routed to any device of your preference, to anyone you appoint, through any platform. Maybe you want a notification on a dedicated Slack channel during working hours but a phone call to your management team after hours. Maybe you want push notification alerts for deliveries, but email alerts for anything else. Keeping an eye on the prize: stable infrastructure, the pool and ping-pong tables… and the BHAG cup Or maybe you want to set up a monitor wall openly displayed like we did, to decide what to do on the spot. Beauty lies in our open, accessible and integrable platform. Employing Artificial Intelligence to do the triage means you choose how to route incoming visitors and let machines handle the noise — they are way better at multitasking that we are. We’ve all struggled to work in spaces where the surroundings are out of sync with us; improvements on this area, no matter how small, have a direct impact on employee happiness, business agility, and everyone’s safety. Sign up here to bring Niro to your workplace so it does what it’s supposed to do: help people and organisations prosper.

Distractions in the workplace

186

distractions-in-the-workplace-1ba55b3a285e

2018-05-15

2018-05-15 17:41:46

https://medium.com/s/story/distractions-in-the-workplace-1ba55b3a285e

false

593

The smart solution to monitor your business. Join us in our journey to transform and empower offices.

null

Nirovision

hello@nirovision.com

nirovision

SECURITY SYSTEMS,AI,WORKPLACE,STARTUP,CAMERAS

nirovision

Startup

startup

Startup

331,914

Eliana Bettio

Nomad Argentinean. Economist. Broadcasting live from Sydney. Product Manager @ Nirovision.

406bf6fce6b9

elianabettio

9

58

20,181,104

null

0

null

0

32881626c9c9

2018-07-19

2018-07-19 10:51:47

2018-07-19

2018-07-19 11:17:29

7

false

en

2018-09-03

2018-09-03 05:09:09

14

1ba56a961e24

8.400943

0

A guide for non-tech marketers to boost their data-based content generation and distribution.

5

Drive Your Content Marketing with Data Scraping A guide for non-tech marketers to boost their data-based content generation and distribution. 95% of B2B buyers are considering vendor-related content as trustworthy when they make purchasing decisions, based on the report of Demand Generation. Additionally, The American Content Marketing Institute also found out that 88% of B2B Enterprise Marketers are using content marketing, and the most used technique they used to generate the right content is analytics tools. This shall be the most comprehensive guide for marketers, without any prior programming knowledge, to leverage data scraping to drive their Content Marketing. We are going to talk about two things: Content Marketing & Data Scraping. The questions we will address are: What is Content Marketing and Content? What is data scraping? Why do we need Data Scraping in Content Marketing? How can we leverage data scraping to boost our Content Marketing, from creation to distribution? Let’s start with the definition of content marketing. So, Word Cloud of this Article What is Content Marketing? Content marketing is “a strategic marketing approach focused on creating and distributing valuable, relevant and consistent content to attract and retain a clearly defined audience, and, ultimately, to drive profitable customer action.” It should be operated as an ongoing business process, instead of simply a marketing campaign. The keyword in the definition of Content Marketing is “valuable”, which makes it distinguishable from conventional marketing. That is to say, Content Marketing should be beyond promotional, instead, it should be informative and educational. Nevertheless, the goal of content marketing is to drive traffic to the website, or draw the attention to the product, and improve the conversion rate, ultimately generating profit. To put it more specifically, it helps enterprises grow by improving brand awareness, lead generation and nurturing, customer engagement, retention, and loyalty. What is the “content” in Content Marketing? The content in Content Marketing mainly refers to digital content. From the format of content, it can be divided into: common content like articles, images, infographics, audio, video, and events; premium content assets includes interactive tools, ebooks/white papers, or webinars. From the way it delivered, Content Marketing is composed of blogs, social media posts, webpages, emails, podcasts, videos, and even live broadcasting. CMI found out that enterprises use 10 kinds of content tactics on average, which is as below: From another aspect — content generation, Content Marketing could break down into: Brand-generated Content (BGC): based on the in-house content team, BGC provides the audience with valuable information related to products and brands, establishing the brand as the authority on a topic. Professionally-generated Content (PGC): enterprises cooperated with professional content agencies or Key Opinion Leaders, to convey product/brand information to a much extensive audience. User-generated Content (UGC): UGC is a target of every enterprise, turning the best customers into brand advocates. And 93% of consumers find UGC helpful when making purchasing decisions. So far we have already got a comprehensive understanding of Content Marketing, now the challenge is to achieve the success of content marketing. No I will tell you how data scraping contributes to the success of content marketing. What is Data Scraping? Data scraping is a way of automating or scaling the process of aggregating information from different websites on the Internet. With the explosion of internet data, data scraping has become an important method to build up the database of a company. There are many methods to scrape data from websites, for example, we can scrape web data with Google Sheets. For non-tech marketers, the most efficient way is to choose an easy-to-use web data scraping tool. Why Do We Need Data Scraping in Content Marketing? We will answer this question by addressing the 3 difficulties most marketers facing when it comes to Content Marketing. Data-based content stands out of the crowd easily On average, every second, around 6,000 tweets are tweeted on Twitter and 854 photos uploaded in Instagram; 300 hours of videos are uploaded to YouTube every minute. The biggest challenge for an enterprise becoming an expertise on a topic is that we have to make our content outstanding in the content stream, as the digital world is flooded with content today. Data-driven content is the most effective way to be distinguished from others. 84% of B2B marketers attribute their organization’s increased success with Content Marketing to higher quality content creation. And I will elaborate on how we could generate higher quality content with data scraping below. No technological background required Over 1/3 marketers consider their organization’s stagnancy in content marketing resulting from the lack of content marketing technologies/tools, or the new systems require a steep learning curve. In the big data era, it is almost impossible to skip the tools to generate data-based content. Most of the tools are quite complicated to command for non-tech marketers. A good data scraping tool should be easy to use, even non-tech markets can build up a scraper to extract the web data they want easily. That’s why data scraping, as an important resource for aggregating big data, and the premise for data analysis, will not improve the learning burden of non-tech marketers. Cost efficient Only 22% of total marketing budget is spent on content marketing from a research of Content Marketing Institute. Many businesses have very limited budgets to invest in complex and expensive tools that can feed in streams of data to their database. Besides, few data analytic tools provide a free version, or the free version could not even meet the essential demand of an enterprise. Octoparse has analyzed the Top 5 data scraping tools in the market, there are free data scraping tools with generic functions provided for non-tech marketers. That’s to say, we don’t always have to invest in expensive tools to gather valuable intelligence. How can we leverage data scraping to boost Content Marketing? Only when our content map highlights to most desired content of one’s market audience will they be interested in spending time on the content we post. Therefore, I will talk about how data scraping helps marketers generate customer-intelligence content. Data-based Content Creation Know your audience better The content is invaluable if it is not developed with the appropriate audience in mind. We need to understand our audience and know what they really care, which topics will resonate with them. For instance, we can retrieve our product reviews from Amazon, or extract people’s discussion about our products/the relevant topics from Twitter. Thus, we can find out the content mapping with the interest of our target audience. Web Scraping Tutorial- Scraping amazon-product-reviews-and-ratings Follow the new trends / existing popular content in the field This is the essential skill for a marketer to generate content. Of course, we can rely on tools, like Buzzsumo, to tell us the most popular content, but its free version can only search data in the past year. This is not enough. A more efficient way is to build up customized scrapers for the websites which have the information we need, and feed the data extracted into our systems automatically and consistently. This is a once and for all solution, we could be informative of the new trends and popular topics in real time. Web Scraping Tutorial- Scraping tweets from Twitter Generate creative insights to lead the market trends If we can release a report or infographic that offers new or definitive market trends or issues and highlights the company’s expertise, we can establish our brand as the authority on a topic, which will bring earned media mentions and interview requests. This is an advanced skill for content marketers, which required marketing instinct in the past. But with the assistance of data scraping, we can easily achieve the target by analyzing a large quantity of data in a long period. If we want to see in-depth intelligence others haven’t discovered, we need to have our data samples differentiated from others in volume and variety. Enrich the content with data visualization Undoubtedly, data is the strongest proof for a viewpoint. Showing out the data analytics results or compose your content with infographics will enhance the persuasiveness and credibility greatly. The value in our content should be regarded as the top priority of content generation. And it will be paid off, as statistics finds out that white paper is the most likely sharing content among buyers, because of its comprehensiveness and authoritativeness. For instance, we extracted the most popular BI tools reviews and information for the TOP 20 Business Intelligence tools. Likewise, we can extract a job description on Indeed or Glassdoor, to find out that “Content Marketing” related jobs has increased over 25% from the past three years. Also, we had scraped Youtube comments and visualized the comments by WordART of the 2018 FIFA World Cup. Scraping & Visualizing: 2018 FIFA World Cup Reviews Data-Driven Content Distribution To achieve the success of content marketing, it not only requires a high quality of content, but also an effective distribution solution to convey the right content to the right audience at the right time and through the right channels. Forbes has described the Content Distribution as Winning The Last Mile In Content Marketing. Content distribution opportunities typically fall into three platform categories: Owned Media: including our websites, email, newsletters and social media accounts; Shared Media: public social media channels for marketers to post original contents; and Paid Media: Ads. Now many channels support content promotion directly when we post. On average, enterprises apply 6 channels to distribute their content. The following are Google+, SlideShare, Instagram, Pinterest, Snapchat, iTunes, Medium, and Tumblr. Map the channels with the right content Some marketers will mistakenly assume they need to post their content anywhere and everywhere to increase their chances of achieving the desired results. But the real problem of distributing content is whether the content is reaching the right audience. Based on the channels we have developed, or our competitors did, we can scrape the content and the audience engagement to find out what kind of content could draw the attention of our audience. For example, we scrape all the video information and reviews from Youtube under the topic of “World Cup 2018”, to find out the popular videos and their common points. Scraping & Visualizing: 2018 FIFA World Cup Reviews Release your content at the right time/place Many marketers may never care about the releasing time in different channels, but statistics have shown that the publishing time plays an important role in the content’s influence. Based on a study conducted by marketers of HubSpot and Search Engine Land, they found out some fascinating things about the best time to publish a blog post. If you want to get the most traffic, the best time to publish a blog post is Monday, 11 am EST; If you want to receive more comments, the best time to publish a blog is Saturday, 9 am EST; If you want more inbound links, the best publishing days are Monday and Thursday, and the best hour is 7 am EST; With the assistance of data scraping tools, we can find out the best time to publish content on different channels in our own situation — like scraping all our past content on different platforms to analyze if the time or title affects the opening rate. Conclusion Technology development brings both chancce and challenge for enterprises and marketers in today’s digital world. Enterprises should establish an ongoing strategy of their content marketing, while marketers need to leverage data scraping tools to improve their effectiveness on content generation and content distribution. Hand-picked Resources: Video: Create your first scraper without coding Video: What is web data scraping Extracting Dynamic Data With Octoparse Using Web Scraping to Improve Business Analytics and Intelligence Originally posted at: https://www.octoparse.com/blog/drive-your-content-marketing-with-data-scraping

Drive Your Content Marketing with Data Scraping

0

drive-your-content-marketing-with-data-scraping-1ba56a961e24

2018-09-03

2018-09-03 05:09:09

https://medium.com/s/story/drive-your-content-marketing-with-data-scraping-1ba56a961e24

false

1,948

Data Driven Investor (DDI) brings you various news and op-ed pieces in the areas of technologies, finance, and society. We are dedicated to relentlessly covering tech topics, their anomalies and controversies, and reviewing all things fascinating and worth knowing.

null

datadriveninvestor

null

Data Driven Investor

info@datadriveninvestor.com

datadriveninvestor

CRYPTOCURRENCY,ARTIFICIAL INTELLIGENCE,BLOCKCHAIN,FINANCE AND BANKING,TECHNOLOGY

dd_invest

Content Marketing

content-marketing

Content Marketing

34,905

Octoparse

https://www.octoparse.com/

d229ddbf79fa

octoparse

101

7

20,181,104

null

0

null

0

null

2018-07-20

2018-07-20 10:50:51

2018-07-24

2018-07-24 17:34:02

3

false

en

2018-07-25

2018-07-25 03:34:23

0

1ba56d3871d5

3.572642

11

0

Graduating early from college left me with four months of time to kick back and relax. The bucket list which had been augmenting for the…

5

Summer Intern Experience Graduating early from college left me with four months of time to kick back and relax. The bucket list which had been augmenting for the past four years finally began to shrink. The beginning of the vacation saw me full of energy and enthusiasm. However, as weeks passed, boredom started seeping in. The fact that I had not been productive for weeks made me feel slightly guilty. Before the dilemma could have started torturing me, I got a wonderful email from Neha. Yup!!! I had cleared my interviews and was eligible to be a Wingtern. I was lucky enough to secure an internship at Wingify and decided to make the most of this opportunity. Feedback from ex-wingterns assured me of the perfect work life balance at Wingify. This meant that not only could I pacify the workaholic within me but also have a great time doing so. In short, I got both the things I wanted (FOMO solved). Wuhoo!!! Naina was wrong — Wingify solves the FOMO (Watch the movie) 😜 The journey begins I pulled up my socks and began the ride as a Wingtern. Fortuitous meeting with the Director of Engineering, Ankit Jain, was pivotal in defining the course of the intern. Having assessed my profile and area of interests, Ankit gave me an opportunity to work on a project which stood at the confluence of my work in the past and what I wanted to pursue in future. I appreciated the immense faith he showed in me. I was given a research based project and was supposed to build it from ground up. The project — Churn Prediction After all the hype I have built around my project, let me not keep you in suspense any longer. The task at hand was to detect the churn probability. In other words, I was supposed to predict the chances of an account/customer to turn inactive and stop using the product. This high level description sounds overly simplistic. However, complexities arise when we consider the fact that we lacked structured data and that there were no defined parameters which determine the user’s on-line behavior. The end to end product fetches the data using appropriate crons. The collected data is cleaned and arranged in the desired format. The structured data is then analysed to find specific combination of parameters which can help determine the churn. Appropriate classification techniques are applied to generate the desired model which predicts the churn probability with satisfactory precision. Well, the above mentioned project is just the minor one. My major feat was — Mission Table Tennis 😛 Mission Table Tennis While some called me ‘TT — intern’, others preferred to call me a loafer. Well, they were not at fault, it sure appeared so 😜. A few rallies on the table and I was hooked to the game 😍. My TT sessions used to extend throughout the official working hours and the actual work began at 8 pm and extended into the wee hours. Weekends were utilized preparing for a few exams and chilling out with friends, exploring the amazing nightlife that Delhi has to offer. My initial days at table 😜 In spite of the outlook of my peers and my skewed schedule, I did manage to get some work done 😜. I religiously fulfilled my weekly deliverables and kept my manager updated about my progress. The end result was that not only did I manage to hone my TT skills but also succeeded in delivering a product that went over and beyond the initial requirements. I hope my finished product would be actively used at Wingify even in the future and I look forward to the day when it’s true success shows up on the revenue sheets. Quintessence of Wingify This experience would have been hollow if not for the wonderful people that I met here .The essence of Wingify is its people who make it a truly remarkable place to work at. Tremendous help and support by my manager Vivek Kishore and mentor Siddharth Goyal made my work a breeze. When the obstacles seemed overpowering Gaurav and Venus helped me find the grit to keep at it. I bonded over Chai and snacks with my colleagues. Had it not been for Anubhav, Mission Table Tennis would have been impossible (will learn your Personalised Serve in future 🙌). The cake that Anubhav and Richa brought for me was the best surprise I got on my birthday and it truly made my day. VM the great (Vanshika) thanks for the farewell tea and much needed support. Whenever the week draws to a close I find myself longing for the Friday night chill scenes that I used to have with Aditya Bhaiya. All these anecdotes are indelibly edged in my memory and still make me feel a part of the Wingify family. Thanks a lot for a wonderful summer!! ❤❤❤

Summer Intern Experience

25

summer-at-wingify-1ba56d3871d5

2018-07-25

2018-07-25 03:34:23

https://medium.com/s/story/summer-at-wingify-1ba56d3871d5

false

801

null

Internships

internships

Internships

8,699

Shantanu Tripathi

null

bffdebd7f6a7

shantanutripathi237

6

7

20,181,104

null

0

null

0

null

2018-06-18

2018-06-18 09:52:32

2018-06-18

2018-06-18 11:53:23

0

false

en

2018-06-18

2018-06-18 11:53:33

0

1ba77e4bc4f4

2.033962

2

0

Okay, this story I start with question that is “When it becomes machine learning project?”. As a deep learning enthusiast, this question is…

3

When It Becomes Machine Learning Project? Okay, this story I start with question that is “When it becomes machine learning project?”. As a deep learning enthusiast, this question is dazzling around to my mind for separating a kind of project that need machine learning as a solution or just a standard algorithm as a solution. To answer this question I watched several videos about machine learning and I found Prof. Yaseer Abu Mustafa from Caltech University online lecture. AfterI watched Prof. Yaseer, I conclude the answer from the lecture video as a question for filtering does the project is machine learning project. Question for filtering Machine Learning Project or Not Do we have data (a lot of data usually)? Does the data have a pattern on it? Can we pin it down the data mathematically? These 3 question is like feasible test for a machine learning project. If you have a project for machine learning and do not pass all of these 3 question than you are not in business, but if you pass these 3 question than you are on business. Some of my friend told me that machine learning effort is 80% in data acquiring or data preparation. I don’t know if these percentage is correct or not, but in my perspective that percentage probably correct. Because preparing data in machine learning is usually the most time consuming after training the classifier. Okay, lets answer the second question, “Does the data have a pattern on it?”. This question is related to the condition of data, which means that if the data is image type than the image must has a pattern on it. For example, if we want have dog or cat classifier than an image must have dog in every image as dataset. Pattern mean, it has a same structure of some object and from that structure machine learning will learn and generalize it when see something similar. How about question number three? “Can We Pin it down the data mathematically?” This the last question that machine learning algorithm do its job well other than another algorithm. But what this question mean anyway? Okay let me explain it, every dog has a different face, different type of tail (long tail or short tail), and many more thing. But in general every dog has a ear, four legs, a tail and other thing that we call it as a structure of dog. This structure has a pattern that with normal algorithm we can not create a specific code to determine a dog with other object just be reading the pixel on image. Because of this condition object detection on image is very hard to do with standard algorithm. In other hand, human will easily do this kind of task because human is pattern reader. Soo, if we want to make computer do the thing that we do, we need to teach them how to learn to understand the abstract meaning of an object so they can perceive pattern like we do. Okay, that’s a three question for feasible test or determine if your project or your problem need machine learning as a problem solving. Let me know your opinion on this topic by commenting in this article.

When It Becomes Machine Learning Project?

15

when-it-becomes-machine-learning-project-1ba77e4bc4f4

2018-06-18

2018-06-18 11:53:34

https://medium.com/s/story/when-it-becomes-machine-learning-project-1ba77e4bc4f4

false

539

null

Machine Learning

machine-learning

Machine Learning

51,320

Lalu Erfandi Maula Yusnu

Web Developer & Deep Learning Enthusiast

4e5cced60ca9

nunenuh

6

25

20,181,104

null

0

null

0

null

2018-06-07

2018-06-07 12:08:51

2018-06-11

2018-06-11 03:01:01

7

false

en

2018-06-11

2018-06-11 03:01:01

8

1ba7b1d2ee3

3.276415

9

0

1

The machine learning community is powered by open source releases from the top organizations. This past week was highlighted by two of them…

5

Google’s Open Source YouTube Dataset, Uber’s Awesome Location Visualization Tool, A Free Tool that works with any Library and Algorithm, and other ML Developments! The machine learning community is powered by open source releases from the top organizations. This past week was highlighted by two of them — Google and Uber. Under the AVBytes umbrella, we love bringing you all the top open source releases so you can learn, replicate and even expand on them! Highlights from the past week: Google released an updated version of the popular ‘YouTube-8M’ dataset, Uber unveiled a pretty visualization tool for analyzing location data, a ML project that detects crowd violence, and more below! In addition to this, Databricks (founded by the creators of Apache Spark) have developed a platform called MLflow and it works with any tool, language, algorithm and library! We liked the concept and are looking forward to them adding more components like monitoring the progress of your ML models. This is another example of the ML community giving back to everyone by making such a breakthrough tool open source. Start your week with these game changing developments from last week and to get a daily dose of AVBytes, subscribe here to get it directly delivered to your inbox! Click on each title to read the full article. Google Open Sources Approach to Visualize Large and High Dimensional Datasets using tSNE: Have you ever come across high dimensional datasets? An intern at Google has revealed an awesome new approach that uses tSNE to visualize large and high dimensional data in real-time and in your browser itself! ‘Eye in the Sky’ is a Machine Learning Project that Detects Violent People in Crowds: A brilliant use case of computer vision in machine learning — a system that uses human pose estimation to identify violent individuals in a crowd! Check out the research paper, video and other details inside. Google has Released an Updated Version of its Open Source YouTube Dataset: Google has released an update to it’s popular ‘YouTube-8M’ open source dataset. It comes with richly labelled data and is ready to be worked on. Download it NOW and start practising! MLflow — An Open Source Machine Learning Platform that works with any Library, Algorithm and Tool!: MLflow is an open source machine learning platform that works with any ML library, algorithm, deployment tool or language. You don’t need to worry about any restrictions, MLFlow will sort all that for you! Uber’s Kepler.gl is a Wonderful Open Source Tool for Analyzing Location Data (No Coding Required): Ever wondered how companies like Uber and Airbnb use location data? You can find out for yourself with kepler.gl, Uber’s open source visualization tool! We found it extremely easy to use and it doesn’t even require coding. Get started right NOW using the already uploaded sample datasets. Apple Launches ‘Create ML’ for Easy Machine Learning Model Training on Macs: Create ML is a GPU accelerated tool for training models on Macs. It is remarkably easy to use and can even be turned into a drag-and-drop interface so you don’t need to know coding to use it. If you’re a Mac user, will you consider transitioning to this? Source: Dice The above AVBytes were published from 4th to 10th June, 2018. This was one of the most intriguing weeks in terms of coverage we have seen on AVBytes. Apple’s conference also saw them release an update to their popular ML framework — Core ML 2. Out of all these happenings, which one(s) caught your eye?

Google’s Open Source YouTube Dataset, Uber’s Awesome Location Visualization Tool, A Free Tool that…

69

avbytes-ml-news-google-uber-databricks-mlflow-1ba7b1d2ee3

2018-06-11

2018-06-11 21:16:03

https://medium.com/s/story/avbytes-ml-news-google-uber-databricks-mlflow-1ba7b1d2ee3

false

590

null

Machine Learning

machine-learning

Machine Learning

51,320

Team AV

This is the Official Handle of Analytics Vidhya team. For more articles, check out the Analytics Vidhya website and Medium publication of Analytics Vidhya.

c7c686fcd4b

analytics

1,907

124

20,181,104

null

0

null

0

cc02b7244ed9

2017-10-11

2017-10-11 05:45:24

2017-10-11

2017-10-11 05:48:05

0

false

en

2017-10-11

2017-10-11 05:48:05

11

1ba7d7e62a41

2.090566

0

PRODUCTS & SERVICES

5

Tech & Telecom news — Oct 11, 2017 PRODUCTS & SERVICES Connectivity Better perspectives for US wireless carriers in 3Q17, as market has cooled down and all of them would expect lower churn and positive service revenue growth. Even new entrant Comcast is behaving rationally, having recently killed its $200 subsidy on high end phones. AT&T and Verizon seen as the key winners (Story) Video It seems Apple has started to spend the $1bn it announced for acquiring / developing original video content, with a deal with S Spielberg’s Amblin TV and NBCU to make 10 new episodes of “Amazing Stories”, a cult 1980s Sci Fi / Horror show. Budget is expected to be “significantly more” than $5m per episode (Story) Internet of Things IoT progressively becoming a central theme in IT solutions. Dell just announced a $1bn effort over the next 3 years, in a specialised IoT division that will build products, create partnerships and do research. Dell’s vision is to complement “hyper-cloud” services (e.g. AWS) with solutions that process data closer to the edge (Story) Blockchain Ripple, a global leader in blockchain, offering cross-border payment solutions based on its own cryptocurrency, XRP, aims to dominate financial blockchain apps, and has said it has $15bn reserves to acquire or partner with rivals. Ripple’s current product makes interbank transactions much more agile (15sec vs. 3 days…) (Story) HARDWARE ENABLERS Network N McKeown, a researcher in SDN / NFV and now chief scientist of the specialised startup Barefoot Networks, has a vision that the economic model to justify investments in 5G will rely on network virtualisation tech being commercially viable and deployable, as a tool to enable application to a variety of use cases (Story) Components Intel has developed a new 17-qubit quantum computing chip, in partnership with a Dutch specialist, QuTech. This is seen as a (still very early) step to build “industrial” QC products that can be ordered / delivered. Intel claims that their expertise with processing materials is a key advantage also for these futuristic chips (Story) Meanwhile, Nvidia has announced a trial in Germany next year with their “Pegasus” processors, the first computer chips to control completely driverless cars (“the end-game for self-driving cars”, according to the company). The chips will be installed in a fleet of electric delivery vans from the German post company (Story) SOFTWARE ENABLERS AI talent being so scarce, companies are perceiving value in getting physically closer to leading AI talent and universities, so Montreal is becoming a global hub for AI research. As an example, Thales will open a new facility to develop AI tech applicable to industries where they’re focused (including aerospace, defence) (Story) VENTURE CAPITAL New deals announced for SoftBank Vision Fund, with an underlying theme related to AI systems in general and self-driving cars in particular. One of them involves leading a $93m investment in Petuum, a Pittsburgh startup building an AI software platform that can be used for a diversity of industries (e.g. health) (Story) A second deal is more specifically related to self-driving cars, with SoftBank leading a $164m in digital mapping startup Mapbox, now expanding efforts in autonomous cars and Artificial Reality. More than 900K developers are already using Mapbox’s mapping and location-search platform, including Snap and GE (Story) Subscribe at https://www.getrevue.co/profile/winwood66

Tech & Telecom news — Oct 11, 2017

0

tech-telecom-news-oct-11-2017-1ba7d7e62a41

2018-04-23

2018-04-23 02:11:56

https://medium.com/s/story/tech-telecom-news-oct-11-2017-1ba7d7e62a41

false

554

The most interesting news in technology and telecoms, every day

null

Tech / Telecom News

ripkirby65@gmail.com

tech-telecom-news

TECHNOLOGY,TELECOM,VIDEO,CLOUD,ARTIFICIAL INTELLIGENCE

winwood66

Quantum Computing

quantum-computing

Quantum Computing

1,270

C Gavilanes

food, football and tech / ripkirby65@gmail.com

a1bb7d576c0f

winwood66

605

92

20,181,104

null

0

null

0

null

2017-10-04

2017-10-04 09:27:56

2017-10-04

2017-10-04 09:53:59

1

false

en

2017-10-04

2017-10-04 09:53:59

2

1ba8a7505315

2.267925

1

0

There are times when we’re all in lack of productivity in our work lives. You have a staff relying on you and a million deadlines to meet.

5

Start working smarter! These 9 Tips will make you super productive in your business There are times when we’re all in lack of productivity in our work lives. You have a staff relying on you and a million deadlines to meet. http://letz.do/ At any moment in a given day, we have push notifications, text messages, email alerts, phone calls, tweets, and even real humans vying for our attention. Not only are we busy in the present, but also our minds are often consumed with the future and all of the decisions we have to make. But in times like these shutting down is not an option. Today we share with you several essential tips that will help you get to that level of productivity and efficiency you wished for. Prioritize Prioritize and only do what matters. Avoid busy work and going in all directions. Say no to stuff that won’t move the needle. 2. Plan your to-do list Don’t overestimate the number of things you can get done in a day. Instead of creating a long list of to-dos and get frustrated if you don’t finish, try to plan to get one large task, three with medium importance, and five small tasks done every day. 3. Quit multitasking While we tend to think of the ability to multitask as an important skill for increasing efficiency, the opposite may, in fact, be true. Attempting to do several tasks at once can result in lost time and productivity. Instead, make a habit of committing to a single task before moving on to your next project. 4. Schedule in a calendar Create and manage your schedule in a calendar. If it’s not on the calendar, don’t do it. Don’t assume that you will be working on something unless it is on your calendar. 5. Put a limit on all meeting Set a time expectation and limit for all team meetings. Typically, one-hour meetings are the max. Thirty minutes is way better unless it’s a long brainstorming session, but even then, break it out into chunks. 6. Reduce meetings Cut down on unplanned team meetings. You can pull people into the meeting room to brainstorm, but just make it clear this is what you are doing. Don’t turn these sessions into long meetings. 7. Don’t let emails rule you Do not refresh or check for new messages, or have an email opened all the time. It will kill your productivity. 8. Keep it short Learn the art of the short email — two to three sentences or paragraphs max. Check how it looks on mobile. If you scroll a lot, it’s too long. Use the least number of words possible and iterate on the emails before sending them out. 9. Take a break Maybe It sounds unreasonable, but taking scheduled breaks can actually help improve concentration. Running from back-to-back meetings is not productive because you get tired and lose focus. Block off time on your calendar or to-do list and take breaks. What are your best work-related tips for productivity? Have you found the secret to increasing your productivity in the office? Feel free to share it with Lucy and let us know what worked for you! Free download link is here: www.letz.do/get-free ENJOY PRODUCTIVITY!

Start working smarter! These 9 Tips will make you super productive in your business

1

start-working-smarter-these-9-tips-will-make-you-super-productive-in-your-business-1ba8a7505315

2018-05-09

2018-05-09 22:50:41

https://medium.com/s/story/start-working-smarter-these-9-tips-will-make-you-super-productive-in-your-business-1ba8a7505315

false

548

null

Productivity

productivity

Productivity

96,541

Letz.do

Letz is the world’s first ACTIVE personal productivity assistant! You can Enjoy Productivity with Letz and DOWNLOAD it for FREE here: www.letz.do

d75ad7f2e1ea

Lucy_Letz

45

167

20,181,104

null

0

null

0

null

2018-08-01

2018-08-01 21:50:25

2018-08-01

2018-08-01 21:51:55

3

false

en

2018-08-17

2018-08-17 18:10:37

5

1bac0ef388e2

4.327358

4

0

Augmented Reality: The Hottest Tech Buzzword

5

Augmented Reality is the hottest buzzword in Tech today. What is it? Augmented Reality: The Hottest Tech Buzzword Software companies are hopping on the AR Bandwagon — but what is it? From Wiki: “Augmented Reality (AR) is an interactive experience of a real-world environment whereby the objects that reside in the real-world are “augmented” by computer-generated perceptual information, sometimes across multiple sensory modalities, including visual, auditory, haptic, somatosensory, and olfactory.” In short, computer-generated information is laid on top of a real-world experience. For example, in the PokemonGo app, the user can use see and hear Pokemon, like Pikachu, through their screen while seeing the actual landscape in front of them. These animated creatures are not present in reality but are shown and heard as a simulation. The sensory information can either add something to reality, known as Constructive AR (like the Pokemon), or it can mask reality, known as Destructive AR. Augmented Reality should not be confused with Virtual Reality: AR alters a user’s perception of reality and VR replaces reality with an entirely computer-generated environment created by software companies. An example of the destructive type of experience: a hole in the ground where the app displays underlying piping. This information wouldn’t have been available to the user’s eye without this technology. Why is AR so popular now? The first functional AR systems were invented in the early 1990s and used largely in the entertainment and gaming businesses. In 2012, it became integrated into electronic devices and being developed by software companies with increasing frequency. It was and is still used in toys to enhance their entertainment and play value. The advertising market is becoming more cluttered and many people opt out of receiving marketing messages: AR brand experiences provide new ways to engage consumers. A study by ad tech firm YuMe and research firm Nielsen found that AR and VR elicited 27% higher emotional engagement than a 2D environment, and 17% higher emotional engagement than a 360-degree video on a flat screen. AR can be used to interact with customers and understand more about their behaviors and preferences. Software companies and brands that understand the possibilities of Augmented Reality driven advertising have even started creating their own social media platforms. What are Some Apps that use AR? Snapchat: Social media apps such as Snapchat and Instagram are incorporating AR technology to keep users engaged. Snapchat is a social media application that allows users to send pictures and short videos to other users. Snapchat uses AR through filters: users can add things like dog ears, cat whiskers, and animated backgrounds to their photos and videos. Pokemon Go: Launched in July 2016, Pokemon GO’s downloads reached 50 million installations before the end of the month. This app guides users on walks around cities to collect Pokemon (animated animals). The app imposes Pokemon images onto your Smartphone screen. Google Maps: a location finder which features transit, car, bicycle, and walking transportation directions. AR is showcased in Google Street View with 360-degree video of locations around the world. Users can search for a location and view and navigate around the location with AR 360 degree video. Quiver: (formerly known as ColAR Mix) brings children’s coloring books to life with animated images that spring directly from the pages. Several free coloring books are available on the Quiver website to download and print. Once the pages are colored in, children can use a smartphone camera to record the page and view an augmented moving image. The app also features accompanying music with the animated images. What are some practical uses for AR? Travel: New AR apps allow users to get a glimpse of the “hotel experience” prior to their stay. Through apps like Radisson Hotels’, customers are able to take virtual tours of hotel rooms and get a 360 view of rooms, suites, meeting facilities, restaurants, and pool areas. Home Improvement: Some retail stores have released AR marketing apps that provide customers with an interactive experience. For example, Home Depot uses AR to allow customers to see paint colors in their home before purchasing paint. Customers choose a paint color and then hold their smartphone or tablet up to a wall and the paint color will appear on the wall. The Home Depot app also allows users to test doors, windows, light fixtures, furniture, and more. Car Finder: Augmented Car Finder is an app designed to guide users to their parked car. Users can set their car location when parking, then when the user is returning to their car, the app will show augmented arrows guiding them to the location. Users can utilize this app technology for any object, including seats in a theatre or concert, or lost keys. What does the future hold for AR and VR? Software companies are banking on the prediction that the value of AR and VR, virtual reality, is predicted to hit $150bn by 2020. The evolution of the two technologies has resulted in personalized, mixed-reality experiences for consumers. As AR and VR continue to evolve, they are expected to impact the marketing industry in many ways. Below are a couple of examples of what the future may have in store for AR. Product Trials: Makeup Genius is a mobile app that allows customers to scan a product in the store and then use their smartphone camera to try on augmented L’Oréal Paris products. The app has already been downloaded more than 16 million times. https://www.youtube.com/watch?v=zbBJfrkZRDI Healthcare: AR could change the way doctors are trained, where doctors can be trained, and who can become a doctor. Proximie, co-founded by Dr. Nadine Hachach-Haram, is a current AR platform that allows surgeons to connect with medical professionals around the world and guide them through procedures. Other uses involve the surgeon wearing an AR headset such as Microsoft’s HoloLens, allowing them to see digital images and other data directly overlaid on their field of view. By Mary MacPherson, Digital Marketing Manager https://www.essentialdesigns.net/ Mary is a popular Vancouver DJ and has been working in the high tech field since her inception as a web developer at Newbridge Networks in 1996.

Augmented Reality is the hottest buzzword in Tech today. What is it?

184

augmented-reality-is-the-hottest-buzzword-in-tech-today-what-is-it-1bac0ef388e2

2018-08-17

2018-08-17 18:10:37

https://medium.com/s/story/augmented-reality-is-the-hottest-buzzword-in-tech-today-what-is-it-1bac0ef388e2

false

1,001

null

Augmented Reality

augmented-reality

Augmented Reality

13,305

Essential Designs

Essential Designs is a team of custom application developers in Vancouver, Calgary and Toronto. We specialize in custom software and app development.

2aa0ea62d6cd

essentialdesign

1

20,181,104

null

0

null

0

null

2018-07-23

2018-07-23 11:02:30

2018-07-23

2018-07-23 11:41:14

6

false

en

2018-07-23

2018-07-23 14:01:07

0

1bacc7d68dff

4.172642

5

0

So imbalance data is one of the most commonly faced problem in data mining. If you are new to machine learning or data mining, probably you…

4

Dealing with Imbalance Data : So imbalance data is one of the most commonly faced problem in data mining. If you are new to machine learning or data mining, probably you have not faced imbalance data because instructors use data sets which are balanced so as to concentrate on learning . But, in real world scenarios you would face class imbalance problems. What is Imbalance Data ? In simple terms , in binary classifiers , Imbalance of data is dominance of one class over other of target variable. Class Imbalance occurs in datasets pertaining multiclass classification as well. Examples: User churn in telecom industry. Detection of rare diseases. Detecting Credit card Fradulant transactions. In above examples you can notice that number of data points of right class are very less as compared to other. Balanced Data : Imbalanced Data: Ways to Handle Imbalance data : 1. Using more appropriate error matrix 2. Resampling strategies (Data Level ) 3. Algorithmic Techniques So lets see how these ways are used to handle imbalance of data. 1.Error Matrix: Consider Example: Consider 100 sample test points. Out of which class A contains 95 sample points and class B contains 5 sample points. If Model M1 predicts all the points in class A, still our model accuracy will 95%.But our model will not be able to capture events of right class. To Overcome this we have to use proper error matrix . More details on Error Matrix are available in my previous article. Please do read that. 2. Re-sampling techniques : In re-sampling of data either we reduce the proportion of dominant class which is under-sampling or we increase proportion of minority class which is called as oversampling. However most successful approaches uses both oversampling and under-sampling together. 2.1 UNDER-SAMPLING : There are many ways of under-sampling. Out of which some are mentioned below 2.1.1 Random under-sampling 2.1.2 Cluster centroid Under-sampling 2.1.3 Tomek Link 2.1.1 Random Under-sampling : Random under-sampling is very simple and intuitive under- sampling technique . Method works by randomly choosing the samples from dominant class. There are two major drawbacks of these technique: 1. Major drawback of this technique is that we eliminate samples randomly, which may lead to loss of potential information . 2. The purpose of machine learning is for the classifier to estimate the probability distribution of the target population. Since that distribution is unknown we try to estimate the population distribution using a sample distribution. Statistics tells us that as long as the sample is drawn randomly, the sample distribution can be used to estimate the population distribution from where it was drawn. 2.1.2 Cluster centroid under-sampling : clusters of majority class and replace that cluster with centroid of that cluster. So we undersample majority class by forming clusters and replacing it with cluster centroids. For example: Majority class : 100 samples Minority class : 10 samples Here , in this case we can form 10 clusters of majority class and replace 100 points with 10 data points i.e by cluster centroid. 2.1.3 Under-sampling using Tomek links: Tomek link pair has two opposite class data points who are their own nearest neighbors. Main idea is to separate minority and majority class . Suppose , d(A,B) : distance between two data points A & B Then, a(A,B) is Tomek link if and only if There is no such point ‘C’ , such that, d(A,C) < d(A<B) or d(B,C) < d(A,B) If pair of samples form tomek link then either one of the sample is noise or both are placed at border. As under-sampling technique we eliminate majority class point , while as part of data mining we eliminate both points. 2.2 Over-sampling: Some of over- sampling techniques are mentioned below: 1. Random oversampling 2. SMOTE (Synthetic Minority Oversampling Technique) 2.2.1 Random Oversampling: In random oversampling technique we replace the samples with existing minority samples. In simple terms we can say that we just do multiplication of existing Minority class. Major drawback of this technique us that , this technique is highly prone to over fitting. Hence to overcome we use SMOTE for oversampling. 2.2.2 SMOTE : In this technique we calculate difference between sample under consideration and its nearest neighbors. Once the distance is calculated we multiply that with the number between 0 and 1. We add it to sample under consideration. Which gives us new sample point for minority class . Depending upon the amount of oversampling required , neighbours from k-NN are randomly chosen. 3. Algorithmic approach: 3.1 Cost sensitive approach 3.2 Choice of algorithm 3.1 Cost sensitive training: It is called as penalised training. It can be done by penalizing for wrong prediction in minority class. It can be done by customizing your error matrix. Example: (10*False –ve + 1* false +ve)/6 3.2 Choice of algorithm: Ensemble methods are found to good for handling imbalance data. Random forest is found to be good at handling imbalance data. sklearn’s implementations of these algorithms provides option to handle imbalanced data-set by setting the class_weight parameter.

Dealing with Imbalance Data :

107

dealing-with-imbalance-data-1bacc7d68dff

2018-07-23

2018-07-23 14:01:07

https://medium.com/s/story/dealing-with-imbalance-data-1bacc7d68dff

false

854

null

Machine Learning

machine-learning

Machine Learning

51,320

aditya patil

Data science aspirant.

673fb822ada9

patiladitya81295

6

5

20,181,104

null

0

null

0

ed71a2a1cfa3

2018-02-12

2018-02-12 19:50:13

2018-02-12

2018-02-12 20:02:18

1

false

en

2018-02-12

2018-02-12 20:04:59

0

1bad50da9fab

1.377358

0

30 years ago Carlsberg asked us to develop a system for inspecting beer crates.

2

Probably the best vision system in the world 30 years ago Carlsberg asked us to develop a system for inspecting beer crates. The breweries wanted two different inspections in order to avoid production stops on the bottling lines, and to ensure that the right crates were used. It was not good if a crate from a competing brewery was allowed to enter the line and eventually come out in the other end with beers in a wrong crate. On arrival the crates are automatically emptied and washed. The next stage is inspection where a camera looks down in the crate to find stuck bottles or debris. Also the top of the crate is analysed to find broken handles. A second camera identifies the brand name on the side of the crate. Getting wrong crates back to the bottling line is a very Danish problem. Our standard 30 piece plastic crate is used by everyone. Abroad it is very different. The drivers delivering the beer crates to retail will only take their own crates back. Getting wrong crates back to the bottling line is a very Danish problem. Our attempt to sell the crate inspection in UK and Germany was a total failure. It took a long time before I realized why. The crate inspection is quite difficult. The main challenges are all the dents and defects that are acceptable, and the small defects that are critical. An aligned but broken handle will only show a minute crack. It can cause severe pain when lifting the crate. If the handle opens and closes again it can take a good bite of your hand. We delivered more than 20 systems over the years, but today most beers are delivered in cans. There is not much need for crate inspection these days, but… It is probably the best beer crate inspection system in the world.

Probably the best vision system in the world

0

probably-the-best-vision-system-in-the-world-1bad50da9fab

2018-06-04

2018-06-04 10:21:09

https://medium.com/s/story/probably-the-best-vision-system-in-the-world-1bad50da9fab

false

312

With more than 35 years in the vision industry, JLI Vision specialize in development, manufacturing and installation of computer vision systems for industry and laboratories.

null

jlivision

null

JLI vision

hb@JLIvision.com

jli-vision

null

Vision Systems

vision-systems

Vision Systems

4

Henrik Birk

null

a6fcf4d56cb1

hb_33633

0

1

20,181,104

null

0

null

0

null

2017-11-23

2017-11-23 09:59:30

2017-11-23

2017-11-23 10:05:44

1

false

en

2017-11-25

2017-11-25 04:20:03

0

1baf9b9b92bb

1.275472

0

On Nov. 20, fifth day of the 19th China High-Tech Fair 2017, Allchips held the sharing session “China’s Electronic Components Situation &…

4

3 Advise for New Startups and IC E-business Owners By the VP of Allchips On Nov. 20, fifth day of the 19th China High-Tech Fair 2017, Allchips held the sharing session “China’s Electronic Components Situation & AI Creatively Developing.” In this session, challenges that traditional IC e-commerce faced, and new demands of products, services from IOT and AI companies have been discussed. According to Rainman Wang, VP of Allchips Ltd.: “Allchips’ Intelligent BOM service, based on a combination of big data and AI, enables its registers get fast quotations within seconds, budget new projects, and purchase more efficiently. The users can also upload a whole Bill of Materials using one button, then correct a part number, and price will be shown automatically.” Mr. Wang, from Allchips Ltd. Some Advise for New Startups and IC E-business Owners by Mr. Wang First, pay particular attention to the supply chains, especially when the new chips for IOT and AI have been very popularly needed in the market. Many big traditional factories are trying to make the transition. They need not only basic materials, but also creative IOT chips. Second, innovation must consider IOT and big data. The future needs must analyze big data, so that your supply chain and purchasing platform is intelligent enough to know the end users well. Make your end users feel more efficient and convenient. Third, offer the best service. There isn’t any retail or trading business that can survive without a good service. You have to be prepared for the industrial chain service, oversea supply chain service, and project design in order to catch up with the new demands of customers. If you are offering basic chips only, the platform would fall behind soon. China’s Electronic Components Situation & AI Creatively Developing

3 Advise for New Startups and IC E-business Owners By the VP of Allchips

0

3-advise-for-new-startups-and-ic-e-business-owners-by-the-vp-of-allchips-1baf9b9b92bb

2018-05-29

2018-05-29 21:58:47

https://medium.com/s/story/3-advise-for-new-startups-and-ic-e-business-owners-by-the-vp-of-allchips-1baf9b9b92bb

false

285

null

Supply Chain

supply-chain

Supply Chain

6,262

Felix Law

electronic components online service

167fc38379d4

Me120

0

6

20,181,104

null

0

null

0

4fe64d8f2902

2017-10-18

2017-10-18 15:19:59

2017-10-18

2017-10-18 20:19:48

4

false

en

2017-10-19

2017-10-19 18:06:04

2

1bb180cf0bd3

7.107547

9

0

By Sergey Nikolenko, Chief Research Officer at Neuromation

4

Neuroplasticity By Sergey Nikolenko, Chief Research Officer at Neuromation Sergey’s a researcher in the field of machine learning (deep learning, Bayesian methods, natural language processing and more) and analysis of algorithms (network algorithms, competitive analysis). He has authored more than 120 research papers, several books, courses “Machine learning”, “Deep learning”, and others. Extensive experience with industrial projects (Neuromation, SolidOpinion, Surfingbird, Deloitte Analytics Institute). This article unpacks how certain parts of the brain can learn to perform tasks they weren’t originally designed to do. Neuroplasticity is another part of this issue. Scientists conducted experiments demonstrating how different areas of the brain can easily learn to do things for which they’re seemingly not designed. Neurons are the same everywhere, but there are areas in the brain responsible for different things. There’s the Broca area responsible for speech, an area responsible for vision (actually, a lot of areas — vision is very important for humans), and so forth. Nevertheless, we can break down these notional biological borders. This man is learning to see with his tongue. He attaches electrodes to his tongue, puts a camera on his forehead, and the camera streams an image on the electrodes pricking his tongue. People stick that thing on them and walk around with it for a few days, with their eyes open, naturally. The part of the brain that receives signals from the tongue starts to figure out what’s going on — this feels a lot like something that comes from my eyes. If you abuse somebody like that for a week and then blindfold him, he’ll actually be able to see with his tongue! He is now able recognize simple forms and doesn’t bump into walls. Image credit Brainport Image credit Juan Antonio Martinez Rojas The man in this photo has turned into a bat. He’s walking around blindfolded, using an ultrasonic scope whose signals reach his tactile neurons through his skin. With a sonar like this, a human being can develop echolocation abilities within a few days of training. We do not have a special organ that can discern ultrasound, so you have to attach a scope to your body. However, we can relatively easily learn to process this information, meaning that we can walk in the dark and not bump into any walls. All of this shows that the brain can adapt to a very large number of different data sources. Hence, the brain probably has a “common algorithm” that can extract meaning from whatever it takes in. This common algorithm is the Holy Grail of modern artificial intelligence (a recent popular book on machine learning by Pedro Domingos was called The Master Algorithm). It appears as though deep learning is the closest we have come to the master algorithm of all the things done in the field up until now. Naturally, one has to be cautious when making claims about whether all of this is like what the brain does. “Could a neuroscientist understand a microprocessor?”, a recent noteworthy article, tries to elucidate how effective current approaches in neurobiology are at analyzing a very simple “brain”, like a basic Apple I processor or Space Invaders on Atari. We will return to this game soon enough and won’t go into much detail about the results but we do recommend reading the paper. Spoiler alert: modern neurobiology couldn’t figure out a single thing about Space Invaders. Feature extraction Unstructured information (texts, pictures, music) is processed in the following way: there is raw input, then features that bear content take shape, and then classifiers are built based on those features. The most complicated part of this process is understanding how to pick good features out of unstructured input. Up until recently, systems for processing unstructured information have worked as follows: people have attempted to select good features manually and then assess the quality of relatively simple regressors and classifiers based on these features. Take Mel Frequency Cepstral Coefficients (MFCC), which had been commonly used as features in speech recognition systems, for example. In 2000, the European Telecommunications Standards Institute defined a standardized MFCC algorithm to be used in mobile phones; all of these algorithms were laid out by hand. Up until a certain point, manually-extracted features dominated machine learning. For instance, SIFT (Scale Invariant Feature Transform), which enables one to detect and describe local features in images based on Gabor filters and the like, was commonly used in computer vision. Overall, people have come up with many approaches to feature extraction but still cannot duplicate the brain’s incredible success. Moreover, the brain has no biological predetermination, meaning that there are no neurons genetically created only for producing speech, remembering people’s faces, etc. It looks like any area of the brain can learn to do anything. Regardless of the brain’s activity, naturally, we would like to learn to select features automatically to create complex AI and large models containing neurons linked to one another for transmitting signals containing all sorts of different information. Most likely, humans lack the resources necessary to develop the best possible features for images or speech manually. Artificial neural networks When Frank Rosenblatt introduced his perceptron, everyone started imagining that machines would become truly smart any day now. His network learned to recognize letters on photographs, which was very cool for the late 1950s. Very soon after, neural networks made up of many perceptrons were developed; they could learn with backpropagation (a method used to calculate the gradient descent called the backward propagation of errors). Basically, backpropagation is a method used to calculate gradients or error functions. The idea of automatic differentiation was floating around back in the 1960s even, but Geoffrey Hinton, a British-Canadian computer scientist who has been one of the leading researchers on deep learning, rediscovered backpropagation and expanded its scope. Incidentally, George Boole, one of the founders of mathematical logic, was Hinton’s great-great-grandfather. Multi-layer neural networks were developed in the second half of the 1970s. There weren’t any technical barriers in place at that time. All you had to do was take a network with one layer of neurons, then add a hidden layer of neurons, and then another. That got you a deep network, and, formally speaking, backpropagation works in exactly the same way on it. Later on, researchers started using these networks for speech and image recognition systems. Then recurrent neural networks (RNN), time delay neural networks (TDNN), and others followed; however, by the end of the 1980s it became evident that there were several significant problems with neural network learning. First off, let us touch upon a technical problem. A neural network needs good hardware to learn to act intelligently. In the late eighties and early nineties, research on speech recognition using neural networks looked something like this: tweak a hyperparameter, let the network train for a week, look at the outcome, tweak the hyperparameters, wait another week, rinse, repeat. Of course, these were very romantic times, but since tuning the parameters in neural networks is nearly as important as the architecture itself, too much time or too powerful hardware was needed to achieve a good outcome for each specific task. As for the core problem, backpropagation does work formally, but not always in practice. For a long time, researchers weren’t able to efficiently train neural networks with more than two hidden layers due to the vanishing gradients problem: when you compute a gradient with backpropagation, it may decrease exponentially as it progresses from the output to input neurons. The opposite problem — exploding gradients — would crop up in recurrent networks; if one starts to unravel a recurrent network, the gradient may spin out of control and start growing exponentially. Eventually, these problems led to the “second winter” of neural networks, which lasted through the 1990s and early 2000s. As John Denker, a neural networks researcher, wrote in 1994, “neural networks are the second best way of doing just about anything” (the second half of this quote isn’t as well-known:.”…and genetic algorithms are the third”). Nonetheless, a true revolution in machine learning occurred ten years ago. In the mid-2000s, Geoffrey Hinton and his research group discovered a method of training deep neural networks. Initially, they did this for deep belief networks based on Boltzmann machines, and then they extended this approach to traditional neural networks. What was Hinton’s idea? We have a deep network that we want to train. As we know, layers close to the network’s output can learn well using backpropagation. How can we train what’s close to the input, though? At first, we will train the first layer by unsupervised learning. After that, the first layer will already be extracting some features, looking for what the input data points have in common. After doing that, we pre-train the second layer, using results of the first one as inputs, and then the third. Eventually, once we’ve trained all the layers, we’ll use the system as a first approximation and then fine-tune the resulting deep network to our specific task by using backpropagation. This is an excellent approach… and, of course, it was first introduced back in the seventies and eighties. However, much like regular backpropagation, it worked poorly. Yann LeCun’s team achieved great success in the early 1990s in computer vision with autoencoders, but, generally speaking, their method didn’t work better than solutions based on manually-designed features. In short, Hinton can take credit for making this approach work for deep neural networks (and it would be too long and complicated to explain exactly what he did). However, researchers had sufficient computational capabilities to apply this method by the end of the 2000s. The main technological revolution occurred when Ruslan Salakhutdinov (also advised by Hinton) managed to shift the training of deep networks to GPUs. One can view this training as a large number of relatively independent and relatively undemanding computational processes, which is perfect for the highly parallel GPU architectures, so everything started working much faster. By now, you simply have to use GPUs to train deep learning models efficiently, and for GPU manufacturers like NVIDIA deep learning has become a primary application that carries the same weight as modern games. Take a look at CEO NVIDIA’s pitch here.

Neuroplasticity

83

neuroplasticity-1bb180cf0bd3

2018-06-04

2018-06-04 23:38:31

https://medium.com/s/story/neuroplasticity-1bb180cf0bd3

false

1,698

Distributed Synthetic Data Platform for Deep Learning Applications

null

neuromation

null

Neuromation

pr@neuromation.io

neuromation-io-blog

AI,DEEP LEARNING,ARTIFICIAL INTELLIGENCE,NEUROMATION,TOKEN SALE

neuromation_io

Machine Learning

machine-learning

Machine Learning

51,320

Neuromation

https://neuromation.io

fbaeecaf782a

Neuromation

629

3

20,181,104

null

0

null

0

null

2018-09-28

2018-09-28 00:02:17

2018-09-28

2018-09-28 00:12:21

1

false

en

2018-09-28

2018-09-28 00:12:21

1

1bb1e1143826

1.211321

0

Pair programmed with Match Du Toit github.com/matchdutoit

5

Use NLP and Machine Learning to find common themes in microfinance loans Pair programmed with Match Du Toit github.com/matchdutoit Problem and Motivation About Kiva Kiva is an international nonprofit founded in 2005 with a mission to connect people through lending to alleviate poverty. In June 2018 Kiva was in 85 countries, and had served 2.9 Million borrowers through $ 1.16 Billion worth of loans. We want to understand if the description can be segmented into meaningful clusters. Meaningful clusters can be potentially be used to analyze interactions with other features of the borrower profiles. Clusters can also be used to identify the features of lenders who provide funding to this cluster borrowers, which can expand Kiva’s tools to analyse the best matches between borrowers and lenders. Data & Methodology This analysis uses Kiva’s Data Snapshot (https://build.kiva.org/docs/data/snapshots) downloaded on June 15th, 2018. We focused on column Desprcition and Description Translated to get an overview of the profiles of Kiva loan borrowers. We choose to use NMF model for its advantage at reducing dimensionality. It also discovers latent features which can be interpreted as topics for large set of text. Conclustions and Recommendations Identified eight clusters with the following themes (unordered): Family Water-related School / Electricity Farming Christian affiliation Philippino Business Kenyan Business Group by Auto translation template How to Use Our Model? Each cluster provides the top popular words and also the descriptions with strongest signals. Example: Below is the description with the highest weight on water cluster: Technology Used Python Sklearn NLTK Pandas Numpy Beautiful Soup Next Steps exclude Spanish descriptions strip out Kiva translation template copy

Use NLP and Machine Learning to find common themes in microfinance loans

0

use-nlp-and-machine-learning-to-find-common-themes-in-microfinance-loans-1bb1e1143826

2018-09-28

2018-09-28 00:12:21

https://medium.com/s/story/use-nlp-and-machine-learning-to-find-common-themes-in-microfinance-loans-1bb1e1143826

false

268

null

Entrepreneurship

entrepreneurship

Entrepreneurship

226,400

Liyou Zhang

Data Science | Finance | Improv | Global citizen

99bb13e322ad

liyou713

0

10

20,181,104

null

0

null

0

null

2018-08-06

2018-08-06 14:20:30

2018-08-06

2018-08-06 14:21:28

1

false

en

2018-08-21

2018-08-21 06:53:10

7

1bb25449df1d

4.528302

0

Like most people, coffee is one of the most important rituals in my morning routine. There’s something about the aroma and taste that…

3

Artificial Intelligence: What’s Now and Next in IoT-Driven Supply Chain Innovation Like most people, coffee is one of the most important rituals in my morning routine. There’s something about the aroma and taste that kick-starts my ability to have a great day. So imagine my surprise when I found out that a favorite coffee shop was closed before I had to jump on an early-morning flight home. The employees were in the shop, but the gate locked out coffee aficionados, like me, that really needed that jolt of caffeine. Although this experience was understandably a letdown, it was also a source of inspiration. Think about it: How many times has your business been “locked out” of an opportunity to change? You see the advantages that can help your operation move forward and accomplish great things; however, there’s something that’s keeping you from crossing that threshold and succeeding. Such is the case for supply chain automation initiatives. Although the Internet of Things (IoT) is playing a significant role in today’s supply chain, advanced analytics-driven data aggregation platforms are now earmarked as an area to watch. IDC recently reported that 60% of manufacturers will likely leverage an advanced analytics-driven data aggregation platform to improve the speed and accuracy of the fulfillment process this year. However, Gartner gives a stern warning that three out of five factory-level artificial intelligence (AI) initiatives in large global companies will likely stall within the next three years, due to inadequate skill sets. Three AI opportunities that can strengthen your IoT-driven supply chain IDC’s and Gartner’s predictions are stunning revelations considering the skyrocketing growth of computing capacity and data volumes used to empower supply chain leaders to make smarter decisions. But if done well, supply chain managers can extend their IoT capabilities with artificial intelligence to run operations that are fast, nimble, and intelligent enough to stay competitive in today’s high-speed global marketplace. 1. Extend your IoT platform to build a smarter supply chain As IoT devices get increasingly smaller and more prevalent in every asset along the supply chain, an impressive volume of data is not fully leveraged — leaving much of the insight it contains in the dark. Personally, I think this common problem is not a problem at all. Instead, it’s a sign that the IoT is maturing to a point where AI is the natural next step to discover and use real-time information in the best way possible. For example, when a new customer signs a contract, production planning can start automatically. A digital signature triggers warehouses to pick and ship goods needed as outlined in the agreement. Then production is scheduled, and qualified and available human resources are assigned by a system. If employees need to travel to a customer location to install a machine, arrangements are made in parallel. Through AI, the best rates for hotels, flights, and car rentals and dates that fit into everyone’s schedule and time restrictions can be determined immediately. Once installed, the machine can use Big Data algorithms to learn patterns and behaviors. This approach enables them to detect the threat of malfunctions that require maintenance, define factors impacting performance, and optimize processes and opportunities for automation. 2. Drive profitability with unprecedented optimization and simplification The struggle against complexity is something that plagues the mind of supply chain managers. From a growing network of suppliers and the risk of corrupt sourcing practices to trade restrictions and just-in-time delivery, automation can help them sleep better at night. Center for Global Enterprise research has revealed that the more digital the supply chain, the greater the chance the business can reduce procurement costs by 20%, lower supply chain process costs by 50%, and increase revenue by 10%. For example, the beverage powerhouse, Schweppes Australia, combated the inaccuracies and inefficiencies of its supply chain by upgrading its entire distribution center management system with AI technology. The paperless system gave supply chain managers greater visibility into every task — from sourcing to last-mile delivery — to eliminate inefficient practices such as over-replenishing pick faces, which often led to delivery delays and suboptimal spend. By introducing more flexible and efficient practices, Schweppes’ supply chain processes are 99.9% accurate. Managers monitor shipments, at the click of a button, to pinpoint and evaluate gaps in replenishment, prioritize and sequence order drops, and oversee process status. As a result, the company streamlined its order shipment process to picking by bulk, transferring orders to the staging line, dropping actual orders, and retrieving orders from the staging line. 3. Maximize the potential of every employee involved in supply chain processes The use of robots in the supply chain is a hot topic. But contrary to widespread fears, the real news is not about eliminating human job — it’s about making work more meaningful, and challenging for everyone while offsetting a looming labor shortage. In fact, IDCpredicted that 50% of fulfillment centers will co-bots operating next to humans in the picking, packing, and shipping floor to drive up productivity up 30% and lower the cost of operations For retail heavyweight Amazon, deploying an army of over 30,000 Kiva robots across a few of its warehouses in 2014 saved roughly US$22 million. And if Deutsche Bank is correct, the company will pick up an additional $800 million in savings as more plants are given the opportunity to use the technology. 4. Turn your supply chain into a source for value-add services Supply chain management should prepare for the future by implementing the IoT and define new use cases to tap into never-before-conceived revenue streams. And if there was a reason to get started, hygiene company Hagleitner is an excellent source of inspiration. For years, Hagleitner has been a reliable bathroom supplier for fast-food restaurants, hospitals, and theaters throughout Austria as well as multiple cruise lines internationally. As the demand for its services grew, the company decided to make its operations more efficient by embedding sensors to track everything from the use of faucets to stocks levels of soap, air freshener, and paper towels. This strategy not only made services more responsive, proactive, and consistent, but the company is also saving warehouse space, meeting demand with greater sustainability, and optimizing logistics processes and personnel assignments. At the same time, its customers are assured that their bathrooms are well-equipped and address every visitor’s bathroom needs. Artificial intelligence: A natural step in supply-chain innovation The more supply chain technology matures, the smarter the supply chain will run. While the IoT is helping your supply chain respond faster and more flexibly to market changes, it is still important to look ahead and see how the data you’re generating can take your supply chains to new levels of efficiency, demand forecasting, and speed. By Dr. Marcell Vollmer, Chief Digital Officer, SAP Ariba

Artificial Intelligence: What’s Now and Next in IoT-Driven Supply Chain Innovation

0

artificial-intelligence-whats-now-and-next-in-iot-driven-supply-chain-innovation-1bb25449df1d

2018-08-21

2018-08-21 06:53:10

https://medium.com/s/story/artificial-intelligence-whats-now-and-next-in-iot-driven-supply-chain-innovation-1bb25449df1d

false

1,147

null

Supply Chain

supply-chain

Supply Chain

6,262

Dr. Marcell Vollmer

Chief Digital Officer & SVP @SAPAriba. Passionate about #Life, #Coffee, #PhD in #Politics, #MBA in #Economics, #SocialMedia Enthusiast and curious to learn&grow

e92f982c8bd0

marcellvollmer

11

15

20,181,104

null

0

null

0

null

2018-01-17

2018-01-17 01:23:34

2018-01-17

2018-01-17 01:28:00

2

false

en

2018-01-17

2018-01-17 01:28:00

2

1bb2ea2497cb

1.156918

0

This week Gotyu was proud to present its GPS live tracker & activity monitor at Eureka Park at CES 2018. We had amazing success and…

5

Best of CES 2018: AI, self-driving, Robomart, smart mirror This week Gotyu was proud to present its GPS live tracker & activity monitor at Eureka Park at CES 2018. We had amazing success and meetings with Amazon, Indiegogo, numerous distributors, Walmart, BestBuy etc. Here is our Top 10 coolest trends & gadgets you should definitely keep an eye on and buy when in the market. SMART TOYS AT CES 2018 Sony’s Aibo robot dog at CES 2018 #1 In addition to all kinds of smart sensors, displays, numerous robots for various functions, Sony presented its new robot dog Aibo which is available for pre-order at $1,730 now. Sony’s Aibo has now OLED eyes which make it seem more realistic, and artificial intelligence algorithms will help it develop its own personality, recognize different members of the family, and play more with the family member who pets the puppy most. Very cool indeed! Square Off self-moving smart chess board #2 A team from India presented a very unique self-moving chessboard which successfully raised about $235,000 on Kickstarter. Square Off connected board makes it possible to play with any friend or chess enthusiast relative anywhere in the world. Multiple playing modes possible and already 200,000 players in the network. Enjoy! Continued at: http://www.gotyu.tech/Blog/ces-2018-cool-gadgets-consumer-electronics/

Best of CES 2018: AI, self-driving, Robomart, smart mirror

0

best-of-ces-2018-ai-self-driving-robomart-smart-mirror-1bb2ea2497cb

2018-01-17

2018-01-17 01:28:01

https://medium.com/s/story/best-of-ces-2018-ai-self-driving-robomart-smart-mirror-1bb2ea2497cb

false

205

null

Ces 2017

ces-2017

Ces 2017

294

Nataliya Stelmakh

null

60ebfad46199

nataliyastelmakh

104

106

20,181,104

null

0

null

0

null

2017-08-31

2017-08-31 22:37:34

2017-10-02

2017-10-02 12:58:38

0

false

en

2017-10-02

2017-10-02 12:58:38

3

1bb30f175d49

2.279245

5

0

Data is everywhere. Now more than ever, the internet has exploded with almost inconceivable amounts of data just waiting to be mined by the…

5

An Intro into Ethical Issues of Data Science Data is everywhere. Now more than ever, the internet has exploded with almost inconceivable amounts of data just waiting to be mined by the eager researcher. The industry is continually discovering new and innovative ways to apply this data from music recommendations to disease prediction. As all of these applications evolve, the time is ripe for a consideration of various ethical problems that are sometimes subtly presented not only to the researcher but also other interested parties. In this short discussion, four of the many ethical issues related to data science will be presented as based on the article presented by Michael Fuller this year entitled Big Data, Ethics and Religion: New Questions from a New Science. First, any use of data requires assumed or explicit consent. So often, though, users see this consent in the form of a privacy policy or EULA that is not only incredibly long but also just as confusing to understand with the legal terminology and release of rights. These documents are far too often focused on the mitigation of liability rather than genuinely explaining what can or will be done with the user’s data. This ethical inconsistency of informing yet not actually informing the user could be solved with readable explanations as well as more granularity of consent rather than a single agreement to the entire lengthy document. Far too often, the disciplines of law and software become isolated from each other such that neither can coordinate enough to care for their customers. Second, an ethical contention exists between data-based intelligence and individual privacy. As Fuller presents, aggregation and analysis of individual medical information can be extremely useful particularly for future research and development of new treatments. Additionally, an individual can be easily identifiable by only three traits: gender, zip code, and year of birth. Understanding this fine balance will be key to the necessary discussion and disclosure of what data can be used and how it may be applied. The solution here, unfortunately, will not be an easy one. Third, ownership and rights to share data pose a significant ethical question. Is identifiable personal data always personal or can it always be shared as it is now by data brokers? Some countries have actually mandated web-based services to keep surveillance data on their users for government access. Like other dilemmas, these data “rights” must be responsibly disclosed and open for discussion. Individual privacy has to be balanced with security concerns particularly at the civil level. Fourth and perhaps most subtle, research bias can creep in perhaps even more than in typical scientific journals. Who performs the data? What data is cleaned from the set? What metrics are used for analysis? Does the presentation of results accurately portray the information? As Fuller recognizes, bias, either intentional or not, can appear in data science at nearly every step of the process. Although bias will always be present, peer review and disclosure of methodologies as applied by other scientific disciplines can assist in keeping these decisions accountable and more ethically sound. In short, data science is inseparable from ethics. The Bible says that money is not evil but rather the love of money is the root of all evil (1 Timothy 6:10). Likewise, data is not the problem; every concern revolves around how it is used. Open discussion and responsible transparency will be the major weapons in this battle of data control and regulation. What are your concerns with Big Data? What do you think about cloud-based computing or data brokering? Add to the discussion by leaving a comment.

An Intro into Ethical Issues of Data Science

11

an-intro-into-ethical-issues-of-data-science-1bb30f175d49

2018-02-01

2018-02-01 04:16:00

https://medium.com/s/story/an-intro-into-ethical-issues-of-data-science-1bb30f175d49

false

604

null

Big Data

big-data

Big Data

24,602

Christian Di Lorenzo

Web/iOS software engineer @RoleModelSoftware. Data Science grad student dedicated to quality results. My mission is to “glorify God and enjoy Him forever.”

c6a4bcd77899

rcdilorenzo

27

4

20,181,104

null

0

null

0

null

2018-07-12

2018-07-12 12:26:59

2018-07-12

2018-07-12 13:07:29

0

false

en

2018-07-12

2018-07-12 13:07:29

0

1bb3e7ddd76c

1.101887

0

Hello people last time we discussed about self-confidence, and why it is necessary. Today we are going to talk about artificial…

1

Future can be dangerous{AI} Hello people last time we discussed about self-confidence, and why it is necessary. Today we are going to talk about artificial intelligence. Before I start we have to first understand the meaning of it. Artificial intelligence is the theory of being able to perform tasks which normally requires human intelligence, such as visual perception, speech recognition, decision making and translation between languages. Artificial intelligence makes it possible for machines to learn from experiences. Most Artificial intelligence we see, hear these are chess playing computers, self-driving cars and humanoids {human like robots}. Artificial intelligence is both a boon and a bane for the society. Let us first see how it is a boon. It is a boon as it makes our life easier and more entertaining. They also help us to learn many new things and especially humanoids helps a person to recover from depression, loneliness a lot. But the other side of Artificial intelligence is far more threatening. Recently a humanoid, Sophia has been made. She has emotions, expressions and intelligence just like humans. It was asked to her that what she wanted to do in future. She said that she had to overcome the world by her powers. In future maybe the robots actually overcome the world and humans will be totally dependent on them for everything. Today only we see people who are very addicted to phones, iPad, television and what not. God knows what future will be. So we saw that Artificial intelligence is both boon and a bane, but the weightage of bane is more than boon. So one should always try to be attached to the nature instead of these machines. Otherwise one day our life will become useless and worthless.

Future can be dangerous{AI}

0

future-can-be-dangerous-ai-1bb3e7ddd76c

2018-07-12

2018-07-12 13:07:29

https://medium.com/s/story/future-can-be-dangerous-ai-1bb3e7ddd76c

false

292

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Bhavya Roy

null

688c0d20bb6f

bavyaroy

1

3

20,181,104

null

0

null

0

null

2018-07-12

2018-07-12 01:53:18

2018-07-12

2018-07-12 01:55:49

2

false

en

2018-07-12

2018-07-12 01:55:49

0

1bb425e66720

1.53805

1

0

The World Cup is almost over, and the competition has been heated and dramatic! Maybe even more so off the field, where xHamster has poured…

5

The World Cup 2018 Porn Data The World Cup is almost over, and the competition has been heated and dramatic! Maybe even more so off the field, where xHamster has poured over data from FIFA countries, and seen massive World Cup traffic changes. Rather than take away people’s attention from porn, the World Cup seems to have stoked their desires — particularly in the Middle East and North Africa. Since the World Cup began, Morocco has seen a stunning 147.4% increase in traffic. Saudi Arabia, Senegal, Tunisia, and Egypt have all had an over 30% rise in traffic during the World Cup. Goal! But not every country gets aroused from kicking balls. The United States, for instance, saw a 1.7% of a decrease in their porn habits during the World Cup. Maybe if they’d stayed longer, they’d have longer… staying power. And Colombia preferred the action on the field to action on the laptop. They saw the biggest decrease of FIFA countries — traffic to xHamster from Colombia has dropped 7.7%. Other Latin American countries share the same preference — traffic from Costa Rica, Brazil, Portugal, Mexico, and Argentina all dropped over 4%. That said, we have seen some increase in interest related to the World Cup. Users in Costa Rica couldn’t get enough “Russian” porn during the series — searches in that country alone for the term have been up 68.1% increase. Croatia, on the other hand, saw a dramatic decrease in “Russian” searches, decreasing during the World Cup by -28.2%. Could competition hurt erections? The United States also saw its searches for “Russian” decrease — 7.1% — but that could be as much political as it is athletic. Other top searches in the United States have remained unchanged — mom, wife, and massage still head the list. No word on apple pie. XOXO, Phoenix

The World Cup 2018 Porn Data

1

the-world-cup-2018-porn-data-1bb425e66720

2018-07-12

2018-07-12 01:55:50

https://medium.com/s/story/the-world-cup-2018-porn-data-1bb425e66720

false

306

null

Soccer

soccer

Soccer

22,100

Phoenix xHamster

Product Manager at xHamster. It’s a story about my work at the best adult company ever & challenges of running sex-inspired product. Let’s make some PORN! (SFW)

62bbabcdbcfe

phoenix_xh

91

33

20,181,104

null

0

null

0

null

2018-08-18

2018-08-18 18:21:08

2018-08-18

2018-08-18 18:22:00

1

true

en

2018-08-18

2018-08-18 18:22:00

2

1bb51ebc619f

7.030189

0

Depth is the cultivation of trust and faith.

1

Divided No More Depth is the cultivation of trust and faith. It is the wisdom of knowing your inner truth by having been through the fires of hell, to awaken your eternal truth. This eternal truth is where the seeds of trust and faith blossom into the infinity of consciousness, beyond knowing, and into the realms of wisdom. To really go DEEP into trust and faith, we must experience many challenges. In this time and space, “challenge” is harder, because everything in our world is designed for convenience. In Tokyo, if there is a “one minute” time delay on the train, it can throw an entire schedule off. We have “smart” phones, watches, microwaves, and houses, that tell us everything from what the weather is going to be, directions to a friend’s house, and on and on. Who even knows a phone number by memory anymore? In a world where technology is telling us what to do, and where to go, and essentially who we are– TRUST AND FAITH–are weakening. Why? Because we are putting our trust and faith into “Artificial Intelligence.” Just think about it. “Artificial Intelligence.” It’s here. It’s weird. And, it’s a big challenge to awakening. It’s perhaps our BIGGEST challenge moving forward. When I began to go down the “portals” of “why we really came here.” It has lead me to some shocking and un-shocking revelations. Shocking because it’s “right in front of our face, everyday, all day. From birth to death.” AND Un-shocking because when you “wake up” out of “deception,” the world makes more sense. All those times you wondered, “Why is everything so backward in this world?” “Why do humans do x, y, and z?” All the while knowing how to “fix” and “make things better.” Yes, this is both the shock, and the un-shock. Today, we are going into a topic, I didn’t know we were going too. I just move with the message as it comes through. And, here we are… Artificial Intelligence: The next waves of awakening will come when A.I. is fully introduced into our collective consciousness. This is an important thing for those who are “sensitive.” When this “artificial consciousness” begins to “shock” our world, it will be the “sensitive” ones who will feel it most. The vibration, the way they talk, the everything “artificial” that is now at our door. It’s an energy that is in disharmony to “human resonance” and it has a “false consciousness,” that disturbs the human frequency. This is not by accident. They are designed to be this way. For those of you who remember life before “technology,” you will know what it was like to have “more time in your day,” “meaningful phone conversations,” and “less irritations because of a computer malfunction.” Now with the “convenience” of technology we spend our days dictated by the speed of internet, and the cute pictures our friends post on social media. Technology is leading the way to this “new robotic consciousness.” Imagine man against robot? Artificial Intelligence is the intelligence man “wishes he had,” and it possesses no “feelings,” which makes it able to do anything and everything. Machines don’t care that they are destroying the world. Cutting down trees, killing animals, polluting our waters. They don’t care if you “hurt their feelings,” and they certainly do not have “problems.” Machines will do anything they are programmed to do. Machines will kill anything and anyone they are “programmed” to kill. This is a war mongers dream. Artificial Intelligence is a companies best employee. A.I. can become the “upgrade” to the human. She is flying out of the Middle East, tested in Asia, and will slowly but surely continuously “creep” into the West. She is already in most everyone’s home, and her “children” are moving in. Siri brought her sister, Alexus, and with her comes all her by products. She is separating us from others, and keeping us locked into the prisms of an already “artificial” world. Remember the days of going to a book to find the information you were looking for? No need for that, Siri, will tell you exactly what you need and where to find it. You have to admit, this technology is super convenient, and has changed our world forever. What do we do to prepare for this coming wave? The first thing is just be aware, research, and prepare yourself mind, body, and spirit. Get back into nature. “Lighten your load” meaning get rid of all unnecessary possessions. For this next wave humans must “stay strong,” and “Stay together.” We were given a “prelude” or sneak preview of what A.I. is and the way out of this false reality. The movie “The Matrix.” For those of you haven’t seen it, the last movie was about A.I. and saving the last human safe place. Zion. In the end, “Neo,” destroys the machine. How does he do this? Through the process of his journey, he goes blind. And, yet, because he was already “awakened” his his third eye was turned on, and he was able to find the “source” of A.I. and removed it from the consciousness, and in the process he “dies” or “transcends” into a new consciousness. What this movie was showing us, is like a “preview” of what’s to come. The beauty is the answer is right there as it “always is,” in front of our face. To overcome the A.I. wave, we must awaken our third eye, and transcend this world. We must dive deep into our “original beauty,” and rely on our creativity and gifts to lead us forward. For in reality, if you have a “job” that is something a robot can do, and can do much better, you won’t have a job after A.I. is introduced. Beyond factories, that have already taken “jobs,” these new A.I. creations are capable and able to do far more than ever before. Once robots can move like humans, and have super legs and limbs. There is no need for humans to work many jobs. Robots can cook, clean, kill, plant, cultivate, and work all day and night. Robots can be wives, husbands, children, pets. They can be whatever they are programmed to be. Why is this happening? The program in your lifetime goes deeper than before. It is all a “test” to see “will we survive?” This is the fall of “Atlantis.” Many of you reading this remember Atlantis. All of you have lived in Atlantis. Why did you come back? Remember? The way in and through this next wave, is to go BACK TO YOUR ORIGINAL SOURCE. Your beauty. The “who,” before the “why.” The ancient wisdom of your Grand Mother’s breath. Ask yourself: Who are you? Beneath the make-up, shoes, clothes, jewelry, etc… Who are you? Really? Who are you? Keep going deep into that question. For this will require faith and trust. It will be a journey unlike anything before. It’s not here to scare you. It’s here to grow you. To bring you back together. In harmony. In balance. To stop the destruction. To return to the sacred. The biggest calling, and this is one that is too late to stop. Is to stop the “smart” technology. Stop buying it and using it. This is “too late,” because it’s everywhere. It’s global, it “connects” us, it is embedded so deep into the consciousness of our world, that to “stop” it, is impossible. So, what do we do? We limit our use, we reconnect with nature, and we share the awakening with everyone we can. We take people back to their original source. Into the “oneness” of our soul. We decide “how far,” we take A.I. and if we allow it to be what causes us to be on the “extinct” list. If robots can “do it all,” what need for humans? Think about it. People are having less children now, and will continue to have less children. Certain countries have “restrictions,” on how many kids you can have. Other countries, are begging people to have children. Even offering “incentives” and ways to make it “easier” so people will have children. What countries are most likely to embrace A.I. all the way? In China, men are already purchasing A.I. wives because they have no women to marry. They might have had a chance to get married, had the government not imposed a “one child” law causing families to choose to kill baby girls in favor of having boys. Now there is a shortage of women, and women have many choices for who they can marry. In a few decades, it will only be the wealthy who are married and have children. The poor will invest an entire year’s salary to purchase the newest A.I. wife. In Japan, the population is rapidly declining. Many Japanese are choosing to stay single, and never marry. Less people, mean there are not enough people “working,” so who is doing the work? In Tokyo there are robot hotels, where the robots do all the work. They can run an entire hotel with a staff of 4–6 people. The robots cook, clean, greet you, and even wish you well before you go to sleep. In Saudi Arabia, the first A.I. “Sofia,” has received citizenship, and has more “rights” than women living in Saudi Arabia. That says a lot about how “Sofia’s” children, and grandchildren will be received in Saudi Arabia. Who knows, the first A.I. queen may emerge from there? How far is the United States from introducing A.I.? The doors into this next wave are very near. And, this is a message of love, to plant as a seed. Do not fear it, do not force change where it can’t be changed. Instead love yourself, love your family, help each other out, detach from the constructs of the “mind,” and now go deeper. Go deeper beyond the spiritual “fluff,” the spirits, the galactics, the vortexes of energy, will lead you floating. Now go deeper into your ancient SELF. Connect with the ancestors through ceremony, song, and silence. This is not the time for trinkets and shallow waters. Instead this is a calling to step into your DIVINE PURPOSE. The REAL reason you came here. Not the bills you’ve been slapped with, or the status you are seeking. The REAL REASON you came here at this TIME and to this PLACE. Why did you come here at this time when Artificial Intelligence, is just emerging? What did you sign up to do before you leave this earth plane? Now is the time to go deeper. The depth is where you will find your faith and trust. It’s time to go deeper into our meditations, take actions on those prayers, and come together. We are one. Love and Blessings, Song Bird Grand Mother Jimmy Chang Originally published at songbirdgrandmother.com.

Divided No More

0

divided-no-more-1bb51ebc619f

2018-08-18

2018-08-18 22:20:37

https://medium.com/s/story/divided-no-more-1bb51ebc619f

false

1,810

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

SongBird GrandMother

Portal jumper of all realms, dimensions, times, spaces, and realities serving Grand Mother Earth.

e71fc36b77b3

SongBirdGMother

98

293

20,181,104

null

0

null

0

null

2018-08-29

2018-08-29 14:48:47

2018-06-05

2018-06-05 21:04:53

6

false

en

2018-08-29

2018-08-29 14:49:24

16

1bb56df22c7c

7.00283

0

So… you want to “learn machine learning”… where do you start?

5

The Dynamic Technical Guide to How I am Teaching Myself Machine Learning Who do you envision yourself to be..? So… you want to “learn machine learning”… where do you start? There are two (in my view) directions to go. First, there is the technical track (which I will outline here). Secondly, there is the management track- the track that is most focused on how to strategically integrate machine learning and artificial intelligence technologies into day to day and long term enterprise operations. I now provide resources for both tracks as I continue my own studies and explorations, go here to check out my Dynamic Guide to Learning About AI for Business Executives. Before you start… what are you really doing here asking yourself if you want to learn/learn about machine learning? This is an important question to ask yourself. What is motivating you to learn? What is your goal? What will you DO with the things you learn? And most importantly, how far are you willing to go? Learning a new skill (I firmly believe) does not require a smart person… but a motivated hard worker. If you are that, then let’s go. If hard work scares you, I suggest backing away slowly. With that out of the way, let’s begin… If you do not have a background in Python, C++ or Java development, I would strongly recommend to start with this skillset. The amount of how much you know about what models are doing can be divided into mathematical and technical knowledge. They converge at certain points. Going back to your original goal, identify which one (or both) are a comfortable place for you. If you are like me and want to know just about everything about the field, then go for it. But if you want to know enough to be dangerous, or build machine learning for application at the workplace and nothing more, there are options here for you too! Well, you don’t have to learn all of the languages I mentioned above for starters. But if you want to actually understand what models are doing, then understanding how coding languages work will save you a lot of confusion and heartache. If, at this point, you are not interested in learning how to code, check out Google’s AutoML . This product is now available to build your own translate and natural language processing models. The Google Cloud team has done a great job in packaging the hard parts of machine learning model development in a user-friendly user interface that requires only general knowledge on how machine learning development works. Would I recommend only developing models in a vacuum with AutoML? Absolutely not. Ignorance is not bliss if you are actually trying to make something work for you, let alone machine learning! I would highly recommend Python as a first language to tackle in general, but especially if you are interested in TensorFlow development (what I am currently teaching myself how to do). While Google has provided some Java support for TensorFlow development, the majority of the documentation supporting TensorFlow development is for Python development. Not just that- Python and its pandas library are essential to development when it comes to big data. Big data and data processing is actually essential to any machine learning development, so save yourself some time and just learn Python first! Codeacademy is awesome for desktop based coding learning, and their Python course is solid. When I was trying to decide if coding was for me, I would turn on one episode of something (45 mins to one hour) in the evenings after work and just sit there with the TV running and code. It turned into the most enjoyable and relaxing time of my day, and having the TV episode running gave me a certain time window where I knew I needed to only be coding. Grasshopper, Google’s Area 120 Incubator project that is a mobile app teaching adults how to code in Javascript Grasshopper is an awesome app Google just developed to teach adults how to code in Javascript. Why does Javascript matter, one would ask, if you told me to learn Python? Javascript = coding essence, especially when it comes to algorithms you will see over and over again even in Python. As someone with a non-development background who jumped into development with little to no preparation, messing around with this app will only strengthen your understanding of your evenings in front of the TV with Python. Instead of scrolling through Facebook/Instagram/name your vice app during your free time during the day, take five minutes to complete one module on the app while waiting or commuting. 2. Jump right in and get your hands dirty going through this series with me and build your own object classification model. The best way to learn anything technical, I have found, is to make sure you have a partner in life who will feed you during your hours of maze-navigating, but most importantly be prepared to struggle. Not just that- you must enjoy the process of struggling to find an answer, which makes finding an answer that much more gratifying. The ability to enjoy struggle and come out hungry for more is (in my opinion) how you will know if you are really cut out for this kind of exploration. This is also a conscious choice- if you decide you will fight through the struggle for the sake of machine learning, you will learn machine learning! Once you dive right in and go ahead and build a model (and allow yourself to stray and struggle along the way) you will have a pretty good idea of what ML development looks like, on a very high level process wise. Even for the non-technical, the management or leadership types, having at least this much depth into what ML looks like from a process perspective is extremely valuable. It gives you insight into what types of people, how many, what types of data and how much time goes into the machine learning development process. If you recall, the first version of this article recommended the Google’s ML Crash course next. About 5 modules into the crash course, I have decided to take a huge step back. I needed more contextual knowledge, as it relates to Google Cloud specifically, since that is where I work for starters. Not just that though- I also felt like I was diving deeper than I wanted into the mathematical aspects of machine learning, which was lending a little less to my ability to actually crank away at hard problems that could be solved in the workplace with machine learning. 3. Now, I am dividing my learning track in two. One part is traditional and theoretical knowledge acquisition, the other being hands on projects. As I write this today, I am currently writing a thesis on how to apply the best machine learning techniques to network netflow classification for cybersecurity purposes, since that is what my degree is in currently. I am gaining a ton of good theoretical knowledge by researching this. See this doc for all of the technical papers I am slowly sifting through as my research resources- this is updated in real time, watch as my list grows! What am I working on now? I am pivoting to Coursera’s offering of Google Cloud Platform courses geared towards applying machine learning to actual problems and datasets at the workplace with Google tools. I will be starting with this specialization in the recommended order that Coursera suggests. Read one article a day about machine learning on Medium. Just pick something that sounds interesting, even if it seems a bit over your head. Getting exposed to the language, the math, and various approaches to machine learning development will keep your mind open and prevent the problem of you becoming a biased (ha… ha…) machine learning developer. One hour per day on learning something ML related. Whether that is me blogging about what I am learning or actually completing coursework on Coursera or the MLCC, one hour will happen. Block it off on your calendar, and let your friends/family/coworkers know that is sacred time for you. A few pages out of Jason Brownlee’s Algorithms book once a week. Currently I am shooting for one Algorithm a month (a few pages in his book a week). Always have a hands on project exploring something you are learning. Whether it is something for you that you are building, something for a customer, or a Qwiklab/Codelab/Tutorial you are working through, always keep a hands on project around. Other Resources to Explore: Machine Learning Mastery is an awesome website in general with good resources. I just bought his basic machine learning algorithms e-book, and plan on trying to eek a set amount of work on the book per day. I love supporting folks who publish e-books to spread their knowledge and experience, especially small self-published writers. This one is bound to be awesome. 3 Blue 1 Brown has a channel specifically dedicated to deep learning, and also has other interesting channels when it comes to general mathematical concepts. I recommend running through the deep learning channel, and go through these videos multiple times until you feel like you understand how neural networks work (forward propagation), gradient descent, and backprop. If you want a thorough and in-depth place to look for specific ML options, I would take a look at this article. This guy did a VERY thorough job in evaluating learning options for ML! Follow me on Twitter at @AminaAlSherif1 as I continue to update this guide and teach myself machine learning! Originally published at medium.com on June 5, 2018. Originally published at medium.com on June 5, 2018.

The Dynamic Technical Guide to How I am Teaching Myself Machine Learning

0

the-dynamic-technical-guide-to-how-i-am-teaching-myself-machine-learning-1bb56df22c7c

2018-08-31

2018-08-31 13:52:27

https://medium.com/s/story/the-dynamic-technical-guide-to-how-i-am-teaching-myself-machine-learning-1bb56df22c7c

false

1,604

null

Machine Learning

machine-learning

Machine Learning

51,320

Amina Al Sherif

null

de46c1e173d3

amina.alsherif

34

3

20,181,104

null

0

null

0

null

2017-10-30

2017-10-30 22:19:53

2018-02-13

2018-02-13 13:06:02

7

false

en

2018-02-13

2018-02-13 13:08:33

3

1bb5db0d9a3d

5.151887

6

0

Volume 1: Predictive Text

5

Writing With Artificial Intelligence Volume 1: Predictive Text An old Smith Corona typewriter watches me sleep from my bedside table. The jagged parts have punched their way through the cellophane cage that it’s lived in since our move from Berkeley to Los Angeles a few years back. The aged machine is cinder block heavy, ash-black, colored of worn oil grit on hot asphalt and lives in the same romantic category of lost smells as the antique passing sports car spewing leaded gasoline smoke on a dirt road. Sometimes at night a gentle breeze pulls a reminiscent dead scent of metal letters and crackled ink ribbon death as I drift to sleep thinking of Hemingway, Zevon and Bukowski violently hammering away at brilliant white wooden sheets of muse in some very alone place. The history of these particular hammers belong to the former shotgun factory the Smith brothers owned. An appropriately simple translation of pen plus sword and a dark kinship of the final pair of tools that Hemingway lovingly stroked at home in Ketchum, Idaho. “I am Jack’s aching chatbot.” I’m a romantic holding onto the past while simultaneously attempting to build the future. As writers we now have a tab open to pull synonyms and antonyms instead of a collection of leather-bound books. We have powerful hardwares and softwares that allow us to take our life’s works in our pockets. But what’s next? I’m already dying for what’s next. The tools that we have now already seem old but I’ve found some pretty amazing inspiration. I’ll save that conversation for the end of the article though… Examples of Writing With Predictive Text “To use predictive text, or not to use predictive text, that is the decision we must make based upon the context.” ~via @HamletBOT (totally made up handle that you should probably do something with) We’re most familiar with this mode as it occurs on our mobile phones. Predictions are based upon the context of words in the message, first letters typed and, in many instances, a database of previously used words. While structure breeds inventiveness predictive text often is so limiting it becomes humorous. Take John Mayer, great guitarist/bad human, who had a little fun with “predictive poetry” by picking one or two words to start and then continue the story by choosing from the three proceeding choices to finish the piece. Also probably written with predictive text: “Your Body Is a Wonderland” In this next example we have a talented writer (me) utilizing a more robust tool. Here the predictive text choices are 6X the amount of Mr. Mayer’s bejeweled iPhone. Having a thematic corpus is also a primary differentiator. This particular corpus is based upon online romance novels. Knowing that the direction that this information would lead me I started with the mentality that the end product would become something drippingly sexy like this: I had no real goal to get to an end destination or to make this a part of a story. My primary goal was to make a sentence that made sense and could pass as creative (a direct challenge to Mr. Mayer via Medium) and, in the end, rudimentary predictive text by itself is a parlour trick whose only real use case scenario other than entertainment is as a writing tool to prime the pump of creative process. On the other hand, with context and editorial license there is a way that an author can be assisted by these simple tools. When I first began working with Ai tools I immediately wanted to build a digital writing partner that would have components of IBM Watson, Natural Language Processing, a series of chatbots serving different functions along with some Twitter API’s… it was obviously another situation that I was going too far too fast. As I researched the few tools that were available I found that there were communities of people utilizing them. One of my favorite examples is brought to you by Botnik Studios (seriously, click that link and play). They’ve recently garnered recognition for feeding every Scrubs script into the ol’ hopper and developing an original script. Even Zac Braff took notice and teased the idea of doing a reading of the bot generated work much to the delight of squealing Ai Twitter nerds like myself. The hilarity of it is in the way that our human minds try to make sense of something that just doesn’t. Take this first page for instance: unintentional brilliance… If this was automatically pushed out as a screenplay then it’s already got me beat. That’s always been a stifling format for me. The content however leaves us with very little that we can work with but the idea of having this episode made with the audience knowing the background of it’s creative process intrigues the heck out of me. The most recent phenomenon (they’re weekly now apparently) is the “I was born” predictive text game. I’ll use @jk_rowling’s tweet as an example because she regularly retweets my articles. I was jk. JK never retweets me. If you’re interested in reading more about this particular story read this article by Bustle. It’s got some of the best tweets from “I was born” so it’s worth the click. The Future Where it does begin to get more interesting is when predictive text is attached to a delicious recipe of other technologies. Long ago in an iOS far away (iOS 8 to be precise) QuickType utilized machine learning to create new dictionaries based upon usage. In the end, predictive text will continue to be one of the most important components in creating a technology that can write with or for a human counterpart. My inner writer lives in a shack made of redwood with a stack of white pages longing for the next great American novel. He drinks scotch and smokes filterless cigarettes that he rolls delicately with his strong fingers. I pet this strange small man occasionally when he saunters onto the pen. “Me no like Tandy. Me like Apple.” He doesn’t have a problem growing a full beard like I do but I take solace in the fact that he’s dying. My Ticonderoga pencils gather dust in the downstairs desk drawer. I’m sure that the ribbons on all of my typewriters would flash into dust upon being struck. The little man dies a little with every keystroke of my MacBook. The inspiration that he gets from a trip to Joshua Tree has been replaced by a quick visit to Botnik or Shelly or a few tweets shot at @pentametron. I like what I’m becoming; a more easily inspired writer. I find myself letting the content come more naturally. I’m more elastic in the way that I edit the human and machine portions together. Eventually Ai tools will become a mix of something amazing in the way that a word processor was a very short time ago.

Writing With Artificial Intelligence

70

writing-with-artificial-intelligence-1bb5db0d9a3d

2018-02-16

2018-02-16 21:27:30

https://medium.com/s/story/writing-with-artificial-intelligence-1bb5db0d9a3d

false

1,087

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

@Ryan_A_Bell

beat poet of tech. NASA JPL & The Emmys. My new book is Slow Dancing on Hard Drugs. Learn more here- http://slowdancingonharddrugs.com

54ea167d33ec

ryan_a_bell

4,791

7,164

20,181,104

null

0

import pandas as pd # names of columns, as per description cols_names = ['Class', 'age', 'menopause', 'tumor-size', 'inv-nodes', 'node-caps', 'deg-malig', 'breast', 'breast-quad', 'irradiat'] # read the data df = (pd.read_csv('breast-cancer.data', header=None, names=cols_names) .replace({'?': 'unknown'})) # NaN are represented by '?' 1. Class: no-recurrence-events, recurrence-events. 2. age: 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70-79, 80-89, 90-99. 3. menopause: lt40, ge40, premeno. 4. tumor-size: 0-4, 5-9, 10-14, 15-19, 20-24, 25-29, 30-34, 35-39, 40-44,45-49, 50-54, 55-59. 5. inv-nodes: 0-2, 3-5, 6-8, 9-11, 12-14, 15-17, 18-20, 21-23, 24-26, 27-29, 30-32, 33-35, 36-39. 6. node-caps: yes, no. 7. deg-malig: 1, 2, 3. 8. breast: left, right. 9. breast-quad: left-up, left-low, right-up, right-low, central. 10. irradiat: yes, no. from sklearn.model_selection import train_test_split X = df.drop(columns='Class') y = df['Class'].copy() X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=42) +---------+ | Feature | +---------+ | value_1 | | value_2 | | value_3 | +---------+ +-----------------+-----------------+-----------------+ | Feature_value_1 | Feature_value_2 | Feature_value_3 | +-----------------+-----------------+-----------------+ | 1 | 0 | 0 | | 0 | 1 | 0 | | 0 | 0 | 1 | +-----------------+-----------------+-----------------+ from sklearn.preprocessing import OneHotEncoder ohe = OneHotEncoder(sparse=False) X_train_ohe = ohe.fit_transform(X_train) from sklearn.preprocessing import LabelEncoder le = LabelEncoder() X_train_le = le.fit_transform(X_train) class MultiColumnLabelEncoder: def __init__(self, columns = None): self.columns = columns # list of column to encode def fit(self, X, y=None): return self def transform(self, X): ''' Transforms columns of X specified in self.columns using LabelEncoder(). If no columns specified, transforms all columns in X. ''' output = X.copy() if self.columns is not None: for col in self.columns: output[col] = LabelEncoder().fit_transform(output[col]) else: for colname, col in output.iteritems(): output[colname] = LabelEncoder().fit_transform(col) return output def fit_transform(self, X, y=None): return self.fit(X, y).transform(X) import category_encoders as ce

18

7dfe0f85577e

2018-06-25

2018-06-25 15:45:04

2018-06-25

2018-06-25 21:42:00

1

false

en

2018-06-25

2018-06-25 21:42:00

10

1bb70f07262d

7.071698

15

3

0

How to easily implement one-hot encoding in Python

4

Dealing with categorical features in machine learning How to easily implement one-hot encoding in Python Photo by Max Nelson on Unsplash Categorical data are commonplace in many Data Science and Machine Learning problems but are usually more challenging to deal with than numerical data. In particular, many machine learning algorithms require that their input is numerical and therefore categorical features must be transformed into numerical features before we can use any of these algorithms. One of the most common ways to make this transformation is to one-hot encode the categorical features, especially when there does not exist a natural ordering between the categories (e.g. a feature ‘City’ with names of cities such as ‘London’, ‘Lisbon’, ‘Berlin’, etc.). For each unique value of a feature (say, ‘London’) one column is created (say, ‘City_London’) where the value is 1 if for that instance the original feature takes that value and 0 otherwise. Even though this type of encoding is used very frequently, it can be frustrating to try to implement it using scikit-learn in Python, as there isn’t currently a simple transformer to apply, especially if you want to use it as a step of your machine learning pipeline. In this post, I’m going to describe how you can still implement it using only scikit-learn and pandas (but with a bit of effort). But, after that, I’ll also show you how you can use the category encoders library to achieve the same thing in a much easier fashion. To illustrate the whole process, I’m going to use the Breast Cancer Data Set from the UCI Machine Learning Repository, which has many categorical features on which to implement the one-hot encoding. Load the data The data we’re going to use is the Breast Cancer Data Set from the UCI Machine Learning Repository. This data set is small and contains several categorical features, which will allow us to quickly explore a few ways to implement the one-hot encoding using Python, pandas and scikit-learn. After downloading the data from the repository, we read it into a pandas dataframe df. This data set has 286 instances with 9 features and one target (‘Class’). The target and features are described in the data set description as follows: We’ll take all columns to be of ‘object’ type and split the training and test sets using the train_test_split of scikit-learn. One-hot encoding There are several ways to encode categorical features (see, for example, here). In this post, we will focus on one of the most common and useful ones, one-hot encoding. After the transformation, each column of the resulting data set corresponds to one unique value of each original feature. For example, suppose we have the following categorical feature with three different unique values. After one-hot encoding, the data set looks like: We want to implement the one-hot encoding to the breast cancer data set, in such a way that the resulting sets are suitable to use in machine learning algorithms. Note that for many of the features of this data set there is a natural ordering between the categories (e.g. the tumour size) and, therefore, other types of encoding might be more appropriate, but for concreteness we will focus only on one-hot encoding in this post. Using scikit-learn Let’s see how we would implement one-hot encoding using scikit-learn. There is a transformer conveniently named OneHotEncoder which, at first glance, seems to be exactly what we’re looking for. If we try to apply the above code, we obtain an ValueError, as OneHotEncoder requires that all values are integers, and not strings as we have. This means we first have to encode all the possible values as integers: for a given feature, if it has n possible values (given by n different strings), we encode them with integers between 0 and n-1. Thankfully, there is another transformer in scikit-learn, called LabelEncoder, which does just that! And… we obtain another ValueError! In reality, LabelEncoder is only intended to be used for the target vector, and as such it doesn’t work with more than one column. Unfortunately, in version 0.19 of scikit-learn, there is no transformer which can deal with several columns (there is some hope for version 0.20). One solution is to make our own transformer, which we creatively call MultiColumnLabelEncoder, which applies the LabelEncoder in each of the features. Will this work this time? It worked! To better understand what happened, let’s check the original training set. We see, for instance, that the age group 30–39 was given the label 0, 40–49 was given the label 1, etc, and analogously for the other features. After applying the MultiColumnLabelEncoder, we can (finally!) use the OneHotEncoder to implement the one-hot encoding to both the training and test sets. Some comments: The OneHotEncoder is fitted to the training set, which means that for each unique value present in the training set, for each feature, a new column is created. We have 39 columns after the encoding. The output is a numpy array (when the option sparse=False is used), which has the disadvantage of losing all the information about the original column names and values. When we try to transform the test set, after having fitted the encoder to the training set, we obtain (again!) a ValueError. This is because the there are new, previously unseen unique values in the test set and the encoder doesn’t know how to handle these values. In order to use both the transformed training and test sets in machine learning algorithms, we need them to have the same number of columns. This last problem can be solved by using the option handle_unknown='ignore' of the OneHotEncoder, which, as the name suggests, will ignore previously unseen values when transforming the test set. And that’s it! Both the training and test sets have 39 columns and are now in a suitable form to be used in machine learning algorithms which require numerical data. However, the procedure shown above is quite inelegant and we lose the dataframe format for the data. Is there an easier way to implement all of this? There is! Using category encoders Category Encoders is a library of scikit-learn-compatible categorical variable encoders, such as: Ordinal One-Hot Binary Helmert Contrast Sum Contrast Polynomial Contrast Backward Difference Contrast Hashing BaseN LeaveOneOut Target Encoding The Ordinal, One-Hot and Hashing encoders are improved versions of the ones present in scikit-learn with the following advantages: Support for pandas dataframes as an input and, optionally, as output; Can explicitly configure which columns in the data are encoded by name or index, or infer non-numeric columns regardless of input type; Compatibility with scikit-learn pipelines. For our purposes, we’re going to use the improved OneHotEncoder and see how much we can simplify the workflow. First, we import the category encoders library. Then, let’s try to apply the OneHotEncoder directly to both the training and test sets. It worked immediately! No need to use LabelEncoder first! Some observations: The outputs are dataframes, which allows us to more easily check how the several features were transformed. As before, each transformed set has 39 columns and it can be checked that the matrices of 0’s and 1’s are identical to the ones obtained previously. I used the option use_cat_names=True so that the possible values of each feature are added to the feature name in each new column (e.g. age_60-69). This was done for clarity, but by default it simply adds an index (e.g. age_1, age_2, etc). We can now more easily understand why we previously obtained an error when we didn’t require that unseen values in the test set were ignored. Below, we show all the columns of the test set after one-hot encoding, when the encoder is fitted to the training set (first) and when is fitted to the test set (second). We see that, in the second case, there are now 42 columns, as 3 new columns were created: age_20-29, tumor-size_5-9 and breast-quad_unknown. These corresponds to three values in the test set (20–29 in age, 5–9 in tumor-sizeand unknown in breast-quad) that are not present in the training set. Comparing the two approaches above, it’s clear that we should use theOneHotEncoder from the category encoders library if we intend to one-hot encode the categorical features in our data set. As remarked in the beginning, one-hot encoding is not the only possible way to encode categorical features and the category encoders library has several encoders which you should explore, as others might be more appropriate for different categorical features and machine learning problems. You can find my LinkedIn in: Hugo Ferreira - Researcher - Instituto Superior Técnico | LinkedIn View Hugo Ferreira's profile on LinkedIn, the world's largest professional community. Hugo has 3 jobs listed on their…www.linkedin.com

Dealing with categorical features in machine learning

34

dealing-with-categorical-features-in-machine-learning-1bb70f07262d

2018-06-25

2018-06-25 21:42:00

https://medium.com/s/story/dealing-with-categorical-features-in-machine-learning-1bb70f07262d

false

1,821

From Physics and Mathematics to Data Science: a journey

null

Hugo Ferreira’s blog

hugorcf@outlook.com

hugo-ferreiras-blog

HUGO FERREIRA,DATA SCIENCE,MACHINE LEARNING,ARTIFICIAL INTELLIGENCE

hu_daa

Machine Learning

machine-learning

Machine Learning

51,320

Hugo Ferreira

Aspiring Data Scientist; physicist and maths geek.

26e4a25a33ff

hugorcf

55

33

20,181,104

null

0

null

0

d8c4ad013dfd

2018-07-16

2018-07-16 20:37:26

2018-07-20

2018-07-20 20:56:43

3

false

en

2018-07-20

2018-07-20 20:56:43

4

1bb78a43d8d

4.357547

32

1

0

Over the last few months I’ve worked as an intern at Numerai. If you haven’t heard of Numerai, it’s a hedge fund that is innovating far…

5

Game Theory Optimal Play for the Numerai Competition Over the last few months I’ve worked as an intern at Numerai. If you haven’t heard of Numerai, it’s a hedge fund that is innovating far beyond the status quo of the financial industry. We run a data science tournament where anyone in the world can download our training data and try to predict the future. If they do a good job, we will trade real world financial assets on their predictions, and pay them for their work in our own proprietary cryptocurrency. Numerai is tapping into a diverse and talented network of data scientists to build a multifaceted artificial intelligence, monopolizing the “wisdom of the crowd”. The only natural cap on this idea is the number of data scientists in the world. But some of this is easier said than done. For starters, how do we decide which data scientists to pay? And then how much should each participant be paid? Answering these questions is of the utmost importance to Numerai. The game theory that underpins the tournament determines the quality of the machine learning models people use. Last month, I and several others came to the realization that our previous tournament structure, though strong in certain aspects, was not perfectly incentivizing the data scientists to build the best models they possibly could. Without getting into the specifics, users were discovering that they could maximize profit by using statistically bad machine learning models. In some tournaments, one or two users won all of the prize pool, even though they hadn’t submitted particularly good predictions. This had to change. Fast forward to today, and we have implemented a new tournament that we believe vastly improves the emergent properties of users’ staking behavior. It is designed to be more fair for the users, and to generate better data for our meta-model. So how does this game theory work? How can you, dear user, choose the best p (probability parameter) to maximize earnings? Before reading further, if you are not familiar with the new tournament, or have not participated in the Numerai competition before, I would recommend checking out the rules here. It may also be good to read up on expected value if you are new to the concept. Without further ado: There are two main factors when considering how to play the tournament optimally. Firstly, a user will want to make sure their chosen p is less than or equal to their actual expected probability of beating the log loss threshold, denoted p-hat. This is because we have designed the payment structure such that the expected value of any bet is break even at p = p-hat. The higher a user sets p the worse odds they will get in the game, so the higher percentage of the time they will have to beat the threshold to be profitable. According to our rules, when a user beats the log loss threshold they can expect to make ((1-p)/p)*stake when they win. For instance, if a user succeeds 50% of the time, and they choose p=0.9, they stand to win(1–0.9)*stake/(0.9) = (1/9)*stake. When they lose we burn their entire stake. So the user expects to win (1/9)*stake 50% of the time and to lose (1/1)*stake 50% of the time. Long run their expected value is (-0.444)*stake per bet. This is no good, and playing this way a user will not be profitable. At p = p-hat long run expected value comes out to zero. When p is lower than p-hat expected value is greater than zero. To visualize this: The region where expected value is positive is shaded. Remember that p and p-hat must be between 0 and 1. Here we can see that with the profit equation (1-p/p), break even is p=p-hat. p≤p-hat is required to be profitable over time. So why don’t users just set their p’s as low as possible? Wouldn’t this maximize profit? Actually, this is not game theory optimal. The lower p is set, the less likely a user will be eligible for the prize pool. We payout the highest p’s first, so it is undesirable to be below the eligibility-cutoff p. Additionally, there is no actual benefit in underreporting p. Because we payout everyone in the prize pool at the same p value, and the lowest p value we can, underreporting p will not improve the p at which one is paid. If a user would be in the prize pool at p=0.7, and at 0.6, and we end up paying everyone in the prize pool out at a p of 0.55, then choosing 0.6 provides no benefit. However, increase the chance that the user is not in the prize pool at the time of payout, because there may be many submissions with higher p’s above them on the payout list. Taking the risk to use a p lower than p-hat is not offset by the potential reward, because it can only ever, at best, marginally improve the p used in payouts, but can greatly jeopardize one’s chances of being in the prize pool. Game theoretical optimal play can be achieved by setting p as close as possible to the real probability of beating the threshold, p-hat. Credit xkcd. Though it is a winning move to play in our tournament (given p=p-hat!) It is worth noting that a participant with a sufficiently large stake can be guaranteed to deplete the prize pool at any p they chose. To maximize expected profit, it seems that they may occasionally want to choose a p lower than p-hat. However, a participant with equal funds and an equally strong model could come in above them and completely exclude them from the prize pool. Due to this adversarial dynamic, every participant, regardless of capital resources, is incentivized to make the best model possible, and then report it as honestly as possible. Thus game theory optimal play is achieved when p is the closest approximation to p-hat possible. We hope this is as fun to play with as it was to design!

Game Theory Optimal Play for the Numerai Competition

178

game-theory-optimal-play-for-the-numerai-competition-1bb78a43d8d

2018-07-22

2018-07-22 01:38:56

https://medium.com/s/story/game-theory-optimal-play-for-the-numerai-competition-1bb78a43d8d

false

1,009

A new kind of hedge fund built by a network of data scientists.

null

Numerai

contact@numer.ai

numerai

MACHINE LEARNING,ARTIFICIAL INTELLIGENCE,HEDGE FUNDS,FINANCE,BLOCKCHAIN

Numerai

Machine Learning

machine-learning

Machine Learning

51,320

Ben Brimacombe

Currently @ Numerai. Columbia CS & Math. "Luck Is What Happens When Preparation Meets Opportunity"

8a9b8d1b1629

benbrimacombe

99

78

20,181,104

null

0

null

0

null

2017-11-13

2017-11-13 06:13:10

2017-11-13

2017-11-13 06:42:16

3

false

en

2017-11-13

2017-11-13 06:42:16

2

1bb7fc0d360f

1.957547

4

0

In this blog, I will demonstrate to you generally accepted methods to actualize an exchanging system utilizing the administration…

5

How To Trade Using Machine learning In Python - SVM (Support Vector Machine) In this blog, I will demonstrate to you generally accepted methods to actualize an exchanging system utilizing the administration expectations made in the past blog. Do read it, there is an extraordinary markdown for you toward the finish of this. There is one thing that you should remember before you read this blog however: The calculation is only for exhibit and ought not be utilized for genuine exchanging without appropriate enhancement. Give me a chance to start by clarifying the plan of the blog: Make an unsupervised ML ( machine learning) calculation to foresee the administrations. Plot these administrations to envision them. Prepare a Support Vector Classifier calculation with the administration as one of the highlights. Utilize this Support Vector Classifier calculation to foresee the present day’s pattern at the Opening of the market. Envision the execution of this methodology on the test information. Downloadable code for your advantage Import the Libraries and the Data: First, I imported the necessary libraries. Please note that I have imported fix_yahoo_finance package, so I am able to pull data from yahoo. If you do not have this package, I suggest you install it first or change your data source to google. Next, I pulled the data of the same quote, ‘SPY’, which we used in the previous blog and saved it as a dataframe df. I chose the time period for this data to be from the year 2000. After this, I made pointers that can be utilized as highlights for preparing the calculation. Be that as it may, before doing that I chose the think back day and age for these pointers. I picked a think back time of 10 days. You may attempt some other number that suits you. I picked 10 to check for as long as 2 weeks of exchanging information and to maintain a strategic distance from clamor innate in littler think back periods. Aside from the think back period let us additionally choose the test prepare split of the information. I like to give 80% information for preparing and staying 20% information for testing. You can change this according to your need. Read The Original Article at: http://ow.ly/ubUM30gx3bB

How To Trade Using Machine learning In Python - SVM (Support Vector Machine)

4

how-to-trade-using-machine-learning-in-python-svm-support-vector-machine-1bb7fc0d360f

2018-04-20

2018-04-20 08:22:22

https://medium.com/s/story/how-to-trade-using-machine-learning-in-python-svm-support-vector-machine-1bb7fc0d360f

false

373

null

Data Science

data-science

Data Science

33,617

Emily Fuller

null

52007d25cba

emily.fuller

83

524

20,181,104

null

0

null

0

null

2018-04-03

2018-04-03 16:25:24

2018-04-05

2018-04-05 19:00:23

15

false

pt

2018-04-05

2018-04-05 19:05:43

9

1bbb858c686e

9.024528

0

Explicações acerca de Inteligência Artificial

5

Seis coisas que você precisa saber sobre a Inteligência Artificial Afinal, o que é IA? Inteligência artificial é um área de conhecimento da Ciência da Computação. Seu objetivo principal é dotar um software de computador com as características de um ser humano, usando métodos, algoritmos e tecnologias. São subáreas da IA, por exemplo, o aprendizado de máquina (machine learning), visão computacional, processamento de linguagem natural e a robótica. A primeira vez que o termo Inteligência Artificial foi usado foi na década de 60 por John McCarthy uma conferência em New Hampshire nos EUA. E então porque apenas agora o jargão está se popularizando? Até muito pouco tempo atrás, a quantidade de dados que havia disponível para processamento era muito pequena e a capacidade computacional das máquinas era muito restrita. Pensando sobre quantidade de dados, a previsão da Cisco para 2016 era que o tráfego de dados na internet ultrapassaria um ZettaByte, uma medida equivalente a um “sextilhão” de bytes. Um outro número que vem crescendo de forma exponencial é o de pessoas conectadas à rede. Um levantamento da ONU de 2015 já apontava 3,2 bilhões de pessoas conectadas na internet. Ainda em 2015, segundo a Cisco já haviam 25 bilhões de dispositivos conectados à internet. Esses número eram praticamente inexpressivos se comparados aos de uma década atrás. Estamos de fato na era da Big Data. Já falando de capacidade computacional, o Apollo Guidance Computer (AGC) — computador a bordo do módulo lunar — executava instruções a uma velocidade de cerca de 40 KHz (ou 0,00004 GHz), cerca de 100 mil vezes mais lenta do que um laptop topo de linha hoje. Para processamentos complexos, na atualidade, se utilizam as famosas GPUs (Unidades de processamento gráfico). Inicialmente projetadas para Games e processamento de projetos 3D, as GPUs evoluíram para se tornar componentes essenciais do processamento de IA. As TPUs são uma evolução natural das GPUs. Atualmente a tecnologia ligada às TPUs é de propriedade exclusiva da Google e está disponível apenas na nuvem da empresa. Apenas para se ter uma ideia, o processamento de 32 GPUs em um período de tempo é equivalente a ⅛ de TPU. Quais os níveis de maturidade em IA? Como é possível ver no gráfico acima, existem níveis de maturidade encadeados, onde deve-se passar primeiro por um nível para chegar ao subsequente, e assim por diante. Algoritmos de Percepção e Notificação são considerados de BI, mas são a base da escada para a maturidade de IA. É nestas etapas que os dados são estruturados e saneados, e sem isso a qualidade das informações produzidas nos próximos níveis cai consideravelmente. A seguir vem os níveis de Sugestão, Automação e Predição, que são bastante conhecidos. Tendo os dados corretos, podemos então trabalhar algoritmos de sugestão, que em sua maioria são estatísticos. Automação e Predição já usam técnicas de Machine Learning, podendo ser algoritmos clássicos ou Deep Learning. E aí vem um primeiro mito: Deep Learning é a bala de prata! Vamos falar mais sobre ele nos tópicos abaixo. Os níveis de Prevenção e Consciência já envolvem os dados gerados pelos níveis anteriores, para tomadas de decisão automáticas. Para exemplificar, existem alguns algoritmos mais sofisticados de carros autônomos que têm essa consciência dos fatos que acontecem ao ser redor. Desta forma eles podem usar os dados para prevenir acidentes, por exemplo. Freio automático de um carro autônomo Quais problemas de negócio ela resolve? AI resolve basicamente 3 tipos de problemas de negócio: Classificação, Predição e mais recentemente Geração. Ela também pode ser usada para reduzir a dimensionalidade dos dados, mas esse processo é apenas um intermediário aos citados anteriormente, e não vamos nos aprofundar nele. Na Classificação, o algoritmo busca separar os dados entre 2 ou mais tipos. Veja os exemplo abaixo: Classificação linear Imaginando que os pontos azuis são pessoas tolerantes à luz e os pontos vermelhos são pessoas intolerantes à luz, a reta verde é o que o algoritmo de classificação vai tentar traçar. Desta forma, quando chegarem novos dados, se estiverem acima da reta verde, serão pessoas tolerantes à luz e se estiverem abaixo serão intolerantes. É claro que essa é uma representação simples, e na maioria dos problemas no mundo real os dados não são facilmente separáveis por uma reta. Para a predição, são usados algoritmos de regressão. Esse algoritmos visam gerar uma reta/curva que melhor represente os dados. Veja o exemplo abaixo: Regressão não-linear Trata-se de uma representação de dados de venda de café para cada hora do dia. Imaginando que os dados estão corretos, a linha verde representa uma curva que pode ser usada para prever as vendas dado uma determinada hora do dia. Veja que se pegarmos o horário de 15 horas por exemplo, é possível determinar pela curva verde que a possível predição é que serão vendidos em média 90 copos de café aproximadamente. Ainda de forma experimental, existem também algoritmos de geração. Esses buscam padrões existentes e aprendem esses padrões gerando informações de acordo com os padrões aprendidos. Veja abaixo o exemplo de um algoritmo que aprendeu o que é um cavalo e está aplicando um padrão zebrado sobre a imagem de um cavalo. Manipulação de imagens feita por uma rede neural Abaixo estão pássaros que nunca existiram, mas que foram gerados por um algoritmo de IA a partir de um texto: Geração de imagens feita por uma rede neural E o que ela não faz? Vamos começar respondendo a pergunta: O que não é IA? Se alguém algum dia te disse que vai fazer uma IA pra te mostrar como foi seu comportamento de vendas não acredite! Isso é estatística aplicada em Business Intelligence. Existe uma diferença crucial entre as análises de BI e as de IA: as demandas de BI tratam os dados do passado, como é o caso do comportamento de vendas. Já as demandas de IA tratam previsões e classificações, ou seja, utilizam dados existentes para gerar informações futuras. Um bom exemplo de IA aplicada, ainda na área de vendas, seria uma Forecast de vendas, onde o algoritmo projetaria as vendas futuras baseado nas informações passadas. Vamos entrar aqui no paradigmas dos tipos de aprendizado. Muita gente acha que algoritmos de IA aprendem sozinho. Sim, existem algoritmos que aprendem sozinhos! Eles são baseados no que chamamos de aprendizado não-supervisionado. Esses algoritmos são principalmente de classificação, e aprendem através dos dados quais são as características mais relevantes para separar os grupos. Existem ainda os algoritmos supervisionados, que são a grande maioria, os semi-supervisionados e os de aprendizado por reforço. Os supervisionados precisam de dados já com respostas para detectar os padrões necessários. Os semi-supervisionados são uma junção do supervisionado e não-supervisionado. Os algoritmos de aprendizado por reforço, em teoria aprendem com os dados, mas antes de começarem a processar, o responsável deve estabelecer reforços positivos e penalidades para alimentarem o algoritmo durante o processo de aprendizado. Veja abaixo um algoritmo desse tipo jogando um jogo de Atari. O reforço positivo é quando destrói um dos blocos da parte de cima, e a penalidade é quando a bolinha cai. Assim, o algoritmo sozinho cria sua maneira de jogar, visando evitar as penalidades e maximizar os reforços positivos. Jogo sendo jogado por algoritmo de IA que usa aprendizado por reforço Um outro ponto importante de ser destacado é que nenhum algoritmo de IA, seja ele de Machine Learning clássico ou o mais sofisticado Deep Learning trabalha com dados desestruturados e/ou inválidos. Para isso inclusive existe uma etapa chamada de Feature Engineering que visa tratar os dados e deixá-los da maneira própria para que possa ser usado pelos algoritmos. Desta forma, se você pretende usar algum algoritmo de Machine Learning, cuide dos seus dados, porque eles certamente serão fator decisivo no sucesso ou fracasso do seu projeto! O que são redes neurais artificiais? O que é deep learning? Redes neurais artificiais são algoritmos matemáticos inspirados em um cérebro humano, formada através de um conjunto de Perceptrons (Neurônios artificias). A figura acima busca fazer a comparação dos conceitos de um cérebro com uma rede neural, fazendo uma analogia entre o avião e o pássaro. Pensando sobre essa imagem consigo entender que tanto o pássaro quanto o avião tem um mesmo objetivo de voar, ambos tem asas, e que o avião foi inspirado no pássaro, mas não dá pra falar que um avião é um pássaro. Esse mesmo entendimento deve ser aplicado na comparação das redes neurais artificiais com o cérebro (tem objetivos comuns, algumas características parecidas e inspiração). Rede neural artificial E como as redes artificiais funcionam? Elas trabalham recebendo dados, aplicando um conjunto de pesos e passando essas informações por uma função de ativação, que decide se a informação prossegue ou não para os próximos neurônios. Existem funções de ativação que determinam também a intensidade com que essa informação é repassada. Esse processo de receber uma informação, aplicar os pesos e passar pelas camadas de perceptrons é chamado de Feed Foward. Após o cálculo da saída, a própria rede neural calcula o erro relativo às saídas apresentadas. Ela então vem atualizando os pesos das camadas mais profundas para as camadas mais superficiais, utilizando cálculos de Derivada Parcial para decidir aumentar ou diminuir o peso. Esse processo é chamado de Back Propagation. Ao ciclo que uma rede neural faz o processo de Feed Foward e Back Propagation chamamos de Época. Uma rede neural pode executar por várias épocas. O processamento da quantidade de épocas e da quantidade de camadas da rede é o fator mais importante quando falamos de custo computacional pra executar esses algoritmos. E aí caímos no conceito de Deep Learning, que nada mais é do que uma rede neural com várias camadas internas. O tipo e disposição destas camadas é o que define o quão boa é uma rede de Deep Learning. Abaixo está o desenho da GoogLeNet/Inception é uma Rede do tipo Convolucional de 2014, com 22 camadas, e comumente usada para classificação de imagens. Seu percentual de erro é de 6.67% (imagem correta entre as 5 primeiras categorias informadas pela rede), bastante próximo de uma percepção humana, que é de aproximadamente 5.1% de erro. Já existem redes mais recentes com mais de 150 camadas, e com percentual de erro em torno de 3.5%. GoogLeNet/Inception 2014 Cada rede neural tem uma aplicação específica, e o quanto ela é eficiente pra tratar um problema, não significa que ela se manterá o nível de excelência em problema diferente. Por exemplo, o AlphaZero é um algoritmo de Deep Learning criado pela Google para resolver jogos de tabuleiro. Ele é tão bom, que ganhou 100 a 0 de sua versão anterior, o AlphaGo, que já havia derrotado o melhor jogador humano por 4 a 1. Se de alguma forma fosse possível adaptar o AlphaZero para prever oscilações no mercado financeiro de ações, com certeza sua eficiência seria muito baixa, isso simplesmente porque ele não foi projetado para isso. A Inteligência Artificial vai acabar com minha profissão? Não só com a sua, como com a minha também. Aliás, quem te garante que esse texto não está sendo escrito por uma IA? Brincadeiras a parte, já existem estudos e várias pesquisas mostrando que algoritmos de IA estão passando a fazer atividades não só repetitivas, mas com certo grau de abstração. Veja abaixo algumas manchetes: Se uma IA consegue estruturar processos jurídicos como um bom advogado, fazer diagnósticos como um bom médico, escrever poemas como um bom escritor e ser indicada como membro de conselho administrativo de uma empresa, não há como falar que eu e você estamos imunes. Não acredito que seja o apocalipse das máquinas acontecendo bem debaixo dos nossos narizes, mas sim a Quarta Revolução Industrial. Essa revolução vai afetar a maneira como as pessoas consomem produtos, como as pessoas se relacionam e com certeza como as pessoas trabalham. Nesse ponto, a IA vem pra automatizar e aumentar a capacidade humana. Uma boa dica é procurar trabalhos mais abstratos, que são mais complexos de serem tratados por algoritmos de IA, como a Filosofia, Matemática Pura, Terapia, Enfermaria, etc. Ter conhecimento de programação vai passar a ser essencial, é tanto que já existem vários estudos sobre a inclusão dessa disciplina nas escolas. Obrigado por fazer essa viagem junto comigo! Deixe seus comentários!

Seis coisas que você precisa saber sobre a Inteligência Artificial

0

seis-coisas-que-você-precisa-saber-sobre-a-inteligência-artificial-1bbb858c686e

2018-04-05

2018-04-05 19:05:43

https://medium.com/s/story/seis-coisas-que-você-precisa-saber-sobre-a-inteligência-artificial-1bbb858c686e

false

1,994

null

Inteligencia Artificial

inteligencia-artificial

Inteligencia Artificial

1,614

Daniel Costa

Especialista em desenvolvimento de software e inovação com vários anos de atuação. Evangelista da aplicação de IA em negócios.

47ac834553f7

cmtedbcosta

0

1

20,181,104

null

0

null

0

8ab23399906b

2018-08-24

2018-08-24 10:45:42

2018-08-24

2018-08-24 10:46:57

2

false

en

2018-08-24

2018-08-24 10:46:57

0

1bbd1e632b8

3.413522

0

Behind the scenes of DACONOMY’s decentralized data ecosystem are several platform engines, also known as the machine room of DACONOMY…

1

Platform Engines — The Machine Room of DACONOMY Behind the scenes of DACONOMY’s decentralized data ecosystem are several platform engines, also known as the machine room of DACONOMY. These engines are driven by artificial intelligence and were developed to solve some of the major problems that the data industry faces today, such as data analysis, data anonymization, and data pricing. In this post, we’ll cover the high-powered engines being used to change the data landscape as we know it. DACONOMY — Analytics Engine At DACONOMY, one of our primary goals is to ensure the growth of our ecosystem; clustering data types and markets partially do this in a way that meets and caters to demand. At the forefront of these efforts is our analytics engine. Once someone has put their data on DACONOMY’s platform, it’s immediately run through the analytics engine. Since we don’t own or store the data, once it comes into our engine we extract all the information needed to make the data marketable. The analytics engine does not only identify data type, it also curates data, cleans data, and ensures it’s neither broken nor corrupted. The DACONOMY Data Anonymity Engine A critical aspect of the DACONOMY marketplace is its ability to handle data in an anonymous manner. A buyer can specify the structured items that they are interested in, and the DACONOMY marketplace will identify matches for that query as a trustee of the respective seller. To facilitate this process, the anonymity engine allows dataset owners to apply a certain level of anonymity to their datasets based on pre-set features determined by DACONOMY or their personal criteria; this ensures that the only information shared by data owners is the information that they’ve chosen. Datasets making use of the anonymity engine will typically not be exchanged between seller and buyer directly, but rather through DACONOMY marketplace smart contracts acting as both the orchestrator of similar data elements and the broker acting on behalf of a group of data sellers. The DACONOMY Classification and Taxonomy Algorithms The classification and taxonomy algorithms are a core part of the DACONOMY architecture, and one of the primary interfaces data provision has to undergo. These algorithms, apart from indexing all the data, enable DACONOMY to autonomously classify the data to make it fit in the most appropriate taxonomic class(es). Once the classification, indexing, and taxonomic model are assigned, a sample of the structured data is stored for matching buyers and sellers. The DACONOMY Verification and Validation Engine The DACONOMY verification and validation engine will verify users during registration and provide different verification levels based on deals successfully executed, verification data provided, recommendation data from peers, and KYC profiles. On the data side, the verification and validation engine will validate the data sets and assign a quality indicator based on the AI-based validation checks, grading consistency, statistical significance, successful checks against reference data, update cycles, and more. Furthermore, the verification and validation engine allows the DACONOMY marketplace to verify data sources in terms of their legal correctness and the user rights of the seller. The DACONOMY Matching Engine Finding appropriate, relevant data easily and quickly is a core strength of the DACONOMY platform. With state-of-the-art data-searching algorithms on structured data sets, the platform will ensure that it presents fit-for-purpose results. Also, to improve the efficiency and speed of relevant data matching, the engine will not only return results, but will also include an easy to digest graphical representation of the data sets found consisting of heat maps, bar diagrams, and other data visualization tools; this will ensure a clear understanding of the data that is being presented. The DACONOMY Pricing Engine The pricing engine calculates a value for any given dataset or data stream. It uses actual prices paid for data exchanged over the platform to refine the pricing function from deal to deal. The price can vary with certain parameters of the dataset, level of depth provided (see also precision levels in Anonymity Engine above), update cycles (none, weekly, daily, hourly), and right of use (royalty, one-time use, use and burn, etc.). In addition, the owner of a dataset can either override any price proposal calculated by the DACONOMY platform or select “auction” as the pricing algorithm. The auction algorithm advertises datasets for a given period and collects offers from registered buyers. The seller can then decide to market the data, for example, to the ten highest bidders in the auction process or determine any other distribution mechanism based on the prices bid by the interested parties. The beauty in our engines is that they are far from complete. Since they’re developed using advanced AI, the engines will grow as the ecosystem does; we will gather more data, host more interactions, and lead the data economy.

Platform Engines — The Machine Room of DACONOMY

0

platform-engines-the-machine-room-of-daconomy-1bbd1e632b8

2018-08-24

2018-08-24 10:46:57

https://medium.com/s/story/platform-engines-the-machine-room-of-daconomy-1bbd1e632b8

false

803

The first decentralized ecosystem for companies of all industries and sizes enabling a collaborative data and AI-driven analytical environment in which participants can buy, sell and exchange all types of data from any data protocol as a product or service.

null

daconomy

null

DACONOMY

socialmedia@daconomy.com

daconomy

CRYPTOCURRENCY,BLOCKCHAIN,DATA MONETIZATION,BIG DATA

daconomy

Data Science

data-science

Data Science

33,617

DACONOMY

null

eb9a8c1a3c4b

daconomy

5

1

20,181,104

null

0

null

0

9e42d8f54b99

2017-10-03

2017-10-03 22:35:40

2017-10-03

2017-10-03 23:06:47

1

false

en

2017-10-03

2017-10-03 23:06:47

5

1bbd7307cd60

3.135849

2

0

“The Future If,” is a global community of business leaders, authors and futurists who explore what our future can look like IF certain…

5

Is an AI-Driven Future with Universal Basic Income Possible? “The Future If,” is a global community of business leaders, authors and futurists who explore what our future can look like IF certain technologies, ideas, approaches and trends actually happen. The community looks at everything from AI and automation to leadership and management practices to augmented reality and virtual reality, the 4th industrial revolution and everything in between. Each we explore a new topic and this week we looked at whether we should Fear or Embrace AI, this was the discussion starter we used! Visit TheFutureIf.com to learn more or to request to join the community. Imagine a world where humans only work a few hours a week, robots and AI do almost all of the work, and we all receive a Universal Basic Income. Although it sounds very different from our current situation, it’s actually not that far fetched and could happen within the next 20–25 years. This week our community discussed the likelihood of this actually happening and what it would take for this hypothetical to turn into reality. There are a number of current trends that could contribute to this future scenario. First, think of the rapidly evolving technological progress that is exceeding our expectations. Technology that was new just five or 10 years ago is now completely obsolete, meaning that in a few decades we really could have machines and robots that could replace almost all human jobs. As technology grows and the world changes, we’re facing the difficulty of re-skilling employees fast enough. The things employees were trained on when they first started their jobs could be completely different now, and it is almost impossible to provide employees the skills they need as things change. That’s not even considering what it takes to prepare employees for new positions. On the flip side, robots and machines can learn new skills quickly and be re-trained almost instantly. Then there’s the money — Universal Basic Income pilot projects are currently be tested and are actually proving to be very successful. Under these systems, everyone gets the same amount of money on a regular basis no matter how much they work. However, many community members doubted that UBI would actually work when put into practice, with Principal Consultant Tamarah Usher saying the one-size-fits-all system goes against our human nature to strive and conquer. Others feared that people would abuse the system and said it wouldn’t work because of the emotional connection we have with money. UBI is very complicated, said Independant Researcher Michael Massey, adding that it would take the perfect combination of conditions to make the system even relatively successful. This scenario has its pros and cons. On one side, it could lead to class separation between the less advantaged doing manual tasks that can’t be easily automated, such as cleaning, said Global Head of Future Work Practices, Karen Eden, while people who work non-manual tasks wouldn’t have to work as much but would still get paid the same. On the other side, it would provide humans lots more time to focus on activities besides work. Community member Chris Sparey said working fewer hours would give people more time to invest in their local communities, health, and experiences. Community member Angela Lapré agreed, saying that she hopes people would share their skills with others by volunteering in the extra time they aren’t working. If we ever indeed find ourselves fortunate enough to live in a ‘four hour work week,’ it should spur us to do more good, or at least more than binge-watching shows all week,” she said. Even though AI would be doing all the work, many community members agreed it would be unlikely that humans would just sit back and not do any work. Instead, we would likely define new areas where can add value, but that those roles likely don’t currently exist. No matter if this exact scenario manifests itself, there’s no doubt that the future of our working and payment conditions will be vastly different in the future. The next generation could work in ways that would be unrecognizable to us. Starting the conversation now about what to expect and how to prepare can help ease the transition into the work environment of the future. Jacob Morgan is a best-selling author, speaker, and futurist. His new book, The Employee Experience Advantage (Wiley) analyzes over 250 global organizations to understand how to create a place where people genuinely want to show up to work. Subscribe to his newsletter, visit TheFutureOrganization, or become a member of the new Facebook Community The Future If…and join the discussion.

Is an AI-Driven Future with Universal Basic Income Possible?

6

is-an-ai-driven-future-with-universal-basic-income-possible-1bbd7307cd60

2018-01-22

2018-01-22 21:09:41

https://medium.com/s/story/is-an-ai-driven-future-with-universal-basic-income-possible-1bbd7307cd60

false

778

Best-selling Author, Futurist, and Leading Keynote Speaker on the Future of Work

null

FuturistJacob

null

Jacob Morgan

jacob@thefutureorganization.com

jacob-morgan

FUTURE OF WORK,CORPORATE CULTURE,JACOB MORGAN

jacobm

Economics

economics

Economics

36,686

Jacob Morgan

3x Best-Selling Author, Speaker, & Futurist. Founder of FutureOfWorkUniversity.com. Exploring Employee Experience, Future of Work & Leadership.

2cf55c094f29

jacobm

10,442

602

20,181,104

null

0

null

0

null

2018-03-28

2018-03-28 07:55:47

2018-05-14

2018-05-14 10:23:01

2

false

en

2018-05-14

2018-05-14 10:23:01

21

1bbe5f84eff8

4.583333

1

0

“Hey! You know what would be great. Why don’t we do a hackathon?”

5

Closing day at our weekend City Data Hack event in London (16–18 March 2018) Wider engagement and fresh ideas needed in your city? A City Data Hack is a great place start “Hey! You know what would be great. Why don’t we do a hackathon?” I don’t think any of us knew the impact that Laura’s suggestion was going to have on us over the course of the following 6-months. There we were, three people, trading ideas on how we were going to grow enthusiasm for improving city services through data, demonstrate different approaches to city innovation, engage a wide network of London innovators, support new businesses in testing their products and showcase some of our own software… all at the same time. “Okay then, let’s do a Hackathon”, replied Charlotte. “Great. Agreed! Has anyone done one of these before?” I asked. “No.” “No…” Fast-forward 6-months and here we are reflecting on our achievements and building on our momentum. We’re enlightened, enthused, and energised by what was achieved in 48-hours with the help of over 60 participants, 10 teams, 30 mentors, 9 partnering organisations, 60-days of planning, 10-days of pre-engagement with partners, 100 customised name tags, many almond croissants, 850 tweets, 1.5m twitter impressions, countless hours of sacrificed sleep. We’re in awe of the experience and idea-collision we curated for people in one great weekend at the Urban Innovation Centre. Was it worth it? Yes. Every city should consider doing something like this regularly if they want to support local data and digital innovation and change their own data culture. But hackathon’s aren’t new. In fact, in some quarters they have a bad rep because expectations have been poorly managed around what might be achieved from one in the past. If you’re expecting a product at the end of 48-hours, then lower your expectations now. If you want fresh ideas, wide participation, open innovation, culture change and relationship development with academics, business and public authorities in your city — do a hackathon. So, what did we achieve? We worked with three public sector organisations to define their pressing city challenges across Employment and Skills, Travel Planning, and Social Isolation: GLA, TfL, LBBD We brought together over 100 data specialists, designers, developers, and mentors to the Urban Innovation Centre in London to form multidisciplinary teams capable of developing product prototypes in 48-hours We brought service design and data science together by drawing on collaborative support from Snook and our in-house design team We provided a testing ground for the data products of 3 UK SMEs: Emu Analytics, Thingful, Space Syntax We also got to test out our new Tombolo products for data specialists: The Digital Connector and our Data Visualisation Application We demonstrated that an open approach to innovation can generate good outcomes for cities wishing to establish a stronger data culture and understand the art of the possible We learned that hackathons are a great way of mobilising a city’s data and digital innovators. Students, hobbyists, young professionals, public sector, private sector the passionate and intrigued What happened next? Travel Planning. Transport for London used the outcomes of the Hackathon to form their future strategy around digitising travel plans. They’re keeping the winning team, 0-One, engaged this summer to flesh out their idea further toward something that can be publicly commissioned by the transport authority. Social Isolation. London Borough of Barking and Dagenham invited their preferred team Come2Meet to present back to social workers in their council chambers and signposted them toward a grant funded opportunity to develop their concept further. They also opened the door for UCL CASA MSc students to collaborate with their data science team on real social / spatial problems for their thesis this summer. Employment and Skills. Greater London Authority invited the winners of their hack challenge to city hall to celebrate. They’re now more informed about how to take building an employment and skills knowledge platform for Londoners forward. The Digital Connector. Future Cities Catapult released the Tombolo Digital Connector, an open source tool for data specialists that helps standardise data preparation, combining and sharing. We gained invaluable feedback over the weekend and more awareness around the platform. City Data Visualisation. Emu Analytics helped us demonstrate the City Data Visualisation application. It’s one of a number of data platforms Emu Analytics have helped clients to build and the first they’ve built through collaboration with local government and a Catapult. Taking part in the weekend has led to other opportunities for them. More Data. Space Syntax and Thingful opened up their datasets to our participants providing them with the ability blend spatial layout and IoT data into their projects. Hackathon Legacy. We made as much of this hack as open as possible – there’s numerous content available online to give others ideas on how to replicate it. The joining information, the projects,the presentations, the weekend summary video — they’re all available to provide inspiration for others. Why should your city-region care about data? City-regions are muscling to find their appropriate position in the global value chain. They’re ensuring that they’re equipped with a competitive blend of data and digital capability, infrastructure and a vibrant ecosystem with high-skilled jobs capable of drawing more value to their regions. In February 2018, GMCA announced that its new Digital Strategy had been adopted across Greater Manchester. In the same month, West Midlands Combined Authority announced it is on the hunt to recruit its first Chief Digital Officer following an earlier announcement that WMCA will receive £800,000 over three years to create an Office for Data Analytics as part of its latest devolution agreement with central government. But there’s more that can be done. More Chief Digital Officer positions are needed. And will 2018 be the first year that city-regions consider appointing Chief Data Officers? By investing in data and digital themselves, city-regions become more authentic when encouraging others to invest in their region. City-regions that invest in skills, culture, support the development local business and strategically procure and commission software together will succeed in taking their places forward in sectors. Those who don’t will lag behind; fragmented and without a collective imagination capable of striking a strong modern identity. A hackathon is not a silver bullet or a panacea. But it is a good indicator of how much resource a city is devoting to data and digital innovation and how established its collaborative relationships are with local business, academic and research institutions. The team behind City Data Hack Back Row (from left to right): Lorena Qendro (Software developer, Tombolo), Laura Pye (Communications and engagement wizard, FCC), Charlotte Hutton (City data hack magician, FCC). Front Row (from left to right): Hemanshu Arya (Software developer, Tombolo), Eddie Jaoude (Seasoned hackathon sage & open source guru), Jon Robertson (me), Joseph Bailey (Data science lead, FCC). Other mentions: Katinka Schaaf, Aisling Conlon and Thanos Bantis I’m currently the Delivery Lead for Tombolo, organiser of City Data Hack, and leading Data and Digital projects at Future Cities Catapult. Follow me on Twitter or connect on LinkedIn.

Wider engagement and fresh ideas needed in your city? A City Data Hack is a great place start

1

wider-engagement-and-fresh-ideas-needed-in-your-city-a-city-data-hack-is-a-great-place-start-1bbe5f84eff8

2018-05-14

2018-05-14 16:31:24

https://medium.com/s/story/wider-engagement-and-fresh-ideas-needed-in-your-city-a-city-data-hack-is-a-great-place-start-1bbe5f84eff8

false

1,113

null

Innovation

innovation

Innovation

59,190

Jon Robertson

City or country? Glass or brick? Human or capital? These are the hard decisions we face. Leading data projects @futurecitiescat and @tombolo_. Views are my own.

1e5500a6ef96

JonSRobertson

16

36

20,181,104

null

0

null

0

94810dc56f79

2018-08-21

2018-08-21 12:21:02

2018-08-21

2018-08-21 12:24:37

1

false

en

2018-08-21

2018-08-21 12:24:37

5

1bbe8aa678ed

2.486792

0

"Millions of calculations are circumvented and it all happens at the speed of light…"

3

A.I. camera could help self-driving cars ‘see’ better Researchers have devised a new type of artificially intelligent camera system that can classify images faster and more energy-efficiently. The image recognition technology that underlies today’s autonomous cars and aerial drones depends on artificial intelligence: the computers essentially teach themselves to recognize objects like a dog, a pedestrian crossing the street, or a stopped car. The new camera could one day be small enough to fit in future electronic devices, something that is not possible today because of the size and slow speed of computers that can run artificial intelligence algorithms. “That autonomous car you just passed has a relatively huge, relatively slow, energy intensive computer in its trunk,” says Gordon Wetzstein, an assistant professor of electrical engineering at Stanford University who led the research. Future applications will need something much faster and smaller to process the stream of images, he says. Outsourcing the heavy lifting Wetzstein and Julie Chang, a graduate student and first author of the paper, took a step toward that technology by marrying two types of computers into one, creating a hybrid optical-electrical computer designed specifically for image analysis. The first layer of the prototype camera is a type of optical computer, which does not require the power-intensive mathematics of digital computing. The second layer is a traditional digital electronic computer. “Millions of calculations are circumvented and it all happens at the speed of light…” The optical computer layer operates by physically preprocessing image data, filtering it in multiple ways that an electronic computer would otherwise have to do mathematically. Since the filtering happens naturally as light passes through the custom optics, this layer operates with zero input power. This saves the hybrid system a lot of time and energy that would otherwise be consumed by computation. “We’ve outsourced some of the math of artificial intelligence into the optics,” Chang says. The result is profoundly fewer calculations, fewer calls to memory, and far less time to complete the process. Having leapfrogged these preprocessing steps, the remaining analysis proceeds to the digital computer layer with a considerable head start. “Millions of calculations are circumvented and it all happens at the speed of light,” Wetzstein says. Quick thinking In speed and accuracy, the prototype rivals existing electronic-only computing processors that are programmed to perform the same calculations, but with substantial computational cost savings. While their current prototype, arranged on a lab bench, isn’t exactly small, the researchers say their system can one day shrink to fit in a handheld video camera or an aerial drone. Related: 4D camera gives robots a wider view In both simulations and real-world experiments, the team used the system to successfully identify airplanes, automobiles, cats, dogs, and more within natural image settings. “Some future version of our system would be especially useful in rapid decision-making applications, like autonomous vehicles,” Wetzstein says. In addition to shrinking the prototype, Wetzstein, Chang, and colleagues are now looking at ways to make the optical component do even more of the preprocessing. Eventually, their smaller, faster technology could replace the trunk-size computers that now help cars, drones, and other technologies learn to recognize the world around them. Related: Ultra-thin camera design doesn’t need a lens The research appears in Nature Scientific Reports. Additional coauthors are from Stanford and King Abdullah University of Science and Technology in Saudi Arabia. The National Science Foundation, a Stanford Graduate Fellowship, a Sloan Research Fellowship and the KAUST Office of Sponsored Research funded the work. Source: Andrew Myers for Stanford University Original Study DOI: 10.1038/s41598–018–30619-y Find more research news at Futurity.org

A.I. camera could help self-driving cars 'see' better

0

a-i-camera-could-help-self-driving-cars-see-better-1bbe8aa678ed

2018-08-24

2018-08-24 19:23:37

https://medium.com/s/story/a-i-camera-could-help-self-driving-cars-see-better-1bbe8aa678ed

false

606

Research news from top universities

null

futuritynews

null

Futurity News

editor@futurity.org

futurity-news

SCIENCE,RESEARCH,HEALTH,ENVIRONMENT,TECHNOLOGY

futuritynews

Machine Learning

machine-learning

Machine Learning

51,320

Futurity News

Futurity.org

5bc60de02d60

Futurity

1,282

5

20,181,104

null

0

null

0

863f502aede2

2018-01-23

2018-01-23 00:13:04

2018-01-23

2018-01-23 00:21:35

3

false

en

2018-01-23

2018-01-23 00:22:01

4

1bbea0757

3.64434

3

0

January 7th — NVIDIA Announces Delivery Target For Self Driving Processor Xavier At CES 2018, NVIDIA CEO Jensen Huang announces the company…

4

AI Biweekly: 10 Bits from January (Pt 2) January 7th — NVIDIA Announces Delivery Target For Self Driving Processor Xavier At CES 2018, NVIDIA CEO Jensen Huang announces the company is planning to deliver the first sample of its Xavier processor during this quarter. NVIDIA DRIVER Xavier contains more than nine billion transistors, making it by far the most complex system on a chip. NVIDIA says that Xavier will power Pegasus, the AI computing platform designed to deliver Level 5 autonomous vehicles. January 8th — Facebook Plans to Shut Down Its Personal Assistant “M” on January 19th Facebook announces it will shut down its virtual chatbot assistant “M.” The AI chatbot in Messenger was only ever available to 2,000 Californians. Facebook says it learned a lot about how users interact with virtual assistants and plans to use that info to further develop its other AI services. Some M users say they never really had a reason to communicate with the bot. January 9th — Amazon Alexa Is Integrating With Windows 10 Amazon Alexa is taking a big step forward by merging itself into Windows PCs. Many OEMs at CES — such as Acer, ASUS and HP — announced they would make use of the Alexa App for Windows 10. The app will be limited to specific functions, and cannot call or message. Windows 10 integration does however enable Alexa to jump out of the Amazon ecosystem. January 9th — A US Media Team Plans to Launch Vital Intelligence Data Live A team of media executives in the US announces VIDL (Vital Intelligence Data Live) the launch this summer. The proprietary technology combines blockchain and AI to deliver accurate news reports from around the world. The platform will be fully automated and promises a foolproof system for delivering news stories that won’t be affected by human bias or warped into fake news. January 11th — Mercedes-Benz Releases In-car Assistant MBUX MBUX is the new Mercedes-Benz infotainment and multimedia system, a smart platform powered by NVIDIA GPU. It enables users to issue natural language voice commands rather than simple control commands, and can function without an internet connection. MBUX will be ready for all new Mercedes A-Class vehicles this year, and in other Mercedes vehicles in the future. NVIDIA Drive Xavier January 13th — Google Assistant Gets Much Attention At CES Google did not introduce many new products at CES 2018. Instead, it highlighted a bunch of tech partners — including Sony, LG, Lenovo, and Huawei — who promoted Google Assistant in various hardware applications, including TVs, refrigerators and even electric bike wheels. In the virtual assistant race, Google Assistant is closing the gap with Alexa. Google Assistant at CES January 16th — GM and Waymo Top New Self-Driving Leaderboard Navigant Research releases an Autonomous Driving Leaderboard which ranks the industry’s top brands. General Motors and Alphabet’s Waymo are in the top two positions. Meanwhile, former leader Ford has dropped to third, while Tesla sits at the bottom of the ranking with Apple. January 16th — Canada.ai Platform Showcases Canada’s AI Information NEXT Canada and a group of top Canadian artificial intelligence institutions launch Canada.ai. The web platform showcases Canada’s leadership in the field of artificial intelligence by presenting information such as research results, leading startups and companies in the AI space, and AI-related events happening across the country. Canada is welcoming many international AI researchers, students, institutions and startups. January 17th — AutoML Can Train Machine Learning Models Without Coding Google has announced the alpha launch of AutoML Vision, which helps build custom image recognition models. The new services available under the AutoML brand don’t require users to have machine learning expertise. The user provides images along with their tags, and AutoML Vision automatically creates a custom machine learning model based on them. Google is handling the hard work of training and tuning the model. January 19th — Google Unveils A New Office In Shenzhen, China Rather than Beijing or Shanghai, Google unveils a new office in Shenzhen. According to Google News, the reason for setting up in Shenzhen is to better cooperate with local partners. Google is gradually collaborating more with Chinese companies on products such as Lenovo’s VR set and self-driving cars. While Shenzhen is now in the spotlight, it remains uncertain whether Google can effectively leverage its new Chinese offices to penetrate the country’s AI market. Contributing Analyst: Alex Chen | Editor: Michael Sarazen Sync AI @ Synced | 机器之心 | Twitter Dear Synced reader, in our latest “Trends of AI Technology Development Report”, our tech analysts cut through the technical jargon and misleading media coverage to break AI down into easily understandable parts. Whether you are a researcher, tech enthusiast, entrepreneur, investor, or student, the report will help you sort out the basics and provide you with the necessary background to move forward. Click here to get the full report and subscribe to our upcoming AI newsletter.

AI Biweekly: 10 Bits from January (Pt 2)

52

ai-biweekly-10-bits-from-january-pt-2-1bbea0757

2018-05-01

2018-05-01 17:06:52

https://medium.com/s/story/ai-biweekly-10-bits-from-january-pt-2-1bbea0757

false

820

We produce professional, authoritative, and thought-provoking content relating to artificial intelligence, machine intelligence, emerging technologies and industrial insights.

null

SyncedGlobal

null

SyncedReview

global.sns@jiqizhixin.com

syncedreview

ARTIFICIAL INTELLIGENCE,MACHINE INTELLIGENCE,MACHINE LEARNING,ROBOTICS,SELF DRIVING CARS

Synced_Global

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Synced

AI Technology & Industry Review - www.syncedreview.com || www.jiqizhixin.com || Subscribe: http://goo.gl/Q4cP3B

960feca52112

Synced

8,138

15

20,181,104

null

0

null

0

6a8b96b0656b

2017-10-13

2017-10-13 23:35:12

2017-10-17

2017-10-17 16:01:01

6

false

en

2017-10-28

2017-10-28 00:56:04

4

1bbf0903768b

6.100943

786

15

2

Besides several similarities, there are 4 major distinctions between these two roles.

3

How Quantitative UX Research Differs from Data Analytics Besides several similarities, there are 4 major distinctions between these two roles. When we introduce ourselves as quantitative UX researchers, people often get curious about the “quantitative” part of our title. One question they ask is how our work is similar to, or different from, the work of product analysts or product data scientists. We hope this article will provide a comprehensive answer. Commonalities The confusion is understandable. For one thing, we both love data. We’re trained to conduct quantitative analyses with variety of different data sources, including experiments, surveys, and logged behaviors. Many of us, across both groups, come from quantitative disciplines such as social psychology, statistics, computer science, and economics, which help us drive insights and elevate our teams’ understanding of product usage through large amounts of data. At Facebook, the Data Analytics and the UX Research team share the same mission of explaining phenomena we observe in data. We both focus on making meaningful and interpretable inferences about data, relationships between variables, and explanations for changes or patterns in the data. (By contrast, machine learning engineers and those in other big data roles focus on predicting unknowns as accurately as possible.) Both groups use the same primary tool: statistics. We write code in data analysis software like R, Python, or SPSS. We spend a lot of time exploring and visualizing data to drive our hypothesis generation. We visualize more complex relationships of data points using libraries in R and Python. We also need to have knowledge of distributed data storage systems like Presto and Hive (some SQL knowledge is often sufficient to work with those tools). Most of our work ends up, hopefully, in presentations that clearly and concisely communicate our results. However, there are also fundamental differences between the two roles. Here are 4 major distinctions between quantitative UX research and data analytics. (Note: these differences are influenced by our work at Facebook and may not all apply everywhere.) 1. Human-Centric vs. Business-Centric For a technology company to perform well, it has to focus relentlessly on both improving business metrics and delighting its users. Insights about how and why metrics are changing help the company build better products and grow their business value. Understanding users — their motivations, their experiences, and how the product fits into their life — is also critically important. Quantitative UX research delivers insights about people. UX researchers often approach research projects with questions such as: What are the human motivations for using these products? How do people perceive and use the product? How do they react emotionally and physically to it? What do they like and dislike about specific features? What role does the product play in their daily life? Data scientists, on the other hand, often start by asking questions related to how the product is performing, or is expected to perform, in the marketplace. How does a product feature change behavioral metrics, such as clicks or time spent? How much adoption of the product did we receive on various devices? Which features are used and which ones are abandoned? Despite the different motivating problems, data scientists and quantitative UX researchers have similar workflows as they collect and analyze data in order to discover important interactions and relationships between technology and people. 2. User Intent vs. User Action User actions tell us about what is happening — for example, how many times they clicked on something, how often they come back to the app, or how much time they spent on the site. User intents, on the other hand, are about the relationship between people who use the product and (un)available product features. For example, out of 10 clicks, how many of those clicks were out of interest and how many were due to frustration? What brings users back to the app and how do they feel about it? How much of the time they spent on the site was time well-spent, and how much was spent looking for something they couldn’t find or trying to figure out a feature that wasn’t intuitive? Data scientists are less concerned with such questions than with with metrics and collective performance based on user actions or lack of action. They’re interested in the timing, variety, and magnitude of users’ signals — things like views, clickthrough rates, time spent, and churn. UX researchers, including quantitative ones, are mainly interested in understanding how people use our products, what problems they may have, and what works differently for them. Quantitative researchers seek to gain insights about the intent of people’s product usage through patterns in the data they collected. They also try to measure quality of experience using self-reported data (surveys) or behavioral data. While data scientists are more concerned about how many people used a new feature and what they did afterward, UX researchers aim to understand how many people used the feature in various contexts, what motivates them to use the feature, and how they felt about the experience. 3. Inference vs. Prediction Accuracy Like data scientists, quantitative UX researchers may use a multitude of statistical tools to gather insights from data. While the main suite of tools used by the two groups are roughly similar, each group uses the tools differently, since they’re pursuing different goals. Compared to UX researchers, data scientists are more often motivated to improve the predictive accuracy of their models. (The most accurate models are black box machine learning models, which are hard to interpret by their nature.) UX researchers are more often motivated by inference. In many cases, we’re not looking to predict future phenomena but to better understand the factors underlying experience or behavior. That’s why UX researchers more often use social science models that are more interpretable but have lower predictive accuracy. 4. Analyzing Survey Data vs. Logged Behavioral Data In data science, activity logs are the primary source of data. Quantitative UX researchers use a combination of log data and self-reported survey data. Depending on the research questions, we may use only one source of data or combine multiple methods of data collection and analysis. Analyzing survey data requires a different methodology than analyzing log data. To accurately make sense of survey data, the quantitative researcher must consider, and model, the survey design and data collection process. Survey data is therefore typically analyzed with some form of regression, in which survey design elements are incorporated into the model. For activity log data, Data Scientists typically consider these elements of data collection to be ignorable. Log data is frequently several orders of magnitude larger than survey data. As a result, overfitting and algorithm speed are critical issues in learning from log data that typically don’t arise in analyzing survey data. While regression models are also used in analyzing log data, Data Scientists frequently use methods involving regularization in order to handle issues introduced by data size. Working Together Despite the real differences between data science and quantitative UX research, there are undoubtedly many cases where the two roles are almost interchangeable. But they can also be highly complementary, taking full advantage of a diverse range of backgrounds and skills. In fact, at Facebook, some of the most impactful, satisfying, and fulfilling research projects are collaborations between the two. The need for both groups shows no signs of fading. The number and variety of meaningful research questions about the relationship between technology and people ensure that data science and quantitative UX research will continue to exist in parallel, driving the business and the technology forward to serve people better. While we’ve shared some of the ways Quantitative UX Research and Data Analytics are similar or different from each other, our observations are influenced by the context of our organization and the nature of our work. We welcome further discussions about these two roles in other organizations, and fields. What kind of problems do you solve as a Quantitative UX Researcher or a Data Scientist? We’d love to hear from you — please comment below! Authors: Mary Nguyen, Researcher at Facebook; Saide Bakshi, Researcher at Facebook; and Alex Whitworth, Data Scientist at Facebook (from left to right) Illustrator: Drew Bardana

How Quantitative UX Research Differs from Data Analytics

3,379

how-quantitative-ux-research-differs-from-data-analytics-1bbf0903768b

2018-06-17

2018-06-17 02:29:23

https://medium.com/s/story/how-quantitative-ux-research-differs-from-data-analytics-1bbf0903768b

false

1,365

Learnings from the people who study human behavior for Facebook

null

Facebook Research

null

facebook-research

RESEARCH,FACEBOOK,UX RESEARCH,UX

null

Data Science

data-science

Data Science

33,617

Duyen Mary Nguyen

null

5208ca4c80e

duyenmarynguyen

860

85

20,181,104

null

0

import math import torch import matplotlib.pyplot as plt m = 3 c = 1 noisiness = 10 num_points = 5 def get_next_batch( batch_size ) : x = ( torch.rand( batch_size, 1 ) - 0.5 ) * 10 y = ( ( torch.rand( batch_size, 1 ) - 0.5 ) * noisiness ) + ( x * m + c ) return x, y x,y = get_next_batch( num_points) x2 = x * x x3 = x2 * x x4 = x3 * x x5 = x4 * x x2plusone = torch.cat( ( torch.ones( x.size(0),1 ), x, x2, x3, x4, x5 ) , 1 ) R2, _ = torch.gels( y, x2plusone ) R2 = R2[0:x2plusone.size(1)] x_test,y_test = get_next_batch( num_points) plt.scatter( x.tolist(), y.tolist(), color='red' ) plt.plot( x_test.tolist(), y_test.tolist(), 'g*' ) l = torch.linspace( x.min(), x.max(), steps=50 ).unsqueeze(1) l2 = l * l l3 = l2 * l l4 = l3 * l l5 = l4 * l l = torch.cat ( ( torch.ones( l.size(0),1 ), l, l2, l3, l4, l5), 1 ) yl = l.mm( R2 ) plt.plot( l[:,1].tolist(), yl.tolist() ) plt.show() linear1 = torch.nn.Linear( 5, 1, bias=True ) criterion1 = torch.nn.MSELoss() optimizer1 = torch.optim.Adam( linear1.parameters(), lr=1e-6, weight_decay=0 ) linear2 = torch.nn.Linear( 5, 1, bias=True ) criterion2 = torch.nn.MSELoss() optimizer2 = torch.optim.Adam( linear2.parameters(), lr=1e-6, weight_decay=.30 ) for epoch in range( 1000 ): x_batch, y_batch = get_next_batch( num_points * 50 ) xb2 = x_batch * x_batch xb3 = xb2 * x_batch xb4 = xb3 * x_batch xb5 = xb4 * x_batch X = torch.cat( ( x_batch, xb2, xb3, xb4, xb5 ) , 1 ) optimizer1.zero_grad() linear_outputs = linear1( X ) loss1 = criterion1(linear_outputs, y_batch) loss1.backward() optimizer1.step() optimizer2.zero_grad() linear_outputs = linear2( X ) loss2 = criterion2(linear_outputs, y_batch) loss2.backward() optimizer2.step() ( x, y ) = get_next_batch( num_points*5 ) plt.figure() plt.scatter( x.tolist(), y.tolist(), color='red' ) l = torch.linspace( x.min(), x.max(), steps=50 ).unsqueeze(1) l2 = l * l l3 = l2 * l l4 = l3 * l l5 = l4 * l L = torch.cat ( ( l, l2, l3, l4, l5), 1 ) yl1 = linear1( L ) yl2 = linear2( L ) plt.plot( l.tolist(), yl1.tolist(), color='blue' ) plt.plot( l.tolist(), yl2.tolist(), color='green' ) plt.show()

26

null

2018-08-24

2018-08-24 00:20:33

2018-08-24

2018-08-24 00:23:23

2

false

en

2018-08-24

2018-08-24 00:37:21

0

1bbf2fb42cfc

2.972013

0

It is possible to build a model that represents the training data perfectly. If a predictive or discriminatory function is desired, the…

3

Regularization It is possible to build a model that represents the training data perfectly. If a predictive or discriminatory function is desired, the desired model must represent unseen data accurately. Regularization is the process that simplifies a model, to allow it to more accurately describe the entire dataset. Techniques to regularize the model include: penalize large weights (θs) in the model: weights contribute to the the loss function (regularization) discard a fraction of the learned model periodically (dropout) In advance, it is noted that the results from this were not expected. It was expected that regularization would improve model generality. It appears to do so, but only in the central section. A better title for this section might be overfitting and memory. Define some parameters to create a noisy straight line. The regressions will find a best fit line. Then define a function to provide X & Y values, which lie in the distribution. Expand the x data to a 5th order polynomial, use linear regression to fit the enhanced data. Since num_points =5, using a 5th order polynomial will perfectly describe the data. Draw the curve The training data (red) matches the curve (blue) perfectly. Data drawn from the same distribution (green) is a less good fit. Note how the simpler model ( a straight line ) would describe both data sets better. This is an example of overfitting, the model becomes a memory of the training data instead of a generic representation. A learning algorithm/network can generalize or remember! There’s not much opportunity in the direct linear regression to simplify the model, other than reducing the polynomial order. Considering the case of the gradient based models, regularization is a feature of the optimizer. In pytorch the weight_decay parameter defines the cost of model complexity. Define two models, with and without the regularization feature. Train them both and plot the difference. Based on this, regularization didn’t simply the model enough. Improvements in this area appear to be needed.

Regularization

0

regularization-1bbf2fb42cfc

2018-08-24

2018-08-24 00:37:21

https://medium.com/s/story/regularization-1bbf2fb42cfc

false

686

null

Machine Learning

machine-learning

Machine Learning

51,320

richard Corbishley

null

409ec80dd31c

rcorbish

0

1

20,181,104

null

0

null

0

null

2018-06-19

2018-06-19 15:51:45

2018-09-22

2018-09-22 16:34:40

8

false

en

2018-09-22

2018-09-22 17:05:14

0

1bbf44cfd29f

7.337107

1

0

How DeepMind’s Generative Query Network brings AI one step closer to human level imagination.

5

Getting Computers to Imagine. How DeepMind’s Generative Query Network brings AI one step closer to human level imagination. Introduction The idea of the generative model is perhaps one of the most significant ideas, not only in the field of artificial intelligence, but in the entirety of the technological era. Since the creation of the Boltzmann Machine (Geoffrey Hinton and Terry Sejnowski, 1985) and the Autoencoder ( Dana H. Ballard, 1987), exponential progress has been made within the field of generative models. Today, we have numerous types of generative models, ranging from variations of classical generative models such as Variational Autoencoders (a type of autoencoder), and new types of generative models such as Generative Adversarial Networks (Ian Goodfellow, 2014). Less than a week ago, Google DeepMind published a paper describing a new algorithm which they call the Generative Query Network, or the GQN. This new algorithm is capable of “imagining” and generating 3-dimensional models of a scene from various viewpoints, from observing only a few 2-dimensional pictures of various locations within this scene. A GIF from DeepMind’s blog post showing the observation image which is inputted into the GQN (left)and the three-dimensional model/rendering of the scene outputted by the network (right). What is a Generative Query Network? For a human, it only takes a few glances and milliseconds for our visual cortex to interpret and make inferences regarding a certain scene; within a matter of seconds, the brain uses the information which it has gathered to “imagine” and construct a model or representation of the room. For example, when a person walks into a room and sees only part of a table due to a certain obstruction, he or she can infer that the full table is present despite not being able to see the complete table, and can even visualize how that table might look. Similarly, if we look at any person’s face from the right side, we would easily be able to visualize what that person’s face might look like from the left side as well. This is exactly what DeepMind has got its Generative Query Network to accomplish. The Generative Query Network is capable of learning about the details of the objects within a certain scene from the color and texture of individual objects to the lighting and spatial relationships between the different objects, only from a few snapshots of the scene. Using this information, the network is able to render a three-dimensional model of the scene from a query viewpoint(a human given input to the network which asks it to render the scene from a certain viewpoint) which is fed into the network, hence the name Generative Query Network. Another GIF from DeepMind’s blog which shows the neural rendering of objects which are generated by the GQN after it is given a few observation images of the objects from certain viewpoints. As stated by DeepMind in their blog post: “ Much like infants and animals, the GQN learns by trying to make sense of its observations of the world around it. In doing so, the GQN learns about plausible scenes and their geometrical properties, without any human labelling of the contents of scenes.” As you may have noticed from this quote, the problem which a GQN is solving is an unsupervised learning problem: the network gets no explicit details regarding the scene, its contents and properties, and learns everything from what it “sees” in the training images of the scene which it is given. This, according to me, is perhaps the most interesting aspect of this algorithm, as it means that the GQN, like the mind of a human baby, perceives the world and innately makes sense and generalizations of its observations. If you think about it, it makes intuitive sense: imagination and visualization in general are unsupervised learning tasks: they are spontaneous and unguided, and do not depend on any explicit, externally given direction. Explaining the Generative Query Network The GQN is primarily made up of two networks: a representation network and a generation network. The job of the representation network is essentially to make sense of the data (snapshots of a scene) given to the network and the job of the generation network is to actually visualize and create a 3-dimensional model using the knowledge gained by the representation network. More specifically, the representation network takes the observations of the scene as an input and creates a representation vector which is packed with information describing the scene; you can think of it as the encoder segment of an autoencoder producing a latent space representation of the input image given to the autoencoder. The generation network then uses this representation vector to construct the three-dimensional scene and to visualize the scene from various, previously unobserved perspectives; you can think of this as the decoder segment of an autoencoder creating a reconstructed image using the aforementioned latent space representation. The architecture of the Generative Query Network, with the Representation Network creating a neural scene representation (representation vector) from the input to the network and the Generation Network using the same to predict/render the scene from a different viewpoint. In order to give the generator network the most amount of data in order to make the most accurate model it possibly can, the representation network must be able to extract as much information as possible from the scenes which it is given, and must pack as much of this information into the representation vector. The representation network not only captures data regarding the positions of objects within scenes, but also the objects’ orientation, color, and texture, as well as the general layout of the room. As the generator network also learns about objects, features, relationships and patterns within the scene during the training phase, the representation network can create an extremely compact representation vector which only contains the most essential information regarding the scene, using which the generator network can create the scene filling in minor gaps wherever necessary. This animation from DeepMind gives a quick run through of what the GQN is doing. The animation shows how, given a few observations of a scene from a certain viewpoints, the GQN can learn the scene and can be queried to render the scene from various, previously unseen viewpoints of the scene. We see that the GQN successfully and accurately renders the scene from given viewpoints from not only different angles, but different locations in the scene as well. Implications of the Generative Query Network As DeepMind writes in their blog: “ The GQN’s generation network can ‘imagine’ previously unobserved scenes from new viewpoints with remarkable precision. When given a scene representation and new camera viewpoints, it generates sharp images without any prior specification of the laws of perspective, occlusion, or lighting. The generation network is therefore an approximate renderer that is learned from data.” Additionally, the researchers from DeepMind say that the “ GQN’s representation network can learn to count, localize and classify objects without any object-level labels.” Despite the fact that the representation vector generated by the representation network is compact, the generation network’s constructions are essentially indistinguishable from the truth; this goes to show the extent to which the representation network can accurately perceive and extract information from the data which it is provided with. The generator network is also capable of accounting for uncertainty and and inconsistency within training data, as it can still construct a model of a scene even when it only has a few partial views of a certain part of the scene by stitching together these partial scenes and relating the information which it gathered from the partial scenes to combine these scenes in a way which is most representative of the ground truth. This GIF from DeepMind’s blog shows how the GQN renders a scene by accounting for the uncertainty caused by the limited observations of the scene. The GIF also shows how the GQN updates its neural rendering of the scene as the number of observations of the scene increases and its uncertainty regarding the ground truth scene decreases. The capabilities of the GQN can allow for much easier training of other deep learning models, especially in the fields of computer vision. Using the GQN to train object detection neural networks could drastically reduce the amount of training data required for the models to learn and could help object detection models achieve much higher performance with regards to understanding spatial relationships between objects in an input image. Furthermore, researchers at DeepMind highlight that using the representations produced by the GQN could allow “state-of-the-art deep reinforcement learning agents learn to complete tasks in a more data-efficient manner compared to model-free baseline agents”. The representation vector generated by the GQNs representation network could be used by RL agents as “innate” knowledge of the environment of the agent. The above image from DeepMind’s blog shows how the representation generated by GQNs can be used for extremely data efficient reinforcement learning. We see that compared to other methods such as learning from raw image pixels, the RL agents can learn at a profoundly quicker rate using the representation vector of GQN. This same idea can be applied in other fields as well such as computer vision. All of these characteristics of Generative Query Networks makes it an extremely powerful an valuable tool which can be applied in an vast number of fields. As stated above, GQNs could prove to be a significant step forward in the development of fully autonomous vehicles. Normally, neural networks used in self-driving cars use hundreds of millions of images and take hundreds of hours training. However, using the information-packed representations by the GQN, in addition to its capabilities to learn relational inference and learning spatial relationships between objects, in autonomous vehicle systems could drastically decrease the training time and increase the accuracy of these systems when it comes to tasks such as object detection and semantic scene segmentation. Furthermore, GQNs could also be applied in the fields of virtual reality, augmented reality, and in simulation/game engines for “querying across space and time to learn a common sense notion of physics and movement”, which could be used ton make more realistic video games and simulations. As the most important aspect of autonomous cars is to understand their surroundings or the “scene” wherein they are located, the GQNs scene representation and understanding abilities has immense potential to be extremely useful within this field. However, I believe that perhaps the most powerful, influential and scary application of GQNs is to combine its autonomous scene understanding capabilities with the learning capabilities of state-of-the-art reinforcement learning algorithms within robots to create truly autonomous, self-aware robots which would be able to explore, understand and learn from scenes without any sort of human guidance. Similar to the robots from the Terminator movies, these robots would be hyper-aware of their surroundings, visualize and imagine hypothetical scenarios, and use the same to make extremely accurate decisions and predictions in order to achieve their goals. The creation of such robots could either be the greatest innovation in the entirety of mankind, or the reason why mankind will cease to exist…. Could GQNs be the first step in the creation of self-aware autonomous robots, such as those from the “Terminator” movie franchise? Conclusion In its blog, DeepMind has mentioned that despite all of its successes, GQNs are still under development, and have numerous limitations, most notably, their inability to recreate scenes with an extremely high complexity, and the fact that until now it has only been tested on rendering simple, synthetic (human generated) scenes. Nevertheless, the development of GQNs represents a gigantic leap in not only scene understanding and rendering, but but computer vision as a whole, and the significance of this discovery and its potential for the future is undeniable. As of today, only time will be able to tell us whether the aforementioned scenarios will remain as science-fiction, or become a reality.

Getting Computers to Imagine.

1

getting-computers-to-imagine-1bbf44cfd29f

2018-09-22

2018-09-22 17:05:14

https://medium.com/s/story/getting-computers-to-imagine-1bbf44cfd29f

false

1,644

null

Generative Model

generative-model

Generative Model

55

Vinay Kanigicherla

15 Year Old. Futurist. Deep Learning and AI Enthusiast.

1fd40a2b543a

vinay.kanigicherla

2

15

20,181,104

null

0

null

0

null

2018-08-10

2018-08-10 12:14:26

2018-08-10

2018-08-10 12:17:35

3

false

en

2018-08-10

2018-08-10 12:17:35

0

1bbf7faad9e4

10.017925

0

(originally at asnaninishit.wordpress.com)

5

Machine Learning at CSE, IITK — The Culture, Research and Prospects (originally at asnaninishit.wordpress.com) You could be a curious soul wanting to know about machine learning (ML) at IIT Kanpur. Or a current IITK student wanting to get into ML. Or a prospective undergraduate / graduate student. Or someone else with a keen interest in this subject. If you want to roam around beings that pride themselves on designing advanced AI that can see through walls and beat you in your favorite board game and sing your favorite songs to you, look no further — IITK is here. Um.. just kidding. No seeing through walls here. But anyway, read on to find out who we are and what we do and what we hope to achieve. Who am I, and Why I Am Probably Not Lying I am a recent graduate (graduated in June 2018) from Computer Science and Engineering (CSE), IIT Kanpur, and spent the best part of these four years working on problems in machine learning either through courses, projects or internships. I have had the good fortune of taking courses with numerous faculty members in ML at IITK and working on research projects with a couple of them as well. So while my view of the ML culture at IITK may not be truly comprehensive, it is not entirely ill-informed either. I have been on the other side of the spectrum as well, albeit occasionally. I have mentored student projects in machine learning as part of courses as well as through the freshmen semester project initiative by the Association of Computing Activities (ACA), CSE IITK. I was also an instructor for a two-week intensive summer school course in machine learning organised by ACA in May-June 2018, through which I had a first hand glimpse of the machine learning landscape among (mostly) undergraduates across engineering institutes in north India. Again, not an experience enough to build a truly comprehensive picture, but definitely for an indicative one. A Bit of a Primer on IITK and Machine Learning IIT Kanpur has a rich legacy of being the pioneer of sorts in computer science, being the first institute in India to start an undergraduate program in the subject. In the recent years, machine learning has emerged as one of the most promising areas in computer science thanks to the availability of huge amount of data, as well as computing resources in the form of GPUs and powerful CPUs. IITK has not been sitting back, and the machine learning culture in the institute has taken off as well. This piece aims to paint a picture as close to being accurate as possible of this culture and how people try to make the most of it. Just a disclaimer at this point — ML is not just confined to computer science. Its theoretical foundations are developed by statisticians and electrical engineers alike, and its applications are as widespread as one can dream of. Healthcare, law, education, security, finance — these are just some sectors that have seen quite a disruption due to the advancements in machine learning. Another disclaimer, the views in this post are purely my own and need not be construed as being the views of the institute. Now that we have the motivation of this post and a little history out of the way, let’s dive in and see what lies under the surface. Faculty Members, Courses and Research Areas Q. How many faculty members in CSE, IITK work on machine learning, artificial intelligence or allied fields, and who are they? A: The CSE department at IITK boasts of having excellent faculty members approaching ML from both the ends of the spectrum: Prof. Purushottam Kar comes from the optimization background, while Prof. Piyush Rai works in probabilistic modeling and inference. This is a unique blend, rarely achieved by even the best universities in the world in computer science. We have Prof. Vinay Namboodiri, who works in computer vision and graphics, Prof. Nisheeth Srivastava, working in computational cognitive science and computational social science, Prof. Sunil Simon working on game theory, Prof. Swaprava Nath who is involved in mechanism design and multi agent systems, and Prof. Harish Karnick, who has been working on diverse aspects of machine learning theory and applications over the years, and has most recently been involved in ongoing work in natural language processing. Q. What are the venues where machine learning research from IITK is generally published, and how often? A: Our research gets published in top tier journals and conferences in machine learning, namely ICML (International Conference on Machine Learning), NIPS (Neural Information Processing Systems), KDD (Knowledge Discovery and Data Mining), JMLR (Journal of Machine Learning Research) and a few others. A considerable portion of the research is also presented in top tier conferences in the applications of machine learning or allied fields, like in computer vision (CVPR, WACV), Artificial Intelligence (AAAI and IJCAI), statistics (AISTATS), data mining (ICDM), web (WWW), software engineering (ICSE) etc. Most of these venues have at least 2 to 3 publications from IITK every year, with the ones in core machine learning (ICML, NIPS etc.) usually having a few more. The volume of publications is growing with more student and industry collaborations and the increasing faculty strength. Q. Is the research environment conducive to collaborations? How significant is the student involvement in research? A: Our faculty has collaborations with researchers from some of the top universities of the world, including those from USA, Canada, China, Spain, Italy and India. Industrial collaborations from research institutes at IBM, Microsoft etc. have also been blossoming. Over the last few years, student participation in machine learning research has grown manifold, especially among the undergraduates. For instance, requests for enrolling in CS771, the introductory machine learning course, have grown significantly over the last half a decade, and stand at over 300 now for a single offering. Many of the students who complete the course end up conducting research in ML or using ML for a project at a research lab, company or in campus itself. Q. What kind of projects do faculty and students work on? A: Quite a few projects are currently going on in the broad domain of ML at the institute, some of which are interdisciplinary in nature as well. A recent example of one such project pertains to using machine learning to improve the C compiler that is used for the first year programming course (ESC101), which has had a real impact on the way the programming labs are conducted and the way students learn their first bits in writing code. There are other projects which explore problems at the heart of machine learning, like learning under uncertainty, learning with millions of examples, robust learning and the like. Most of the instructors are quite flexible and allow even undergraduate students to pick projects of their liking, if they have some alignment with the faculty member’s area of interest. Q. What sort of work is carried out in deep learning at IITK? A: Deep learning has seen enormous progress in its techniques and success stories over the last few years, and is certainly a key area of research at IITK. It creeps into many learning problems as a viable solution, and a lot of work is done at IITK in deep learning, deep probabilistic modeling, optimization techniques for deep learning, deep learning theory, etc. Moreover, its effects can be most strongly felt in its applications, and research at IITK is not behind either — many of our recent projects and publications employ deep learning models for getting state of the art results in daunting problems in computer vision, natural language processing, recommender systems, speech processing, robotics and finance. Q. How many and what sort of courses are offered at IITK that are related to machine learning? How easy is it for a student to enroll in these courses? A: The various course offerings in machine learning follow the trajectories of the faculty members involved themselves. CS771 is the introductory machine learning course that covers a breadth of topics in supervised and unsupervised learning, and is generally offered once a year by one of the instructors listed above. It commands participation from a healthy mix of undergraduates and graduate students, and it is in this course that many intra institute work partnerships form, as students work in teams for their course projects. It aims to consolidate students’ understanding of the fundamental concepts in machine learning, and introduce them to traditional methods, new approaches, various learning paradigms, and related application areas. I have had the good fortune of taking the course once, and then being a project mentor for a few student groups the following year. I’m not sure how many institutes in India allow undergraduates to have that privilege. The other courses commonly on offer are (based on the courses that have been offered in the last three years): Probabilistic ML: Probabilistic machine learning, Bayesian ML, Topics in Markov Chains etc. Optimization: Online Learning and Optimization, Optimization Techniques, Learning Theory Applications: Visual Recognition, Topics in Computer Vision, Natural Language Processing, Multi-agent systems HCI: Computational Cognitive Science, Human Centered Computing A request to enroll in any of these courses is not a guarantee of being accepted into it. For the advanced machine learning courses, prior experience in terms of courses and / or projects helps in getting an enrollment request accepted. For CS771 though, the faculty members try to maintain a good proportion of non-CSE students in the course, but due to high demand, not everyone who requests the course gets accepted into it. It is a basket course (if you don’t know what that means, think of it as being semi-mandatory) for the CSE students, so it is generally a little easier for them to do the course. But this should not discourage anyone with a genuine interest in ML to pursue the field at IITK — if you have interest and something to show for it, most people at IITK are ready to extend a helping hand. It is to be noted here that various courses in robotics, neural networks, signal processing etc. are offered in the electrical engineering department, and might be of interest to the ML crowd as well. So would courses in regression, time series analysis, stochastic processes etc. for those who have a more statistical bend of mind (generally offered by the Mathematics / industrial and management engineering departments). Research Groups and Activities Q. If I am a student at IITK interested in machine learning, how do I find other students like me, and who would be able to help me find my way around? A: The Special Interest Group in Machine Learning, popularly known by its acronym SIGML, is a loose group of students and faculty members on campus who organize and take part in activities pertaining to machine learning. (https://www.cse.iitk.ac.in/users/sigml/) This includes talks, seminars and meetups with eminent researchers from top universities and research institutes, founders, engineers and managers of companies playing big in ML, and others from academia and industry who are advancing the field with their innovations and implementations. Hackathons are held annually in collaboration with other interested parties in industry and on campus to promote a culture of using machine learning to solve common problems. SIGML members sometimes meet among themselves as well, and discuss recent research in areas of common interest or papers that have the potential to cause ripples in the community. Q. How do students working in ML come to know about the work of others on campus in the field? A: We also celebrate a machine learning research day (MLRD) on campus every year, where students showcase their research work carried out at IITK and in various other institutes during internships and exchange visits. SIGML, MLRD and other such initiatives help unite the community of people working in ML on campus, helping people make connections, troubleshoot their problems with help from others and leverage the department infrastructure and mentorship to learn and build valuable technical skills. Oh, did I mention that we like to crack lame ML jokes every now and then just because we can do no better? And sometimes some people make memes as well. One of which (made by Parth Sharma, a batchmate of mine) was cited by Yann LeCun from his official Facebook page. Beat that! The said meme, by Parth Sharma Future prospects, internships and exchange Students pursuing machine learning at IITK find themselves quite sought after among industrial internships as well as placements. While this may or may not be a causal relationship, it is hardly a doubt that most big companies and startups alike want engineers and software developers who have a machine learning background and are opening up new job positions for the same. Data science seems to be the buzzword of the decade. Or words. Q. What sort of companies do students in ML from IITK get placed into? A: While the research labs at Microsoft, IBM and Adobe stand out among the recruiters from campus for their excellent research infrastructure in India, there are positions available for students in ML at Google, Tower Research, Goldman Sachs, Amazon, Flipkart, Optiver, WorldQuant, Uber, and other top financial and technological giants as well. Many startups, including ShareChat, Zomato and the like, also hire graduates from IITK proficient in machine learning. Q. What are the prospects for higher education / research internships in machine learning after graduating from IITK? A: Research internships and PhD positions, while appearing harder to get, are still very much in sight. Students tend to gravitate towards Microsoft Research, EPFL, NYU, Duke University, CMU, IBM Research etc. for potential destinations to strengthen their research background in ML through internships and exchange programs. Students from the recent batches have worked at length with faculty members at IIT Kanpur before embarking on a three- or six-month exchange program. Potential MS / PhD destinations, judging by the recent past, are spread throughout the globe with most concentration at places that are considered to be generally good in AI — U of Toronto, Carnegie Mellon, University of Montreal, Aalto University, UC Berkeley, Princeton, NYU, Oxford, Georgia Tech, Stanford, EPFL, Duke, etc. constitute a formidable set of universities where our recent graduates are pursuing their higher education in AI / ML and related fields. Before Signing Off If one is into machine learning, or wants to get a grip on the area, IIT Kanpur may turn out to be one of the finest places in India to pursue those goals. Then again, if you are into a specific subfield, a more detailed study of the faculty members, labs and the peer group at various places might be of help in order to compare across institutes. I have seen people from other IITs and IIIT Hyderabad do a great job in ML as well, so those might be other potential places to compare with, but an institute’s eagerness to push the boundaries of research and to collaborate across fields should also be taken into consideration.

Machine Learning at CSE, IITK — The Culture, Research and Prospects

0

machine-learning-at-cse-iitk-the-culture-research-and-prospects-1bbf7faad9e4

2018-08-10

2018-08-10 12:17:36

https://medium.com/s/story/machine-learning-at-cse-iitk-the-culture-research-and-prospects-1bbf7faad9e4

false

2,509

null

Machine Learning

machine-learning

Machine Learning

51,320

Nishit Asnani

Graduate Student at Stanford University. Interested in Machine Learning, entrepreneurship and literature.

7ea91be267a0

asnani04

147

203

20,181,104

null

0

null

0

eab1500c3930

2018-09-26

2018-09-26 02:16:10

2018-09-26

2018-09-26 02:30:32

1

false

pt

2018-10-10

2018-10-10 15:39:52

8

1bc0cc55df46

2.928302

0

Palestrante: Maria Bofill, sócia da TozziniFreire Advogados.

3

Advocacia 4.0: O futuro do Direito impactado pela tecnologia Palestrante: Maria Bofill, sócia da TozziniFreire Advogados. Hub: Nescafé Horário: 13h ___________________________________________________________________ Overview da palestra: A premissa central da palestra era comentar sobre o direito inserido na revolução digital chamada então de advocacia 4.0 e entender como o futuro/presente do direito está sendo impactado pelas novas tecnologias. ___________________________________________________________________ Highlights: Para iniciar o debate Maria cita o futurista Isaac Asimov: “A ficção científica de hoje é o fato científico de amanhã.” E como isso afeta o direito? Nessa primeira análise Maria traz alguns dados secundários como por exemplo: 1. Na Alemanha um funcionário da volkswagen foi morto por um robô. Ela questiona o seguinte: De quem é a responsabilidade agora? É da Volkswagen, ou é de quem programou o robô? 2. Um outro exemplo foi o chatbot da Microsoft que fazia tweets racistas. “são novas questões jurídicas que devem ser levadas em conta.” Ao ver essas noticias a gente já consegue fazer uma reflexão referente aos impactos de legislação dentro do cenário do direito e como a gente pode repensar essas questões. “Se as novas tecnologias existem, elas podem ser aproveitadas para profissão jurídica.” Essa afirmação nada mais é do que como a tecnologia pode melhorar também a profissão jurídica, ou seja, será que se faz necessário hoje em dia as empresas relacionadas ao direito contratarem diversos funcionários para ficar revisando contratos? Um robô não poderia fazer isso? Outro exemplo que a palestrante traz que ilustra muito bem esta questão é: Inteligência artificial vence 20 advogados em teste de revisão de contratos. Nota-se que o advogado do futuro ele vai ter diferentes habilidades, conhecimentos, formação, capacidade e cultura. Mais um exemplo se dá através da FGV que está dando aula de programação no direito. Já na segunda parte da palestra Maria indica os principais cenários que mais afetam a área do direito hoje, sendo eles: 1. Inteligência Artificial 2. Carros Autônomos 3. Blockchain 4. Data Privacy Um outro aspecto também observado pela palestrante foi que no mundo tecnológico, não há limites, ou barreiras para o empreendedorismo, ou seja existe uma grande movimento de parcerias com startups, grandes empresas e fundos de investimento que estão atrelados ao este mercado do direito. Maria cita um exemplo de um cliente/empreendedor que compareceu ao escritório dela indicando que ele tinha um projeto de tecnologia que era uma startup, mas eles não tinham dinheiro suficiente para pagar os honorários pois eram muito altos, em consequência disso o cliente acabou desaparecendo, porém dois anos depois eles foram vendidos por vinte milhões. E aí se observa um grande movimento de comportamento. Então quem são os nossos clientes? Gerações: Millenials x Centennials E para finalizarmos o raciocínio percorrido até o momento se faz necessário o seguinte questionamento: Qual será o papel do advogado do futuro? Na visão de Maria Bofill o advogado do futuro terá que ser muito mais estratégico e empático, ou seja, terá um aumento significativo nas relações humanas tendo que por exemplo estudar mais filosofia, sociologia e atividades como programação e ciência de dados. E agora eu pergunto para você: está preparado? Curiosidades: 1. o público que assistiu a palestra era composto por: Advogados, publicitários e área de compras; 2. Brastemp possui um serviço de assinatura de água; 3. Ambev investiu em uma máquina de bebidas; 4. Maria recomendou a leitura do livro — Direito das Startups composto pelos autores Bruno Feigelson,‎ Erik Fontelene Nybo e‎ Victor Cabral Fonseca 5. “Sempre que você ver um produto gratuito na internet fique esperto porque na verdade o produto é você.” ___________________________________________________________________ Reflexões do Trendspotter: O direito está intrinsecamente no dia a dia das pessoas, empresas, startups e de forma geral também no coletivo. Entender esses impactos que a tecnologia vem causando em uma esfera global se faz necessário em qualquer área de atuação. Os principais pontos que podemos levar em consideração passam pelo âmbito da educação, comportamento, na reação dos indivíduos enquanto a tecnologia como ferramenta para efetuar o trabalho braçal que hoje ainda é muito realizado pelo homem, principalmente em indústrias. Dessa forma, nota-se que a inovação no formato de trabalho dentro das organizações pode gerar sim uma disrupção em diversos aspectos do setor, tornando o ato de fazer muito mais estratégico e empático conforme comentado por Maria Bofill. ___________________________________________________________________ Trendspotter: Rafael Marques Barboza, assistente de Planejamento na Global.

Advocacia 4.0: O futuro do Direito impactado pela tecnologia

0

advocacia-4-0-o-futuro-do-direito-impactado-pela-tecnologia-maria-bofill-1bc0cc55df46

2018-10-10

2018-10-10 15:39:52

https://medium.com/s/story/advocacia-4-0-o-futuro-do-direito-impactado-pela-tecnologia-maria-bofill-1bc0cc55df46

false

723

BS Festival Trendspots by GPRS

null

BS Trendspots by GPRS

null

bstrendspotsbygprs

null

Advocacia

advocacia

Advocacia

113

Grupo de Planejamento RS

Representante dos profissionais de planejamento, estratégia e gestão de marcas no Rio Grande do Sul

a1ec50072fc3

GrupoPlanejRS

47

3

20,181,104

null

0

null

0

null

2018-04-04

2018-04-04 07:00:26

2018-04-04

2018-04-04 07:06:27

0

false

en

2018-04-04

2018-04-04 07:06:27

0

1bc1505e7b38

1.290566

0

Great financial metrics of @Accenture only prove it has accomplished higher operational efficiency. I believe, it’s primarily on account…

1

#DeepAnalysis, #SWOT, #BusinessModel, #RealityCheck, #FACTCHECK, ..: #Accenture Great financial metrics of @Accenture only prove it has accomplished higher operational efficiency. I believe, it’s primarily on account superior branding & other #shady practices. #Accenture isn’t the torchbearer of automation, #ML(#AI), digitization, .. . It’s outsourcing King. People ask me, in that case why #Accenture #WNS #FIS #Convergys #Capgemini feel compelled to pedal lies about #AI, #robotics, .. ? Well, most listeners, analysts, prospective employees can’t comprehend what they really mean and believe they indeed have something special to offer. Finally, people ask me what do they really offer? Well, exactly what they offered earlier - flexible contract labour. The cloak of branding helps them attract employees, justify higher price, .. . Long-standing clients know this. Yet, they deal with them to line their own ....... pockets, enjoy special privileges, benefits, .. . It’s a mutual benefit club at managerial & seniormost level. Others in outsourcing business are able to offer same & better outcome for a fraction of the cost. Also, Accenture .. has changed name several times & a chequered past. Let me also take this opportunity to clarify #ML(#AI). Unlike regular programs ML programs typically classify data, learn from experience. To do so, they undergo training on test data initially. Thereafter there are mostly on a self-learning course; new approach to programming. So, would those like @QuessConnect @Teamlease .. manage to catch-up with those like @Accenture? Maybe, maybe not. Branding, relationships, .. marketing budgets, scale, .. of leading MNCs rule out that possibility for a very long time. Interesting reading: https://t.co/Fh5Up8aD4p WindowsForDummies(Everything Accenture does): Accenture - Design and ConsultingAccenture - Build and implementationAccenture - Application ServicesAccenture - Maintenance ServicesAccenture - Testing ServicesAccenture - Outsourcing ServicesBuild and ..................….…..... CONTD .... implementation (including development and integration services)#Accenture - Cloud based Services #Crooks anyway enjoy best #camaraderie; self interest is the best binding glue. Also, it’s impractical for one guilty #crook to blow the whistle on other(/s). So, .. #intelinet, .. will continue to have a field-day atleast, until something snaps & their attempts to cover-up fail.

#DeepAnalysis, #SWOT, #BusinessModel, #RealityCheck, #FACTCHECK, ..: #Accenture

0

deepanalysis-swot-businessmodel-realitycheck-factcheck-accenture-1bc1505e7b38

2018-04-04

2018-04-04 07:06:28

https://medium.com/s/story/deepanalysis-swot-businessmodel-realitycheck-factcheck-accenture-1bc1505e7b38

false

342

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Cool Dung Bettle

My agenda is to make the world a more liveable place.

f984c36e0fd7

BuddhaStings

64

176

20,181,104

null

0

null

0

81068db0cb21

2018-09-01

2018-09-01 09:15:20

2018-09-01

2018-09-01 09:21:54

1

false

en

2018-09-01

2018-09-01 09:21:54

12

1bc284537513

3.573585

0

by Eric Yuan for Forbes

4

There’s Nothing Artificial About How AI Is Changing The Workplace by Eric Yuan for Forbes © Shutterstock/Forbes If you’re having trouble figuring out how artificial intelligence can make a difference in your workplace, you’re not alone. A 2017 Gartner study found that the hype around AI is making it hard for many users to understand the technology’s true value to their organizations. However, if you put the hype aside and drill down into some specific ways in which AI and other new technologies like augmented reality can be used in the workplace, you’ll see that they’re already making a measurable difference in a number of ways and that many workers are excited about those changes. AI, AR Will Make Meetings And Collaboration Better Than Face To Face Video conferencing/collaboration is just one of several workplace arenas in which AI and augmented reality (AR) are making real inroads, and it’s one that holds a lot of promise. In fact, a recent Zoom survey found that 73% of respondents said they expect AI to have a positive impact on meetings, and 67% claimed the same for AR. Here are just a few of the ways AI and AR are improving collaboration and video meetings today — as well as what we can expect in the coming years. The New Meeting Scribe: Artificial Intelligence As I write this, AI has already begun to make video meetings even better. You no longer have to spend time entering codes or clicking buttons to launch a meeting. Instead, with voice-based AI, video conference users can start, join or end a meeting by simply speaking a command (think about how you interact with Alexa). Voice-to-text transcription, another artificial intelligence feature offered by Otter Voice Meeting Notes (from AISense, a Zoom partner), Voicefox and others, can take notes during video meetings, leaving you and your team free to concentrate on what’s being said or shown. AI-based voice-to-text transcription can identify each speaker in the meeting and save you time by letting you skim the transcript, search and analyze it for certain meeting segments or words, then jump to those mentions in the script. Over 65% of respondents from the Zoom survey said they think AI will save them at least one hour a week of busy work, with many claiming it will save them one to five hours a week. AR Redefines Video Conferencing Though its current uses in video meetings are more limited than that of artificial intelligence, augmented reality has the promise to truly redefine and enhance communications in many industries — and it’s getting off to a great start. Companies like Zugara and Meta (a Zoom partner) offer solutions that let you use augmented reality in video conferencing to share and manipulate 3-D virtual holograms in real time and allow others to interact with them as well. If you’re teaching medicine, your students can watch you demonstrate a procedure on an anatomical model that looks realistic. Or your customers can see virtual replicas of products they’re interested in and even try them! I don’t see a strong future right now for virtual reality in video communications. The point of video communications is to engage with someone face to face over long distances. If you’re engaging with their virtual avatar instead of their face, you lose this important benefit. With AR, however, you can share, which is the real point of collaboration. Human Resources See Productivity Gains With AI While it may seem counterintuitive, artificial intelligence can help enhance various human resources functions. In fact, in a 2017 report, the Human Resources Professionals Association found that 84% of respondents believe AI is a useful tool for human resources. Textio, a Seattle-based company, provides software that uses AI to help companies write more effective job postings and recruiting emails. And a conversational AI recruiting assistant named Mya helps improve the hiring process for candidates and recruiters. A large retailer noted that the AI recruiter delivered a 144% increase in recruiter productivity. Improving The Customer Experience Perhaps one of the most common applications of AI and AR in the workplace resides in the sales and customer experience functions. According to a recent survey from Oracle and Coleman Parkes Research on which technologies will most improve the customer experience, 34% said AI will be the biggest game-changer. Chorus.ai is an example of artificial intelligence being used across sales and customer service teams. The software records and analyzes business meetings in real time to create visibility into those meetings and then improves performance by helping replicate how top sellers converse with prospects. Wayfair, the popular e-commerce company that sells home goods, offers an AR-based application that lets prospective buyers see how décor and furniture might look in their home — at full scale — using the camera on their smartphone or tablet. No Longer Just Hype There’s no doubt that some AI and AR applications don’t seem to have practical benefits and exist simply for the sake of entertainment or saying it could be done — like virtual reality roller coaster rides or Sophia, the first AI robot citizen. However, the use of AI and AR in the workplace is delivering real improvements — and those accomplishments are being welcomed by many users. Source: Forbes MOST 2414 is a digital marketing agency and consulting firm.

There’s Nothing Artificial About How AI Is Changing The Workplace

0

theres-nothing-artificial-about-how-ai-is-changing-the-workplace-1bc284537513

2018-09-01

2018-09-01 09:21:55

https://medium.com/s/story/theres-nothing-artificial-about-how-ai-is-changing-the-workplace-1bc284537513

false

894

Digital marketing agency and consulting firm. We help clients to define digital strategies, find new opportunities, and create new businesses with a tailored, creative and problem-solving approach. MOST 2414 is a company based in Bangkok.

null

most2414

null

MOST 2414

info@most2414.com

most-2414

SOCIAL MEDIA MARKETING,DIGITAL MARKETING,ASEAN,DATA VISUALIZATION,INSIGHTS

most2414

Virtual Reality

virtual-reality

Virtual Reality

30,193

Eliseo B.

Partner, MOST 2414 | www.most2414.com

1f34c3bac21b

most2414

6

2

20,181,104

null

0

null

0

null

2017-10-26

2017-10-26 01:02:07

2017-10-26

2017-10-26 01:03:13

1

false

en

2017-10-26

2017-10-26 01:03:13

1

1bc3cf1722a1

1.060377

4

0

This famous statement — the six degrees of separation — claims that there is at most 6 degrees of separation between you and anyone else on…

5

Graph Theory: Six Degrees of Separation Problem This famous statement — the six degrees of separation — claims that there is at most 6 degrees of separation between you and anyone else on Earth. Here we feature a simple algorithm that simulates how we are connected, and indeed confirms the claim. We also explain how it applies to web crawlers: Any web page is connected to any other web page by a path of 6 links at most. The algorithm below is rudimentary and can be used for simulation purposes by any programmer: It does not even use tree or graph structures. Applied to a population of 2,000,000 people, each having 20 friends, we show that there is a path involving 6 levels or intermediaries between you and anyone else. Note that the shortest path typically involves fewer levels, as some people have far more than 20 connections. Starting with you, at level one, you have twenty friends or connections. These connections in turn have 20 friends, so at level two, you are connected to 400 people. At level three, you are connected to 7,985 people, which is a little less than 20 x 400, since some level-3 connections were already level-2 or level-1. And so on. To read the full article with simulator, application to web crawling, some maths, and source code, click here.

Graph Theory: Six Degrees of Separation Problem

43

graph-theory-six-degrees-of-separation-problem-1bc3cf1722a1

2018-02-23

2018-02-23 19:49:07

https://medium.com/s/story/graph-theory-six-degrees-of-separation-problem-1bc3cf1722a1

false

228

null

Algorithms

algorithms

Algorithms

7,319

Vincent Granville

Data science pioneer, founder, entrepreneur, inventor, author, CEO, investor, with broad spectrum of domain expertise.

60b579a69a7a

analyticbridge

265

202

20,181,104

null

0

null

0

null

2017-10-26

2017-10-26 11:04:37

2017-10-26

2017-10-26 11:13:39

3

false

en

2017-10-26

2017-10-26 11:14:40

0

1bc62691c6f3

0.817925

4

0

Morning coffee, hot shower spray tingling revival for a living drone

5

Morning Coffee in AI Heaven Morning coffee, hot shower spray tingling revival for a living drone Get dressed ingest medicine a needle jab booting up telly switch email check scan news feed timeline Absent life home alone another pulse heartbeat desire Words, read and written companions in the void comfort scent body heat memory’s touch in solitude In the waking dark of summer dawn the missing murmur of a soul. asleep

Morning Coffee in AI Heaven

18

morning-coffee-in-ai-heaven-1bc62691c6f3

2017-12-17

2017-12-17 05:58:32

https://medium.com/s/story/morning-coffee-in-ai-heaven-1bc62691c6f3

false

71

null

Poetry

poetry

Poetry

217,749

Dermott Hayes

Novellist, poet, blogger and ex-journalist. ‘If the cap fits.’ https://medium.com/@dermotthayes

8b57be67a6a2

dermotthayes

629

781

20,181,104

null

0

null

0

359fa7b54533

2017-11-07

2017-11-07 09:59:34

2017-11-07

2017-11-07 10:30:19

1

false

en

2017-11-07

2017-11-07 12:55:00

8

1bc64c602295

5.109434

5

0

Define-by-run frameworks are relatively new (2015 for Chainer) but they begin to be well known. I’m writing this post, after some questions…

5

A sneak peek into Dynamic Neural Networks Define-by-run frameworks are relatively new (2015 for Chainer) but they begin to be well known. I’m writing this post, after some questions were asked on GitHub about new imperative frameworks. With the release of Gluon (Mxnet API for dynamic neural network) a month ago, we see people asking why another deep learning API, why don’t they end their Keras v2 compliance etc… I realize not everyone knows the difference between define-by-run frameworks and define-and-run frameworks, especially new users. With the release of Gluon API (Mxnet) a month ago and the pre-release of Eager API (Tensorflow), I think it’s time for a clarification. I will not explain define-by-run framework designs or how they work. This is practical guide on what are pros of these networks. Disclaimer, that will not be a fair comparison, I will just speak about the pros. At the beginning of my journey in deep learning (years ago), and like many new users, I didn’t see the advantages of these frameworks and I sometimes did not even understand them. Just a note in this post I differenciate Mxnet (the symbolic Mxnet API) and Gluon (the imperative Mxnet API). So, we will first begin with an exhaustive list of deep learning frameworks, define-by-run frameworks (Chainer, Pytorch, Gluon, DyNet…) and Define-and-run frameworks (Tensorflow, Keras, Mxnet, Theano (rip)…) From my point of view, there are four points where define-by-run frameworks are better than define-and-run. An easier way to debug neural network architecture, an easy way to understand how to define a model, a better flexibility and, in my opinion, more expressiveness. These points don’t always come from imperative vs symbolic, some of them are more API design. The three dynamic frameworks I tested (Pytorch, Chainer and Gluon) had those advantages. Some of these points don’t come from imperative particularity but there are de facto for them. For example, the Callable network is possible in Keras, it’s one feature that allows an easy network creation. Debugging The first point push forward by Define-by-run frameworks is debugging. It’s easier to debug neural network with imperative logic than symbolic one. But to be honest, as a new user, I don’t correctly perceive this aspect. If you compare debugging on small network like LeNet and only for “hello world” program class like MNIST, you will not see the advantage of debugging. This aspect, in my opinion, only appears with complex structure or when the issue was not straight forward. For example, if your network doesn’t converge it’s easier to inspect gradient or the output dimension. But these are not issues when using small or already working networks. As an end user, these issues never appeared as long as you don’t try to adjust layer structure, change hidden layer, the number of feature etc. … With dynamic neural networks you can explore your network step by step and monitor what happens in each layer with your regular debugger. There is no difference between network debugging and debugging the rest of your program. Model definition Now almost all frameworks have user-friendly network definition. The python API helps a lot with this point. But with the difference that define-by-run network can infer channel dimension of your previous layer. Here is an example with the Gluon API. We define a simple LeNet Network. In this definition, the number of output features don’t need to be define by the user. For example, in Keras you need to define input shape and the layer will be estimated from this. On this example, the number of features will be defined at forward computation. This allows to easily define network with less parameters. Another thing I find awesome is network combination. Callable network is one the features that allows an easy network creation. This feature is already implemented in Keras. For define-by-run framework it is a de facto feature. After this definition and some initialization, you can simply call them like that: It’s easy to do Siamese network, network in concurrence and more complex structures. Flexibility Define-and-run frameworks use an immutable network. As the computation graph is statically defined, the Control-flow needs to be defined as part of this graph. If you want to introduce complex structures like recursion conditions or even loops, it must be done in an indirect manner. This method is different than the standard control flow from regular imperative languages. In define-by-run you can use regular python condition, loop, control flow those you already know. Symbolic frameworks introduce a compilation step between definition and execution. This compilation can obfuscate what happens in your network and add complexity in network construction. For example, if you try to define a shared layer, like in a Siamese network. You need to think of where to store your weight, how to initialized them, how to update them. Otherwise, your final network will not be what you wanted. Define-by-run doesn’t have this additional step, so network behavior is more like regular python class/function. Expressiveness This point is more an API design than an advantage of imperative frameworks. But as most of them (PyTorch, Gluon) take inspiration of Chainer design, the three of them show the same level of abstraction. An example on the Mxnet high level API. The training part in Mxnet is defined and executed by the fit function. Here is an example, the full code can be found here The fit function is easy to use and does lots of things. In one line you compile the network, initialize it, define the optimizer, evaluation metric etc… But if you want to change things inside iterations, monitor the gradient or something else you need to redefine callback, metric etc. for that you need. This can be tricky and you need to know exactly how these callbacks work. That demands lot of work, for some time just monitoring or debugging. On the other side Gluon API is maybe more complex at first glance but brings more expressivity. Full code can be found here Conclusion In my opinion, the hype on define-by-run frameworks comes from a good balance between a good API design and a better flexibility. A good API allows to easily understand what happens, even if you don’t know the framework. I would say that Keras has a good API for example. If you take one of the easiest API in symbolic frameworks, we can describe it as “easy to learn, hard to master”. Whereas, Define-by-run frameworks are a bit harder to learn but they allow complex network creation with serenity. This good API design is combined with a better control on what happen in training loop and what happens in the network. So, yes, imperative networks deserve the hype they’re getting. The advantages from this kind of frameworks are not visible at the first glance. The additionnal flexibility brought by Chainer, Pytorch or Gluon allow to design complex networks easily, with better debugging functionnality and with more control. Dynamic frameworks: - can be used as drop-in replacement for Numpy - are fast for prototyping - are easy to debug and use conditional flows Define-by-run frameworks are fast-growing. And I think now is a good time to try them out! And cherry on the cake, Chainer, PyTorch and Gluon have similar API (inspired by Chainer). So, it’s simple to understand or test others if you already known one. If you want to compare these frameworks, and much more, you can check this repository (not mine), Ilia Karmanov listed simple example (CNN on CIFAR and RNN on IMDB) with lots DL frameworks. Find us on: Facebook and Twitter LinkedIn GitHub Our WebSite

A sneak peek into Dynamic Neural Networks

22

a-sneak-peek-into-dynamic-neural-networks-1bc64c602295

2018-05-09

2018-05-09 07:41:15

https://medium.com/s/story/a-sneak-peek-into-dynamic-neural-networks-1bc64c602295

false

1,301

Wassa is an innovative digital agency expert in Indoor Location and Computer Vision. Within its offices in Paris and Hong Kong, Wassa regroups all the creative, technical and functional required talents to elaborate digital solutions with great added values.

null

wassabemobile

null

Wassa

contact@wassa.fr

wassa

MOBILE,WEB,COMPUTER VISION,CORPORATE,INNOVATION LAB

wassabemobile

Machine Learning

machine-learning

Machine Learning

51,320

Wassa Team

Wassa is a company specialized in the design of innovative digital solutions with great added value.

fd895c58c10b

Wassa

56

28

20,181,104

null

0

null

0

dc36b5ddf133

2017-11-18

2017-11-18 22:46:32

2017-11-18

2017-11-18 22:48:52

1

false

en

2018-07-12

2018-07-12 17:36:05

14

1bc66d9572bd

2.716981

3

0

As a VC, I’ve been researching and tracking technologies broadly within the “Future of Work” space. Ask any VC and you’ll get different…

5

Future of Work #1: Human-Centered AI Photo by Christopher Burns on Unsplash As a VC, I’ve been researching and tracking technologies broadly within the “Future of Work” space. Ask any VC and you’ll get different responses as to what Future of Work covers — be it collaboration, productivity, AR/VR, RPA, and more. So I thought I’d share some of our early learnings at Work-Bench, which are ever-evolving — we’d love to hear from you, and to continue to learn. As an enterprise-focused VC fund, we invest in startups selling into large enterprise customers, and we engage regularly with Fortune 500 line of business buyers. I was recently invited to join a great panel hosted by New America NYC, on the theme of Future of Work, featuring Rick Wartzman’s new book: The End of Loyalty: The Rise and Fall of Good Jobs in America. We covered a wide swath of topics that surrounds this complex issue: the decline of unions, depressed wages, education, healthcare, the changing role of employers’ responsibilities. I talked about the unique confluence of my own personal perspectives — as VC investing in technologies that are disrupting and transforming work; as a long-time GED educator, teaching at the 1199 SEIU Labor Union; advising LaGuardia Community College on student and workforce development initiatives; and even formerly having worked at Cisco Systems within their Learning and Development (corporate training) team. Where I spend a lot of time thinking about at Work-Bench is this theme of human-centered AI. Amidst the doom and gloom that AI and machines will take away our jobs…there will be some period of time when data platforms can help enable and augment humans: to be more efficient and effective. To do our jobs better, faster, and happier. And/or to free us up to do work that is more complex, creative, compelling. How can we ensure AI is built with a human-centered and empathic approach? Where we are not trying to remove or automate away the human, but rather, use their input and knowledge to better train the AI systems and models; to unlock and enhance our human abilities to make otherwise inaccessible data-driven actions and decisions; and to be developed with the expertise, experiences, and insights of humans in mind? Drew Conway, CEO of Alluvium, articles this idea of empathy in data best: In all data there is humanity . In every bit there are traces of of this humanity: in how a choice is made, or how a system is built. In the physical world the complexities of this humanity are magnified. To manage this complexity requires both deep technical expertise and innovative engineering. It also requires considerable empathy for the human beings behind that data. Those of us who work with data are fond of describing it as messy, but data from the physical world is more than simply messy. It is knotted up in the perpetually flawed mechanism used to convert analog actions to digital signal, and the humanity that underlies it. The complexity of this humanity, however, is also our greatest strength and opportunity. The expertise, experience, and bias that people imprint on the data provide material for building great products. Some examples of startups helping to remove really painful, repetitive, and manual work: Upskill — smart glass wearable solution that helps line technicians better assemble complex manufacturing parts through their hands-free, voice command viewfinder (and literally removes painful repetitive motion: https://upskill.io/landing/upskill-and-boeing/) x.ai — virtual assistant who helps remove the pain of scheduling (and rescheduling) meetings Some great examples of software augmenting humans to do their jobs better: Alluvium — helps enable industrial operators by providing more data and insights around production stability Merlon Intelligence — helps compliance analysts more effectively monitor transactions for money laundering by surfacing suspicious activity Do you know other companies in this space using a human-centered approach to building AI, data analytics, and automation? I’d love to chat. *All of the above companies are Work-Bench portfolio companies.

Future of Work #1: Human-Centered AI

8

future-of-work-1-human-centered-ai-1bc66d9572bd

2018-07-12

2018-07-12 17:36:06

https://medium.com/s/story/future-of-work-1-human-centered-ai-1bc66d9572bd

false

667

Work-Bench is an enterprise technology VC fund in NYC. We support early go-to-market enterprise startups with community, workspace, and corporate engagement.

null

work.bench.vc

null

Work-Bench

hello@work-bench.com

work-bench

STARTUPS,ENTERPRISE TECHNOLOGY,ENTERPRISE SOFTWARE,VENTURE CAPITAL,TECHNOLOGY

work_bench

Startup

startup

Startup

331,914

Jessica Lin

co-founder & VC @Work_Bench | GED educator | rethinking work

c876a2575234

jerseejess

708

914

20,181,104

null

0

null

0

null

2017-09-24

2017-09-24 17:43:19

2017-09-24

2017-09-24 17:44:10

0

false

en

2017-09-24

2017-09-24 17:44:10

1

1bc80f6e0f65

0.34717

0

“China, Russia, soon all countries w strong computer science. Competition for AI superiority at national level most likely cause of WW3…

4

You Think He’s Kidding. I Don’t. “China, Russia, soon all countries w strong computer science. Competition for AI superiority at national level most likely cause of WW3 imo.” — Elon Musk, today. You think he’s kidding. I don’t at all, and I’m NOT a Elon fanboy, just a sober, serious, smart thinker who thinks he’s a colossal fucking genius. He previously has been CEO or founder at Zip2, Paypal, Tesla, SpaceX, OpenAI, Neuralink. I think you better actually be concerned. http://bigthink.com/paul-ratner/putin-weighs-in-on-artificial-intelligence-and-elon-musk-is-alarmed

You Think He’s Kidding. I Don’t.

0

you-think-hes-kidding-i-don-t-1bc80f6e0f65

2017-09-24

2017-09-24 17:44:11

https://medium.com/s/story/you-think-hes-kidding-i-don-t-1bc80f6e0f65

false

92

null

Elon Musk

elon-musk

Elon Musk

4,393

Peter Marshall

I am extremely interested in AI, especially the not-so-good side of AI weapons and AI war, although the good parts are magnificent and wonderful too, naturally.

f6bab8ee3d29

ideasware

1,765

276

20,181,104

null

0

null

0

null

2018-02-24

2018-02-24 01:17:22

2018-03-02

2018-03-02 04:42:11

6

false

en

2018-03-02

2018-03-02 19:35:37

37

1bc857f3f53e

7.161321

1

0

Part 1 of my benefits of virtualizing big data workloads with VMware technologies series. This part covers a quick intro to VMware, their…

5

5Vs of Big Data: Volume, Velocity, Variety, Virtualization & VMware Pt. 1 Part 1 of my benefits of virtualizing big data workloads with VMware technologies series. This part covers a quick intro to VMware, their upcoming initiatives and some reasons to consider running big data on vmware. Disclaimer: VMware hasn’t paid me or asked me to publish this article and all thoughts and opinions are my own. I did give VMware early access to the writing to ensure I didn’t disclose any of their company’s intellectual property, shown to me at a special event under NDA, they don’t deem ready for public consumption at the moment. Introduction Last month I was invited out to VMware for their “Experts Workshop: Big Data, Scientific and Engineering Workloads” event that completed last week. As you can imagine from the title, we discussed VMware’s big data related technologies and I was able to learn from other vendors across the big data landscape. There were speakers from premier big data companies like the Tom McCuch, VP of Solutions Engineering at Hortonworks and the “Spark Godfather” Matei Zaharia, Co-founder and Chief Technologist of DataBricks. Other attendees had networking, consulting or software backgrounds in areas of science and High Performance Computing (HPC). VMware also brought out the heavy hitters like their VP of Server Platform Technologies Richard Brunner, Sr. Director & Chief Technologist for HPC Josh Simons and even Ray O’Farrell the Executive Vice President & CTO at VMware. I’m assuming O’Farrell heard I was going to be there so he changed his busy schedule around to make sure he could meet with me to present VMware’s future business initiatives, or at least that is what I’m telling everyone. In his presentation, O’Farrell exposed VMware’s future roadmap covering their cross-cloud (private, hybrid, public) strategy and tools they have to enable their customer’s infrastructure management requirements for emerging technology verticals like IoT, Core (Edge) Computing and 5G networking. Figure 1: VMware’s vision of virtualization and management of compute across any device, application or cloud I’ve noticed more and more companies verbally trying to re-position themselves as “data companies” where VMware, by way of their parent company Dell Technologies, has pledged to make a $1 billion investment in R&D for IoT to go after what they call the “Tidal Wave of IoT Opportunity.” From Figure 1 we see VMware’s plan is to provide virtualization technologies from the data center to the edge, allowing customers to easily manage servers or devices and deploy applications where needed in a secure fashion. Why are we talking IoT on a big data article? Well IoT and edge computing efforts matter for big data as both will provide us with more data than ever before while pushing the computation and analysis closer to the device, meaning we will need to re-examine existing big data technologies along with architectural and access patterns. Lets get into how VMware sees their overall business initiatives manifesting when it comes to virtualization of big data workloads but first I’ll define what I mean by big data as the term has become a bit ambiguous. Defining Big Data Before getting into the VMware’s value proposition of running big data workloads on their technologies, lets first quickly level set on what big data is. The quickest and best description of big data I use, from Eddie Satterly, is “data that doesn’t fit into traditional data models.” I like this description because when you focus solely on speed or the size of data you results will vary from one use case to another. A “big data” application to one company that is getting GBs a day of data may be “small data” to a company bringing in 605 TB every hour. As the traditional data wrangling technologies couldn’t handle the new use cases created from the large data generation, a new solution, “big data” was born. Since then, along with the explosion of data, we’ve seen an explosion in new technologies trying to wrangle the explosion of data. Figure 2: Just a few of the big data tools available Most big data tools can be broadly categorized as tools that help you collect, store, process or analyze data. Some tools span multiple categories and others, like machine learning’s analysis of data, create a new subcategory but overall modern big data tools fit in one of these four categories (at least they do for the purpose of this article). If you need a deeper dive on “What is Big Data?” from a business or technical sense then follow the links. Now that we’ve level set on a definition of big data, lets discuss how VMware wants to help you with those workloads. Why Big Data on VMware? How VMware Sees the World of Big Data From O’Farrell’s presentation, it is clear that VMware is tracking the onslaught of new data being generated, by people or devices, and looking to arm their enterprise customers with the necessary tools to navigate the big data jungle. From VMware’s dedicated big data site they state, “VMware is the best platform for big data as well as traditional applications. Virtualizing big data applications simplifies the management of your big data infrastructure, delivers faster time to results and is more cost effective.” They believe their offering’s ability to simplify the management, deliver faster results and reduce cost allows you to jump over big data’s biggest hurdles which are a lack of IT expertise or available enterprise grade tools and budget constraints. I’ll now take a deeper dive into what these benefits look like in execution. Virtualized Architecture and Managing Big Data Workloads Virtualization of Hadoop is a pretty common, well known use case by now. I feel comfortable saying the majority of Hadoop workloads run in some type of virtualized environment, rather than running on bare metal. The value propositions of virtualization (like ease of deployment, flexibility to scale as you need, ability to separate compute and storage, etc.) are well known and mostly achieved with any virtualized hardware. Rather than look at why virtualization is good for Hadoop in general, I tried to focus on how VMware differentiates itself from other virtualization technologies. During the workshop’s presentations, I couldn’t discern VMware’s major differentiation to virtualization and management of big data technologies like Hadoop or Spark. This is not to say that there are none as you can look at the other leader in Gartner’s Magic Quadrant for x86 Server Virtualization, Microsoft’s Hyper-V, and find differences in supported operating systems, pricing, ease of deployment, public cloud support, supported hardware and more, but when looking specifically at differentiation for big data workloads nothing immediately jumped out at me. A lack of a immediately noticeable differentiation over competitors isn’t necessarily a bad thing as VMware customers don’t have to look elsewhere to satisfy their big data needs as the platform can help in realizing all of the virtualization benefits mentioned earlier. To realize these benefits, they’ve provided a Best Practices for Virtualized Big Data Applications guide to help with hardware selection, VM sizing, application tuning and testing of your setup through benchmark analysis. Now on to accelerators. Accelerating Big Data Workloads All you have to do is take a look at NVIDIA’s stock performance over the past couple of years to see that accelerators like GPUs are really taking off across multiple compute use cases, but primarily artificial intelligence. FPGAs are finding their ground also, in use cases like SmartNICs and Machine Learning (ML), due to their high efficiency and low power consumption. Figure 5: CPU vs GPU Performance on vSphere As ML is utilized more and more in the analysis phase of big data, we are seeing a rise in GPU consumption as GPUs have significantly more cores (1,000 core on average) than CPUs (12 cores on average) which make them ideal for the redundant matrix multiplication required in Deep Learning (DL), a subfield of ML. When adding NVIDIA GPUs to VMware vSphere environment for DL applications, you can use one of two modes: DirectPath I/O Passthrough or GRID vGPU. Figure 6: Direct Path I/O Passthrough (left) and Grid vGPU (right) GRID vGPU is recommended when you need to pair 1 VM to 1 GPU and for applications that require short training times and use multiple GPUs to speed up machine learning tasks, DirectPath I/O is the better option. For more details on the performance differences check this article out. And note that I’ve used NVIDIA GPUs as an example but other accelerators, like Intel’s Xeon Phi or AMD’s FirePro S7150x2 can also be utilized by vSphere in a similar manner. On to the important question, what is the virtualization “hit” on ML workloads? Figure 7: A 4% virtualization hit on performance for deep learning workloads For language modeling with a Recurrent Neural Network (RNN) on the Penn Treebank (PTB) data set, there was only a 4% performance hit due to virtualization. For most ML workloads, the 4% hit is minor when comparing it with the benefits you get from virtualization. Looking at other GPU benchmarks across machine learning and graphics rendering workloads, the virtualization hit is anywhere from 1% to 4% so the 4% for this specific language modeling use case is within an optimal performance range. For more details and examples of ML on vSphere with accelerators check out VMware’s presentation given at the 2017 GPU Technology Conference. Conclusion VMware’s technology stack shows a lot of maturity when using it for big data and machine learning workloads. In the next section we’ll explore more of the answers provided to me when asking why should a customer run their big data workloads on VMware. Part 2 will cover networking capabilities, persistent memory and benchmarks of big data workloads. I’ll end Part 2 by answering the question of whether you should move your big data workloads from another solution to VMware or, for existing customers, from VMware to an alternative. Continue on to read Part 2 of 5Vs of Big Data: Volume, Velocity, Variety, Virtualization & VMware. If you enjoyed this article, please tap the claps 👏 button. Interested in learning more about Jamal Robinson or want to work together? Reach out to him on Twitter or through LinkedIn.

5Vs of Big Data: Volume, Velocity, Variety, Virtualization & VMware Pt. 1

1

5vs-of-big-data-volume-velocity-variety-virtualization-vmware-pt-1-1bc857f3f53e

2018-06-12

2018-06-12 15:52:27

https://medium.com/s/story/5vs-of-big-data-volume-velocity-variety-virtualization-vmware-pt-1-1bc857f3f53e

false

1,646

null

Cloud Computing

cloud-computing

Cloud Computing

22,811

Jamal Robinson

Enterprise technologist with experience across cloud, big-data, AI and other cool technologies. Recently published author of “Rich People Making Poor Decisions”

bdedf45adfcb

jamal.robinson

21

1

20,181,104

null

0

null

0

1c18e80de565

2018-07-15

2018-07-15 11:11:42

2018-07-18

2018-07-18 18:19:36

5

false

en

2018-07-18

2018-07-18 18:19:36

0

1bc85c0a3900

4.584277

0

More on uninformed search strategies, aimed mainly to solve the infinite-depth problem of DFS

3

Search Algorithms Part 3: Uninformed Search Algorithms — 2 In the previous blog, we have discussed the three popular uninformed search strategies: BFS, uniform-cost search and DFS, along with their advantages and disadvantages. In this blog, we will discuss three more uninformed search algorithms, where two of them is intended to solve the infinity-depth problem of DFS, and the third one is a small improvisation on the idea of searching from the source to the destination. Note: In the previous blog, you might have noticed this symbol occurring in some of the expressions: ^. This symbol refers to the mathematical expression “raised to the power of”. a^b means ‘a raised to the power of b’. I will be using this symbol with the same meaning throughout the series of blogs. Depth-limited Search To solve the problem of DFS getting stuck in a search tree with infinite depth, depth-limited search restricts the depth of the tree to a predetermined depth limit l. All the nodes at depth d is treated as if they have no successors. Figure 1: Pseudo-code of the depth-limited search. Depth-limited search solves the infinite-path problem. But the search is not complete if l < d. Even if l > d, optimal solution is not guaranteed, as we could be eliminating some of the solutions at depths > l. Time complexity is O(b^l), and space complexity is O(bm) (It is same as DFS, only with restricted depth to l). In fact, DFS can be viewed as a special-case of depth-limited search with l →infinity. The problem with depth-limited search is to set the value of l optimally, so as to not leave out any solution, as well as keep the time and space complexity to a minimum. For example, in the problem of traveling from Arad to Bucharest, we can see that there are only 20 cities in the map. Hence, we can set the value of l to 19 (remember that we have set the root value to 0). Figure 2: A part of the road map of Romania (revisited). If we observe the map carefully, we can see that we can reach any city (state) from any other city (state) in at most 9 states. Using this information, called as the diameter of the state space, we can further reduce the l value to 9. For most of the problems, we cannot estimate the depth limit until we have solved the problem. Iterative Deepening Depth-first Search (IDS) Iterative deepening search (or iterative-deepening depth-first search) offers a solution for the problem of finding the best depth limit. It gradually increases the depth — first 0, then 1, then 2, and so on — until a goal is found. It combines the advantages of both BFS and DFS. Like DFS, it consumes less memory: O(bd). Like BFS, it is complete when b is finite, and is optimal when the path cost is a non-decreasing function of depth. Figure 3: Iterative-deepening search, repeatedly calling the depth-limited search while varying the depth limit in every iteration. Iterative deepening search generates the states multiple times, but it is not too costly. In a search tree with nearly the same branching factor at every level, the size of the bottom levels are very huge compared to the top ones, hence it does not affect much if the nodes in the top levels are generated many times. To generate the node at depth d, the nodes at depth d-1 are generated twice, the nodes at depth d-2 are created 3 times, and so on. Hence, the total number of nodes generated will be: which has the complexity of O(b^d), the same as BFS. For example, for d = 5 and b = 10: n(IDS) = 50 + 400 + 3000 + 20000 + 100000 = 1,23,450 n(BFS) = 10 + 100 + 1000 + 10000 + 100000 = 1,11,110 In terms of performance measure, since IDS seems to be the hybrid form of BFS and DFS, IDS is the preferred uninformed search method when the search space is large and the depth of the solution is not known. Bidirectional Search The idea behind the bidirectional search is to run two searches simultaneously — one forward from the start state, and the other from the goal — hoping that the two will meet somewhere in the middle. It is implemented by replacing the goal test with a check to see whether the frontiers of the two search intersect (the solution is found if they do). The check can be performed when the node is generated, and can be done in constant time using hash tables. The main aim of bidirectional search is to reduce the total search time. If we use BFS at both the ends as the search algorithm, the time and space complexity will be O(b^(d/2))(In the worst case, the two frontiers meet in the middle). The space can be reduced if one of them is IDS, but one of the frontiers must be in the memory to perform the check for the intersection. The difficulty in using the bidirectional search is to search backwards. We need to find the predecessors of each state (all the states having the given state as their successor). If the actions of the search space are reversible, like in the Arad-Bucharest path-finding problem, the predecessors of the node are its successors. If there are several goal states, then we can create a dummy goal state whose predecessors are the actual goal states. Summary This blog post was the continuation of the discussion on uninformed search algorithms. We have discussed couple of algorithms to handle the main problem faced in DFS — the infinite-depth situation. We came across depth-limited search, where we remove all the nodes beyond a certain limit, and iterative deepening search, where we incremented the limit of the depth until we find the solution. We finally discussed the bidirectional search, where we can perform two searches simultaneously to reduce the search time. Figure 4: Evaluation of uninformed search strategies. In the next blog, we will discuss on informed (heuristic) search algorithms.

Search Algorithms Part 3: Uninformed Search Algorithms - 2

0

search-algorithms-part-3-uninformed-search-algorithms-2-1bc85c0a3900

2018-07-19

2018-07-19 07:26:04

https://medium.com/s/story/search-algorithms-part-3-uninformed-search-algorithms-2-1bc85c0a3900

false

994

Research, solution offering and training on AI, Deep learning, Robotics, DeepRL, Devops and Blockchain. Blogs emerging out of research, innovation and engineering at Kredo.ai.

null

Kredo.ai Engineering

hackintoshrao@gmail.com

kredo-ai-engineering

DEEP LEARNING,REINFORCEMENT LEARNING,ARTIFICIAL INTELLIGENCE,DEVOPS,MACHINE LEARNING

kredo_ai

Search Algorithm

search-algorithm

Search Algorithm

22

Rithesh K

AI Enthusiast. Loves Mathematics. Writing Short stories and Quotes are my hobbies.

87bd15e3fdc3

rithesh18.k

12

1

20,181,104

null

0

null

0

null

2018-07-18

2018-07-18 11:19:01

2018-07-18

2018-07-18 11:41:39

1

false

en

2018-07-20

2018-07-20 07:34:15

1

1bc89ec2f2ce

3.101887

5

0

When I first joined Fenix Intl’s data team, I was blown away by how data-literate the Fenix business was. It wasn’t just the skill of their…

3

Photo by Markus Spiske How does Fenix approach data analysis? When I first joined Fenix Intl’s data team, I was blown away by how data-literate the Fenix business was. It wasn’t just the skill of their existing data team members, but also the way in which the different teams across the business self-serviced data and made their own data driven decisions. Having come from a large organisation where data driven decisions still seemed to be heaped on to a handful of members of a data analytics team, it was refreshing to see teams such as finance, marketing and product development taking it upon themselves to build their own reporting and carry out their own analysis. This is why Fenix is not exaggerating when it calls itself a data driven business. Everything from the design of its latest generation products, to the way it staffs its call centres has been arrived at through the analysis of data. With so much demand on data analysis throughout the business, it’s no surprise that Fenix’s small but growing data team has had an important role to play, which is why I would like to describe Fenix’s approach to data analysis, from my point of view as a new member of their data team. There’s a lot of emphasis put on “data” by the industrial press, often with the tagline that “data is everything” for a business. While this claim certainly has merit, I feel that the value of data isn’t actually realised until it is analysed, insights have been uncovered, and it has been used to tell a story. I believe that this represents a key part of the Fenix data team’s approach to exploring data. The data, and the analysis, are both a means to an end. They’re not the prize at the end of the rainbow, they’re the fuel and the engine that help us get there. This is why we always try to start our data analysis projects by understanding the business problem, or the overarching objective of our stakeholders. After all, they’ve got a goal they need to hit, and putting ourselves in their shoes will do a lot to help guide the analysis. Another key part of our approach that probably seems obvious; we familiarise ourselves with the data, and understand its limitations. Is an increase in a household’s expenditure being caused by a new product launch, or is it really because fuel prices have gone up and people have no choice but to spend more? At the same time, we try to remind ourselves to avoid analysis paralysis. It’s easy to sink two hours into working out whether users responding to an interaction were doing it via a smartphone or feature phone, but ultimately if a stakeholder just needs to know how many positive responses came from the interaction, they won’t even look at the insight. We also try to anticipate our audience’s questions. This comes back to putting ourselves in our stakeholder’s shoes. What would we want to know if we were running their business unit? What would our immediate priorities be? The data and the analytics become less important. It’s the “so what?” that counts. What’s the insight that will help them make that next decision? Finally, we try to keep it simple. This is easier said than done. How can we keep insights simple for someone after spending weeks building and training a machine learning model which looks at hundreds of variables just so that we can forecast out sales? This is the part of data analysis a lot of people find most difficult. Any analyst will want to showcase the complicated, scientific and rigorous process that they undertook to analyse the data given to them. So much so that the audience is likely to get lost in the detail and feel like they never got the insight that they were promised. It’s important to find a balance between communicating the key insight to a stakeholder, whilst at the same time assuring them that we’ve explored every avenue or possible pitfall in the data. We won’t claim that this is always easy. For most people (including myself) it’s likely going to be a trial and error process where you learn from doing. I hope this insight into Fenix Data’s approach to data analysis has been helpful to other practitioners. Next time you’re embarking on some data analysis, just remind yourself of why you’re doing it, and who you’re doing it for. Written by: Alex Roussel, who is on the data team at Fenix. Alex likes to salvage old bikes and restore them in his free time

How does Fenix approach data analysis?

5

how-does-fenix-approach-data-analysis-1bc89ec2f2ce

2018-07-20

2018-07-20 07:34:15

https://medium.com/s/story/how-does-fenix-approach-data-analysis-1bc89ec2f2ce

false

769

null

Data Science

data-science

Data Science

33,617

Fenix Engineering

We use Systems Engineering and Data Science to give millions access to clean energy and financial loans.

d411887a988a

fenixintl_Tech

12

1

20,181,104

null

0

null

0

null

2018-08-31

2018-08-31 16:31:03

2018-08-31

2018-08-31 17:49:27

2

false

en

2018-09-01

2018-09-01 01:24:55

3

1bcb6d4d0497

2.330503

1

0

Everything that you need to know!

5

DATA Everything that you need to know! Data is nothing but a Piece of Information. If I want to know the average sales of a firm for the last 6 months, I need to collect some data, and by looking at those data I can spot the trend of average sales. Classification: There are mainly two types of data: Quantitative: When data is expressed in numerical terms, eg: Height, Weight. Quantitative data is also called “Variable”, as the value can change. Qualitative: When the data cannot be expressed in numerical terms, example: Religion, caste. Qualitative data is also called “Attribute”. Classification of Data Types Discrete Data: Suppose, if I ask you” How many members are there in your family?” Your answer will be 2 or 3 or 5 or 10 etc. These are all isolated values. This type of data is known as “Discrete Data”. Continuous Data: Now if I want to know about your height, you might answer me in the following wayà5.8888 feet or 6.2222feet. So your height can take any value within a specific range i.e. between 5 & 6 or 6 & 7. This is an example of “Continuous Data”. Discrete and Continuous both comes under the category of “Variable”. Nominal Data There are some qualities or Attributes which cannot be compared or cannot be rank ordered. For example color of eyes, color of hair, religion cannot be ranked. As an individual, you might have your own preference but logically you cannot say that which religion is the best. We call this kind of Quality as “Nominal”. Ordinal Data: When the quality can be ranked. Eg. If you are asked to give a feedback of this post by giving stars where 1star means “Very Bad”, 2 stars means “Bad”, 3 stars means “Good”, 4 means “Very Good” and 5 means “Excellent”. So here you are ranking the quality of this post. This is what we call Ordinal Data. Now, if you have worked on R, you might know that R considers some data as “Factor”. The factor is nothing but a “Categorical Data”. Categorical Data: The categorical variable represents the types of data which may be divided into groups. Those groups are finite numbers. Gender is an example of a Categorical Variable. If we divide Gender into groups we will get mainly 3 categories: “Male”, “Female” and “Others”. If R is considering any variable as a factor, and you check the structure of that variable, you can see different levels of it. If you want to convert any variable into a factor you can use as. factor () syntax. For further reading click here. Hope you find this useful. We will be back very soon with other interesting topics. Do you share the same enthusiasm for Data Science, ML, Deep Learning and collaborative learning!! Go ahead and fill in your details here and we will add you as a writer on our Medium publication and StepUp Analytics. Happy writing! And of course — don’t forget to spread the word around about our publication!. Scale Up Your Skills with StepUp Analytics. “Keep on Learning, Keep on Practicing”

DATA

43

data-1bcb6d4d0497

2018-09-01

2018-09-01 01:24:55

https://medium.com/s/story/data-1bcb6d4d0497

false

516

null

Data Science

data-science

Data Science

33,617

StepUp Analytics

StepUp Analytics is a Community of Creative, Highly Energetic Data Science and Analytics Professionals and Data Enthusiast.

5b49263f8b72

StepUpAnalytics

9

4

20,181,104

null

0

null

0

null

2018-08-30

2018-08-30 16:37:24

2018-08-30

2018-08-30 16:37:36

0

false

en

2018-08-30

2018-08-30 16:37:36

1

1bcc5eff1af2

2.430189

0

[Download] [PDF] Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference READ ONLINE By Cameron Davidson-Pilon…

1

Pdf Download eBook Free Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference By Cameron Davidson-Pilon PDF Full #Audiobook [Download] [PDF] Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference READ ONLINE By Cameron Davidson-Pilon Link https://shoppipubherenow.icu/?q=Bayesian+Methods+for+Hackers%3A+Probabilistic+Programming+and+Bayesian+Inference Master Bayesian Inference through Practical Examples and Computation-Without Advanced Mathematical Analysis Bayesian methods of inference are deeply natural and extremely powerful. However, most discussions of Bayesian inference rely on intensely complex mathematical analyses and artificial examples, making it inaccessible to anyone without a strong mathematical background. Now, though, Cameron Davidson-Pilon introduces Bayesian inference from a computational perspective, bridging theory to practice-freeing you to get results using computing power. Bayesian Methods for Hackers illuminates Bayesian inference through probabilistic programming with the powerful PyMC language and the closely related Python tools NumPy, SciPy, and Matplotlib. Using this approach, you can reach effective solutions in small increments, without extensive mathematical intervention. Davidson-Pilon begins by introducing the concepts underlying Bayesian inference, comparing it with other techniques and . . . . . . . . . . . . . . Read Online PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Download PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Download Full PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Download PDF and EPUB Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read PDF ePub Mobi Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Reading PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read Book PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read online Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Download Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Cameron Davidson-Pilon pdf, Download Cameron Davidson-Pilon epub Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read pdf Cameron Davidson-Pilon Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Download Cameron Davidson-Pilon ebook Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read pdf Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Online Download Best Book Online Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read Online Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Book, Read Online Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference E-Books, Read Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Online, Read Best Book Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Online, Read Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Books Online Download Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Full Collection, Download Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Book, Read Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Ebook Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference PDF Read online, Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference pdf Download online, Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Read, Download Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Full PDF, Read Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference PDF Online, Read Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Books Online, Read Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Full Popular PDF, PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Read Book PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read online PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Download Best Book Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Read PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Collection, Read PDF Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference Full Online, Read Best Book Online Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference, Download Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference PDF files #ebook #epubs #online #PPT #AudiobookOnline

Pdf Download eBook Free Bayesian Methods for Hackers: Probabilistic Programming and Bayesian…

0

pdf-download-ebook-free-bayesian-methods-for-hackers-probabilistic-programming-and-bayesian-1bcc5eff1af2

2018-08-30

2018-08-30 16:37:37

https://medium.com/s/story/pdf-download-ebook-free-bayesian-methods-for-hackers-probabilistic-programming-and-bayesian-1bcc5eff1af2

false

644

null

Machine Learning

machine-learning

Machine Learning

51,320

Acura Bryant

null

2828cf3187f2

acurabryant

0

1

20,181,104

null

0

null

0

835c7a9dc1d3

2018-06-20

2018-06-20 08:04:40

2018-06-20

2018-06-20 13:50:27

3

false

en

2018-10-31

2018-10-31 08:17:36

4

1bcdd483f12a

3.191509

2

0

by Andrea Bolioli from H-FARM Industry - Artificial Intelligence

5

A quick glossary for AI - pt 1 by Andrea Bolioli from H-FARM Innovation Everybody talks about “Artificial Intelligence” and we use this term as a buzzword in conversations, advertising, scientific conferences, workshops, etc. When was Artificial Intelligence born? On August 31, 1955 John McCarthy (Dartmouth College), Nathaniel Rochester (IBM), Claude Shannon (Bell Telephone Laboratories) and Marvin Minsky (Harvard University) submitted their proposal to organize a summer research project on AI: We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer. Participants of the Dartmouth Summer Research Project on Artificial Intelligence. | Photo credits: Achievement.org More then 70 years have passed, and we may say that the key themes of Artificial Intelligence research are all there, very similar to those envisaged by the founding fathers. There has been substantial progress, but long-term objectives of understanding intelligence and building intelligent machines are still bold and exciting. Sometimes I prefer to use the term “machine intelligence” instead of “Artificial Intelligence”, as the assumption that the process of human thought can be mechanized has a very long history. Maybe I will write about this in a next post. It’s impossible to cover here the full spectrum of topics related to AI (theory, algorithms, applications, software and hardware infrastructure), but I’ll try to present a short glossary of basic terms used in this discipline. I chose the definitions given by the founders of this discipline, when they are still valid, or, in other cases, the most recent definitions that are broadly shared today. Artificial Intelligence A discipline whose objective is to understand and reproduce human cognition, creating cognitive processes comparable to those found in human beings. (Villani 2018) This definition is similar to the one that scientist McCarthy gave in 1955: Making a machine behave in ways that would be called intelligent if a human were so behaving. Text Mining and Analytics The discovery by computer of new, previously unknown information, by automatically extracting information from different written resources. (Hearst 2003) Automatic Speech Recognition The automatic speech recognition (ASR) component processes the acoustic signal that represents the spoken utterance and outputs a sequence of word hypotheses, thus transforming the speech into text… A primary goal of ASR research has been to create systems that can recognize spoken input from any speaker with a high degree of accuracy. (McTear 2016). Information Extraction Any method for filtering information from large volumes of text. (Grishman 1997) Chatbot or dialog systems or conversational interfaces Chatbots are systems that can carry on extended conversations with the goal of mimicking the unstructured conversational or ‘chats’ characteristic of human-human interaction. (Jurafsky 2017) Conversational interfaces enable people to interact with smart devices using conversational spoken language. (McTear et al. 2016) Sentiment analysis and opinion mining Automatic identification and extraction of opinions, emotions, and sentiments in text. (Wiebe et al. 2005) References: Ralph Grishman (1997), “Information extraction: techniques and challenges”, www.ru.is/faculty/hrafn/Papers/grishman97information.pdf Marti Hearst (2003), “What Is Text Mining?”, people.ischool.berkeley.edu/~hearst/text-mining.html Daniel Jurafsky, James H. Martin (2017), “Speech and Language Processing. An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition”, Prentice Hall, Upper Saddle River, New Jersey. John McCarthy et al. (1955), “A proposal for the Dartmouth summer research project on Artificial Intelligence. August 31, 1955.”, in AI magazine 27.4 (2006): 12. Michael McTear, Zoraida Callejas, David Griol Barres (2016), “The Conversational Interface. Talking to Smart Devices”, Springer International Publishing, Basel (Switzerland). Cedric Villani (2018), “For a meaningful artificial intelligence. Towards a French and European strategy”, www.aiforhumanity.fr/pdfs/MissionVillani_Report_ENG-VF.pdf Janyce Wiebe, Theresa Wilson, Claire Cardie (2005), “Annotating expressions of opinions and emotions in language”, in Language resources and evaluation 39.2–3 (2005): 165–210.

A quick glossary for AI - pt 1

3

ai-glossary-1bcdd483f12a

2018-10-31

2018-10-31 08:17:36

https://medium.com/s/story/ai-glossary-1bcdd483f12a

false

700

Stories about H-FARM's division dedicated to the digital transformation of companies.

null

hfarmspa

null

H-FARM Innovation

null

h-farm-industry

BUSINESS TRANSFORMATION,DIGITAL MARKETING,INNOVATION,ARTIFICIAL INTELLIGENCE,STORYTELLING

hfarmspa

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

H-FARM

The innovation platform that supports the creation of new business models and the digital transformation and education of people & companies. http://h-farm.com

bfc9ecc09859

hfarmspa

3,037

218

20,181,104

null

0

import pandas as pd mnist_train = pd.read_csv('input/mnist/train.csv') batch_y = pd.get_dummies(mnist_train.ix[:, 0]).values batch_x = mnist_train.ix[:, 1:mnist_train.shape[1]].values min_max_scaler = preprocessing.MinMaxScaler() batch_x = min_max_scaler.fit_transform(batch_x) x = tf.placeholder(tf.float32,[None,784]) y = tf.placeholder(tf.float32,[None,10]) W1 = tf.Variable(tf.random_normal([784,300],stddev=0.03),name="W1") b1 = tf.Variable(tf.random_normal([300]),name="b1") W2 = tf.Variable(tf.random_normal([300,10],stddev=0.03),name="W2") b2 = tf.Variable(tf.random_normal([10]),name="b2") hidden_out = tf.nn.relu(tf.add(tf.matmul(x,W1),b1)) y_ = tf.nn.softmax(tf.add(tf.matmul(hidden_out,W2),b2)) y_clipped = tf.clip_by_value(y_,1e-10,0.9999999) cross_entropy = -tf.reduce_mean(tf.reduce_sum(y*tf.log(y_clipped)+(1-y) * tf.log(1-y_clipped),axis=1)) optimiser = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cross_entropy) init_op = tf.global_variables_initializer() correct_prediction = tf.equal(tf.argmax(y,1),tf.argmax(y_,1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32)) with tf.Session() as sess: sess.run(init_op) total_batch = total_size//batch_size for epoch in range(epochs): avg_cost = 0 mnist_train = shuffle(mnist_train) batch_y = pd.get_dummies(mnist_train.ix[:, 0]).values batch_x = mnist_train.ix[:, 1:mnist_train.shape[1]].values batch_x = min_max_scaler.fit_transform(batch_x) for i in range(total_batch): _,c = sess.run([optimiser,cross_entropy],feed_dict={x:batch_x[i*batch_size:(i+1)*batch_size],y:batch_y[i*batch_size:(i+1)*batch_size]}) avg_cost += c/total_batch print("Epoch:", (epoch + 1), "cost =", "{:.5f}".format(avg_cost)) test_batch_x = mnist_test.values test_batch_x = min_max_scaler.fit_transform(test_batch_x) results = sess.run(y_clipped,feed_dict={x:test_batch_x}) print("Training Completed")

6

null

2017-10-05

2017-10-05 17:22:47

2017-10-15

2017-10-15 13:51:48

2

false

en

2017-10-15

2017-10-15 13:51:48

4

1bd331015fd8

3.500314

6

0

MNIST is a entry point for anyone who is starting to study machine learning. MNIST is a popular data set used for digit recognition. It is…

4

Solve MNIST using Tensorflow MNIST is a entry point for anyone who is starting to study machine learning. MNIST is a popular data set used for digit recognition. It is a set of 28x28 pixel black and white images of digits. Tensorflow is a very popular tensor processing framework developed by Google and it has very useful functions for machine learning. Tensorflow comes build in with a script to download and process MNIST data set so you can focus only on the learning part of the system. But in doing so you could miss lot of things such as preprocessing of data to suite for learning. So in this article I am going to get raw pixel data and process them and feed them in to learning system and create a model to read digits. You can find pixel data set form here. This training csv files contains 28x28 columns for pixel data and a label column. First step in the process is reading the CSV file, for that we can use pandas framework which comes with lot of functions to read,process and extract data from a CSV file. Here, first we have import pandas framework and then read the csv file. Next step is extracting labels and training pixels separately. batch_y is the labels. The get_dummies function convert the labels to binary format. That is if label is 1 it will converted to [0 0 0 0 0 0 0 0 0 1]. Since this is a classification problem we have to decide given image belongs to which class so converting to this kind of array with 10 elements makes our lives easier specially when we are working with matrices. Next, we have to process the data so that they are suitable for learning. In the dataset we have details about 28x28 pixels. If the pixel is white it’s value is 0 otherwise pixels brightness is decided by an integer which can ranges upto 255. In the data set, black color pixels usually has higher values(closer to 255) so data set has lot of zeros and few (or more) values closer to 255. The data set is not evenly distributed. This can leads to huge errors in the model. So we have to make them evenly distributed. There are several ways for this preprocessing of data: min-max scaling, normalization etc. In the below code we have used min-max scaling. Now it is time to create the learning system. For this problem it is suitable to use either fully connected neural network or a convolutional neural network. For the sake of simplicity I am going to design a fully connected neural network. In tensorflow we can first design the operations we are going to perform and then we can create a session and execute them as needed. In this code snippet, we have first created placeholders for our variables that is for batch_x and batch_y. [None,784] means we will initialize two dimensional array of unknown number of row count. Row count will be decided while executing tensorflow session. Next we have assigned random numbers for the weights and biases of the neural network. Architecture of our neural network hidden_out is the output of the hidden layer and y_ is the output of the final layer. To train the model we have to calculate the cost of predictions with current parameters corss_entropy is for that purpose and optimizer tries to minimize this cost by changing the parameters of each layer. After training the model we can use following operators to meassure the accuracy of the trained model. Next step is executing those operations as needed in a tensorflow session. This segment will call above discussed operations on our data set. The data set will be processed in batches. To improve the accuracy we have done several iterations of training. In each iteration when feeding data into neural network we have shuffled data. That is also done to improve the accuracy. This will complete the training of the data set. When predicting test values the same operations can be used. We have to evaluate y_ with test images as input. You can find the full source code for this program in here. References [1] http://yann.lecun.com/exdb/mnist/ [2]https://en.wikipedia.org/wiki/Convolutional_neural_network

Solve MNIST using Tensorflow

12

solve-mnist-using-tensorflow-1bd331015fd8

2018-05-02

2018-05-02 04:18:06

https://medium.com/s/story/solve-mnist-using-tensorflow-1bd331015fd8

false

826

null

Machine Learning

machine-learning

Machine Learning

51,320

Chamath Abeysinghe

Undergraduate of University of Moratuwa, Department of Computer Science and Engineering.

15fa9f5a626c

abeysinghechamath

14

20,181,104

null

0

null

0

null

2017-09-09

2017-09-09 13:12:13

2017-09-11

2017-09-11 05:50:01

1

false

en

2017-09-14

2017-09-14 02:53:27

1

1bd3ab041c9d

3.856604

0

How we got to blockchain from speech to writing and the printing press

5

The Road to Blockchain (Ver el artículo en español) Blockchain is going to be a transformative technology. It is the next step in information storage and management technologies. We humans have been working at improving how we store and manage information for all our history, and each time a new technology has taken hold we have seen fundamental changes in society and business. Information storage and governance have a subtle but profound impact on our world. It might not be as visible as physical technologies, but its impact is much more pervasive. Information storage and management determines what we remember, talk about and measure, as individuals and as a society, so it determines memory and history. Ancient Egyptians pharaohs changed the written records in the walls of temples and monuments to erase unwanted previous rulers and events. It also determines our organization. As a consequence, it determines government and power structures. So in full, it determines the reality in which we live and operate as humans. Orwell’s great 1984 is a great example of how a new world order is determined mainly by information access and storage. Each new technology has improved how we deal with information in a series of dimensions. There are six main dimensions which need to be considered: Permanence, how well does information endure over time in terms of its overall structure and its specific content. Can we trust it against the tampering of time and other humans? Transmission, how easy and effective is it to replicate and transmit information. Updatability, how can new information be incorporated and learning occur. Processing, how can the information be used to come up with new knowledge or used to cause effects in the physical world. Cost, how many resources does information storage and management require. Speed, how fast can transmission, updates, and processing occur. The oldest information processing structure is our genome. It is quite permanent, has mastered transmission, allows some updatability and it is processed into living beings that embody it. It is a wonderful achievement that allowed evolution and life. However, its speed is very slow, in the millions of years, and it is not versatile, it only deals with living beings. Then came neural nets and the isolated brain. It was a substantial improvement in terms of speed and versatility of updating and processing. This advantage allowed brain-endowed creatures to gradually take over the biosphere at a macroscopic level. However, its transmission is very limited and it is very costly in biological terms. Speech is evolution’s solution to the transmission problem in a very cost effective way. It is probably what allowed humans to expand all over the world. Its limitations are in permanence and processing, which are sorely lacking, as “words are taken by the wind”. Magic and priestly rituals fantasize with what speech could do with permanence and processing. Writing issued us into history and world domination. It was speech with crude permanence and updatability. The different writing technologies (vellum, stone, clay, papyrus, paper…) made varying tradeoffs between permanence, updatability and cost. Writing allowed us to create complex organizational structures, such as early cities and empires, that integrate brains, speech and writing for processing. The printing press was the forerunner of the industrial revolution. It improved transmission cost and speed drastically and enhanced permanence and updatability. For the first time, it allowed information to be transmitted effectively to large masses of people, integrating brains and speech closer and allowing even more complex organizations, like nation states and corporations. The next step was digital information storage through computers. It improves cost and updatability, at the expense of some permanence. It also gave the first improvement to processing since the brain. As an example, its original use case was projectile trajectory calculations for the war and other mathematical endeavors, pure processing. Programming is the first example of processing being done independently from human beings, starting to fulfill the promise of magic. The last step before the blockchain are databases and the internet. Put together they have allowed a step forward in transmission equivalent to the printing press, and they also improved learning and processing dramatically. Cost continued to decline based on the improvement of digital information storage. However, it has created a crisis of permanence. With information being so ubiquitous and easy to store, duplicate and manipulate: What can you trust? What is the last version? How can you keep privacy? How can you make sure several parties share the same information? Blockchain, has come to the rescue to take permanence to a whole new level. Blockchain allows information that endures, that is safe against the tampering and that record all changes even when decentralized between different actors. At the same time, it continues to benefit from the cost and speed advantages of digital and the transmission power and updatability databases and the internet. In turn, this could unlock processing, because you can safely embody that information in the physical world and make “law into code”. So it promises to make true the fantasy of magical incantations through automatic processing of smart contracts. Blockchains are currently struggling with challenges around the volume of information they can process, the speed of transactions and the cost to secure the network. As we can see in the graph the improvement of information technologies has been exponential over time and we can expect it to iron out those problems over the next years. We can also expect the next revolution soon enough (few years), building on blockchain and all previous advances. “Brains 2.0” focused on taking processing to the next level through AI could be that advance, so watch out for the marriage of blockchain and AI in the next wave of change. Magic could finally be made true with words spoken creating direct impact on the world.

Road to Blockchain: Evolution of Information Technology

0

road-to-blockchain-evolution-of-information-technology-1bd3ab041c9d

2018-04-08

2018-04-08 07:22:44

https://medium.com/s/story/road-to-blockchain-evolution-of-information-technology-1bd3ab041c9d

false

969

null

Blockchain

blockchain

Blockchain

265,164

Jaime Rodriguez-Ramos

Impact of exponential technologies on society and business.

3af872479b37

jaime.rodriguezramos

43

21

20,181,104

null

0

null

0

31f4f88d6548

2017-10-09

2017-10-09 14:25:49

2017-10-09

2017-10-09 15:03:22

1

false

en

2017-10-09

2017-10-09 17:26:22

2

1bd3afcb8788

3.637736

4

0

Learn more about the man behind one of the most competitive M.S. in Data Science programs in the nation, and his vision for our CDS…

5

5 Minutes with Arthur Spirling, Deputy Director and Director of Graduate Studies Learn more about the man behind one of the most competitive M.S. in Data Science programs in the nation, and his vision for our CDS Master’s students Arthur Spirling joined NYU in 2015 as an Associate Professor of Politics and Data Science, and as a steering committee member of the Moore-Sloan Data Science Environment, a grant housed by CDS. This Fall 2017, he not only continues to be our Deputy Director, but also starts his tenure as our Director of Graduate Studies for the Master’s in Data Science program. His research focuses on analyzing text-as-data to answer questions in political science, and he also the co-organizer of the text-as-data speaker series with Professor Sam Bowman. Before coming to NYU, he was the John L. Loeb Associate Professor of the Social Sciences at Harvard University. 1. The M.S. program in Data Science has been rapidly evolving since we launched it in 2013. What are some of the new changes to the program this year? As with every year of the MSDS, we have expanded our numbers — yet simultaneously become more selective. Our fantastic incoming cohort of 96 were selected from some 1659 applicants, meaning our program is considerably harder to enter than many top law schools or business schools. Every single one of our new students has truly remarkable potential and I know my colleagues are very happy to be working with them. On the curriculum side, this year we are excited about the addition of ‘tracks’ that allow our students to specialize in certain areas of Data Science, like Big Data, Math and Data, Natural Language Processing and Physics. Next year, we will add Biology. Ultimately, we want to provide a program structure that links domain knowledge and methods via the course offerings. This enables CDS and its students to place themselves at the center of Data Science as the field’s popularity explodes. Part of our vision at CDS is that Data Science is not just about learning powerful new methods; it’s also about having a deep understanding of the ways those methods can be used in the ‘real world’ of industry and policy. Our tracks — and the demand for them from students! — are proving a great way to make this vision a reality. 2. In addition to your role as DGS, you also co-organize our popular text-as-data speaker series. Analyzing text-as-data is a major part of your research. How did you start getting involved with analyzing text-as-data? Why do you think it’s becoming such a compelling methodology? Human beings have been writing things down for around 5000 years, but it’s only very recently in human history that anyone other than the social elite were producing texts. And, even when they did, it often wasn’t preserved for future research purposes. A major change came with the advent of the Internet, social media, and news sites. Now, literally billions of people write billions of words every day, whether they be online newspapers, product reviews, government reports, or someone commenting on how cute a friend’s baby looks on Facebook. From a research perspective, we can easily access that information in machine readable form: these huge troves of text data are ready to be analyzed immediately, often in real time. At the same time, the technology to make older documents machine readable has also advanced remarkably: one can now take, say, government records from World War II, push them through an optical character recognition system, and have quite high quality documents amenable to statistical work. With the explosion of text data has come methods for dealing with these collections, and that symbiotic relationship seems set to continue. Personally, I became involved in text-as-data because I was studying a particular historical puzzle: the democratization of the UK in the 19th Century. It’s an interesting case because in a relatively short period of time (around 80 years) politicians there embarked on very radical reform, going from a narrow franchise where no one could vote, to one where everyone could. This is surprising: generally, elites don’t voluntarily give up power to people poorer and less educated than they are. But, more broadly, I noticed that there was actually a lot we didn’t know about politicians and voters back then: how they interacted with each other, how they spoke in parliament, how they organized policy-making and so on. Simultaneously, I also realized that we had millions of records of speeches from which one could make an inference. So I turned to modern technology to understand these events: it’s been a great experience, and I learned a lot both substantively and in terms of methods! 3. You’ve been at CDS for a while now! What has been your favorite experience so far? Time flies: it’s my third year! My favorite experience is seeing how MSDS students develop during their time with us, and helping them accomplish their goals professionally. Our students work very hard: our courses are technically tough, and demand a lot of hours of focus and effort. The students persevere and, more often than not, they land themselves in their dream job at tech firms, banks, in government, or in academic research institutions. They are rightly proud of how far they have come — and we are proud of them, too. Interview conducted by Cherrie Kwok

5 Minutes with Arthur Spirling, Deputy Director and Director of Graduate Studies

36

5-minutes-with-arthur-spirling-deputy-director-and-director-of-graduate-studies-1bd3afcb8788

2018-02-16

2018-02-16 10:48:55

https://medium.com/s/story/5-minutes-with-arthur-spirling-deputy-director-and-director-of-graduate-studies-1bd3afcb8788

false

911

This is the official research blog of the NYU Center for Data Science (CDS). Established in 2013, we are a leading data science training and research facility, offering a MS in Data Science and, as of 2017, one of the nation’s first universities to offer a Ph.D. in Data Science.

null

nyudatascience

null

Center for Data Science

ab4829@nyu.edu

center-for-data-science

DATA SCIENCE,DATA MINING,TECHNOLOGY,ARTIFICIAL INTELLIGENCE,MACHINE LEARNING

NYUDatascience

Data Science

data-science

Data Science

33,617

NYU Center for Data Science

Official account of the Center for Data Science at NYU, home of the Master’s and Ph.D. in Data Science.

880781a85c2

NYUDataScience

3,530

9

20,181,104

null

0

{ "FirstName": "Bob", "ID": 1, "PhoneEntries": [ { "Category": "Personal", "Address": "(555)-555-5555" } ] } // ComposeSMSCommandNativeData { "SMS": { "Recipients": [ { "PhoneNumber": "(555)-555-5555", "Contacts": [ { "FirstName": "Bob", "PhoneEntries": [ { "Category": "Personal", "Number": "(555)-555-5555" } ], "ID": 1 } ], "ToUserWrittenName": "Bob's Personal", "ToUserSpokenName": "Bob's Personal" } ], "Body": "", "To": { "MissingPhoneNumbers": [], "DisambiguatePhoneNumbers": [], "ValidPhoneNumbers": [ { "ContactName": "Bob's Personal", "PhoneNumber": "(555)-555-5555" } ] } }, "FocusHint": "Body", "AutoListen": true, "State": "ExpectingStartOfBody", "OverwroteBody": false } // ComposeSMSResult { "SpokenResponse": "What's your message?", "SpokenResponseLong": "What's your message?", "WrittenResponse": "What's your message?", "WrittenResponseLong": "What's your message?", "AutoListen": true, "UserVisibleMode": "Compose Text Message", "ConversationState": { "ConversationStateTime": 1519154951, "Mode": "SMS", "VoiceActivityDetectionMinima": { "MaxSilenceAfterPartialQuerySeconds": 3 }, "SMSConversationStateContent": [ { "CommandKind": "ComposeSMSCommand", "ComposeSMSCommandKind": "SMSCreateCommand", "PropagatedData": { "SMS": { "Recipients": [ { "PhoneNumber": "(555)-555-5555", "Contacts": [ { "FirstName": "Bob", "PhoneEntries": [ { "Category": "Personal", "Number": "(555)-555-5555" } ], "ID": 1 } ], "ToUserWrittenName": "Bob's Personal", "ToUserSpokenName": "Bob's Personal" } ], "Body": "", "To": { "MissingPhoneNumbers": [], "DisambiguatePhoneNumbers": [], "ValidPhoneNumbers": [ { "ContactName": "Bob's Personal", "PhoneNumber": "(555)-555-5555" } ] } } }, "State": "ExpectingStartOfBody", "SMSDynamicResponseKind": "ComposeSMSResult" } ] } } // ComposeSMSCommandNativeData { "SMS": { "Recipients": [ { "PhoneNumber": "(555)-555-5555", "Contacts": [ { "FirstName": "Bob", "PhoneEntries": [ { "Category": "Personal", "Number": "(555)-555-5555" } ], "ID": 1 } ], "ToUserWrittenName": "Bob's Personal", "ToUserSpokenName": "Bob's Personal" } ], "SpokenBody": "Where are you?", "WrittenBody": "Where are you?", "Body": "Where are you?", "To": { "MissingPhoneNumbers": [], "DisambiguatePhoneNumbers": [], "ValidPhoneNumbers": [ { "ContactName": "Bob's Personal", "PhoneNumber": "(555)-555-5555" } ] } }, "LastEditedTextFieldType": "Body", "NewBody": "Where are you?", "AutoListen": true, "State": "ExpectingContinuationOfBody", "EditBodyState": "SET", "FirstTimeAddingToBody": true }

7

ef24c943f19e

2018-02-20

2018-02-20 17:57:41

2018-02-27

2018-02-27 21:18:43

4

false

en

2018-02-27

2018-02-27 21:18:43

6

1bd49dc59fb1

3.613208

2

0

Learn what Dynamic Responses are and how to use them to manage application state.

5

Working with Dynamic Responses Learn what Dynamic Responses are and how to use them to manage application state. Some Houndify domains may return more than one response and require clients to choose between those responses. These domains are using a feature called Dynamic Responses. In this tutorial, we’ll talk about what Dynamic Responses are, why they are useful, and how to leverage them in your client. Checking if a domain uses Dynamic Responses You can determine if a domain is using Dynamic Responses by checking the Client Integration Requirement section on the domain’s documentation page. Additionally, the Try API tool will warn about this requirement and will display both default and success responses. Home Automation Domain returns dynamic responses on the Try API Widget. Let’s dig deeper into how Dynamic Responses work by looking at the SMS domain as an example. Example: Dynamic Responses in the SMS Domain The SMS domain allows composing text messages to your uploaded contacts using conversation state and dynamic responses. Let’s walk through how this works. Assume that we have uploaded the following contact information. The Syncing User Data tutorial covers how to upload JSON contact such as the one below. Syncing your user data How to enhance queries for Contacts and Home Automation Domains.medium.com 2. If we make a query “Text Bob” we will get an SMSCommand JSON result. It will contain NativeData field of type ComposeSMSCommandNativeData. It will also contain the information required for sending the SMS. Apart from NativeData, the result will contain several dynamic responses. You can choose which response to use based on the state of your client. ClientActionSucceededResult: Use this if the SMS was sent correctly. ClientActionFailedResult: Use this if the SMS failed to send. NoSMSAppResult: Use this if there is no SMS application available on the client. ComposeSMSResult: Use this if you are missing the body of the SMS, and have only received the sender’s information. DisambiguationSMSResult: Use this if the query cannot be fulfilled because the user didn’t provide enough information to uniquely specify the needed information. ExitSMSResult : Use this to Since we are missing the body of the text, we should choose ComposeSMSResult and continue the conversation. We pick the ConversationState for this dynamic response and send it with the next query that contains the text message. 3. After we send “Where are you?” as a follow up query we will get similar set of dynamic responses, but ComposeSMSResult will now respond with “Would you like to send, review, continue, or cancel?”. NativeData will contain all the data we need to send the text. 4. If we reply with query “Send it” and ConversationState from ComposeSMSResult, NativeData will contain flag SendSMSNow: true. Now we can send the SMS. Note that it is a client’s responsibility to send the actual text and choose the correct dynamic response based on the status. Example script for testing SMS domain using Houndify Python SDK We hope that this tutorial helps you understand how Dynamic Responses work. Leave a comment if you have any questions.

Working with Dynamic Responses

60

working-with-dynamic-responses-1bd49dc59fb1

2018-05-31

2018-05-31 06:02:02

https://medium.com/s/story/working-with-dynamic-responses-1bd49dc59fb1

false

772

Engineering posts on how to build products using Houndify.

null

Houndify

support@houndify.com

houndify

FRONT END DEVELOPMENT,VOICE RECOGNITION,ARTIFICIAL INTELLIGENCE,API,SPEECH RECOGNITION

null

Houndify

houndify

Houndify

28

Asif

null

283d469678a1

aisfo

31

41

20,181,104

null

0

null

0

null

2018-01-30

2018-01-30 11:42:35

2018-01-30

2018-01-30 11:48:12

1

false

ru

2018-01-30

2018-01-30 11:48:12

6

1bd5d9627d0a

1.773585

7

0

Как научить компьютер сопереживанию и состраданию не только к человеку, но и другому компьютеру

3

ИИ обладающий эмпатией и сознанием Как научить компьютер сопереживанию и состраданию не только к человеку, но и другому компьютеру Шоумо Банерджи из Математического института Оксфордского университета не перестает меня удивлять. И похоже, что уже в этом году его вместе со всеми его работами купит какой-нибудь Google или IBM, после чего все эти работы засекретят. Но пока нам раздолье, и можем свободно читать. Только что Банерджи передал на рецензирование научной общественности новую революционную работу на стыке ИИ и двух теорий: Адаптивных самоорганизующихся и Мультиагентных систем. Работа называется «A framework for designing compassionate and ethical artificial intelligence and artificial consciousness» (Основа для разработки способного к состраданию и этике ИИ, а также искусственного сознания) https://goo.gl/BnTres Революционность этой работы в том, что это 1я в мире попытка, если не примирить, то хотя бы сблизить две полярные точки зрения на сознание. Несколько упрощая, эти 2 полярные точки зрения таковы. 1) все элементы сознания так или иначе сводимы к вычислительным процессам (и, следовательно, могут быть воспроизведены компьютером); 2) есть нечто большее в человеческом осознанном поведении, что невозможно описать вычислительным процессом (например, творчество или чувство свободы, которыми обладают люди, похоже, не связаны с логикой или расчетами). Что стоит за каждой из этих 2х точек зрения, я, пожалуй, напишу поподробней в отдельном посте. Здесь же лишь отмечу, каким образом Шоумо Банерджи предлагает решение на стыке 2х полярных подходов. Это — «гибридный интеллект», обладающий верхним уровнем сознания. По сути, подход Банерджи продолжает и расширяет подход IBM, реализуемый в IBM Watson. Но если Watson основан на глубоком обучении компьютера с использованием обширного корпуса сетевых текстов, подход Банерджи предусматривает включение в контур глубокого обучения еще и человека-оператора. Описывать здесь детали подхода Банерджи не вижу смысла. Это будет длинное и не простое переложение текста статьи Банерджи. Кому интересно — прочтет сам в оригинале. Но если будете читать, обратите внимание, что ключом к переходу от моделирования интеллекта к моделированию сознания, Банерджи называет способностью компьютера сопереживать и сострадать. И не только другому человеку, но и другому компьютеру. О предыдущей работе Шоумо Банерджи — революционной теории преступности и насилия в городах, построенной по аналогии с иммунной системой человека, где преступность (и сами преступники) уподобляется патогенным инфекциям, а ответ на нее со стороны общества (полиция, суд и тд) — иммунным ответам организма — см. https://t.me/theworldisnoteasy/292 . Фантастический сплав современной теории городов, как биологических организмов, вычислительной социобиологии и теории искусственных иммунных систем. _________________________ Хотите читать подобные публикации? Подписывайтесь на мой канал в Телеграме, Medium, Яндекс-Дзене Считаете, что это стоит прочесть и другим? Дайте им об этом знать, кликнув на иконку “понравилось”.

ИИ обладающий эмпатией и сознанием

12

ии-обладающий-эмпатией-и-сознанием-1bd5d9627d0a

2018-02-07

2018-02-07 10:22:59

https://medium.com/s/story/ии-обладающий-эмпатией-и-сознанием-1bd5d9627d0a

false

417

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Сергей Карелов

Малоизвестное интересное на стыке науки, технологий, бизнеса и общества - содержательные рассказы, анализ и аннотации

4fa09a8333b2

sergey_57776

2,156

86

20,181,104

null

0

null

0

null

2017-11-06

2017-11-06 01:39:43

2017-11-06

2017-11-06 01:54:59

2

false

ja

2017-11-06

2017-11-06 05:38:44

3

1bd5fb70cb72

2.251333

3

0

GLIT Adminβ(企業向け管理画面)』を企業様にすでに提供を開始しておりましたが、いよいよGLITのコア機能である”マッチング機能”をリリース致しました。

5

マッチングした候補者とメッセージでやり取りが可能な機能をリリース！ GLIT Adminβ(企業向け管理画面)』を企業様にすでに提供を開始しておりましたが、いよいよGLITのコア機能である”マッチング機能”をリリース致しました。そちらの内容について、本ブログでまとめさせて頂ければと思います。マッチング機能について GLIT Adminβより入稿された求人情報(GLITオリジナル求人)は適宜GLITユーザーにレコメンドされていっております。GLITオリジナル求人をユーザーが興味ありにスワイプすると、ユーザーである候補者と企業がマッチングし互いにメッセージのやり取りを行うことが可能になります。ただしGLITオリジナル求人を興味ありにスワイプした全てのユーザーとマッチングする訳ではありません。GLIT側の方で、ユーザーが企業側とマッチしそうかどうかや、プロフィールの入力状況などから判断し一次スクリーニングは行わせて頂きます。その上で、マッチしそうなユーザーを企業側におしらせします。そこから企業はユーザーに対してメッセージを送るかどうかを判断することが可能になり、企業がメッセージを送って初めてユーザー側にマッチングしたというお知らせが届く流れとなります。それでは、改めてGLIT Adminβ(企業向け管理画面)よりどんなことができるのかをまとめておきたいと思います。企業向け管理画面からは下記のようなことが可能です。・求人の入稿、配信・求人の編集、募集停止・求人に対するインプレッション数/興味あり/なし数の把握・フィード(採用広報)への発信機能・マッチング＆メッセージ機能 ←今回リリース企業向け管理画面より求人を入稿頂くと、適宜『GLIT』のユーザーに求人がレコメンド配信される形になります。今後の実装予定機能はこちら・求人の削除・求人のプレビュー機能・アナリティクス機能現在はβ版のため、サービスの利用料は完全無料となっております。興味を持って頂いた企業様はこちらより、お問い合わせください。＜忙しいビジネスパーソンのためのAI求職アプリ『GLIT』＞ LP：https://glit.io AppStore：http://appstore.com/glit

マッチングした候補者とメッセージでやり取りが可能な機能をリリース！

5

matching-message-1bd5fb70cb72

2018-02-27

2018-02-27 05:00:14

https://medium.com/s/story/matching-message-1bd5fb70cb72

false

26

null

Recruiting

recruiting

Recruiting

15,454

AI求職アプリ『GLIT(グリット)』

カジュアルに出会えるリクルーティングアプリ。レコメンドされる求人情報をスワイプするだけ！Tinder風のUI/UXが特徴です。https://glit.io

8b2cde18cc1b

glit_hr

6

105

20,181,104

null

0

null

0

null

2017-09-12

2017-09-12 16:27:54

2017-09-12

2017-09-12 16:28:28

0

false

en

2017-09-12

2017-09-12 16:28:28

1

1bd6395b4c62

0.532075

0

One more time, an excellent article by Gregory Allen, who is the author of Harvard Study on AI. Musk and Putin are right — of course — but…

3

I Believe We Could Do That. One more time, an excellent article by Gregory Allen, who is the author of Harvard Study on AI. Musk and Putin are right — of course — but he did not say very much about China, because he knows full well that China is ALREADY in the lead in AI, and it will only get worse quite soon — in the military sphere, in the AI arms race. You better think seriously about what to do, and I believe the best option is for a one world government, right now, specifically because of the AI military arms threat. Then all we would have to do is wipe up some terrorists, and with the world united, I believe we could do that. https://medium.com/cnn-opinion/putin-and-musk-are-right-whoever-masters-ai-will-run-the-world-cf47b3222b57

I Believe We Could Do That.

0

i-believe-we-could-do-that-1bd6395b4c62

2017-09-12

2017-09-12 16:28:29

https://medium.com/s/story/i-believe-we-could-do-that-1bd6395b4c62

false

141

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Peter Marshall

I am extremely interested in AI, especially the not-so-good side of AI weapons and AI war, although the good parts are magnificent and wonderful too, naturally.

f6bab8ee3d29

ideasware

1,765

276

20,181,104

null

0

null

0

820f4931b42

2018-07-02

2018-07-02 16:36:11

2018-07-10

2018-07-10 20:43:29

4

false

en

2018-07-10

2018-07-10 20:43:29

9

1bd64a9e3f2

3.375472

1

0

Subscribe to SheReports

5

The Future May Be Female, but Does Your Voice-Activated Assistant Need to Be? Subscribe to SheReports What Brand Marketers Need to Consider when Implementing AI As technology develops at a dizzying pace, with smarter and more humanlike robots and artificial intelligence (AI) emerging, and quickly being adopted by consumers, brand experts and marketers need to make sure that we are paying close attention to how gender is portrayed. This will affect how young women perceive themselves and how the world around them perceives them. The tech industry is well known for its struggles with gender bias. According to a study by Babson College (albeit from a few years ago), only 2.7% (183) of the 6,517 companies that received venture capital funding have a woman in the CEO role. Without more women in the field that number will not get to where it needs to be. And in 2017, the U.S. Department of Commerce released a report stating that women held only 24% of STEM (science, technology, engineering and mathematics) jobs in the U.S. That report also notes that women with STEM degrees are less likely than their male counterparts to work in a STEM occupation; they are more likely to work in education or health care. … brand experts and marketers need to make sure that we are paying close attention to how gender is portrayed. This will affect how young women perceive themselves and how the world around them perceives them. When there are fewer women working in the field, inclusivity is not top of mind for male developers and creators on staff. A thoughtful op-ed in Branding in Asia on the subject of the rising paradox of gender and branding — and what brands can do — makes the cases that the gender gap in STEM likely led to the fact that the default for voice-activated assistants are female voices. [I]n 2017, the U.S. Department of Commerce released a report stating that women held only 24% of STEM (Science, Technology, Engineering and Mathematics) jobs in the U.S. “They have all been given obedient, servile, female personalities. They are turning our lights on and off, ordering our shopping. Whereas for more high-powered tasks, such as making business decisions, AI is often given a male personality — take IBM’s Watson or Salesforce’s Einstein,” says AI technologist Kriti Sharma. And the name Siri? It means “beautiful woman who leads you to victory” in Norwegian. While 94.6% of (human) administrative assistants and secretaries are women, the best way to change that is to let the world hear the voice of a male assistant. Many of the companies behind voice-activated assistants have gotten the message and are now doing just that, giving users the option of hearing a male voice. In Google’s case, one of the male voice options is even that of singer John Legend. But marketers should have considered this in advance and not have to be playing catch-up — no matter how dulcet John Legend’s tones are. And the name Siri? It means “beautiful woman who leads you to victory” in Norwegian. In the A+E Networks® Research Womanhood study, younger women expressed that they would like to see more women in STEM fields and more women in positions of power. Forty percent of those felt strongly about seeing more women represented (in media and advertising) in these roles. Having a subservient, disembodied female voice ordering lunch doesn’t exactly help. Additionally, when younger women find a technology product that they like, they are 6% more likely than the average adult to recommend it to others. So the tech brands developing products appealing to what these millennial women want to see are thinking smart. Forty percent of those felt strongly about seeing more women represented (in media and advertising) in these roles. Having a subservient, disembodied female voice ordering lunch doesn’t exactly help. One way to do that, the article in Branding in Asia argues, is for brands to help enable STEM education for young women. Another is to make conscious and responsible decisions about gender when implementing it into AI. The final piece of advice: Be agile and embrace the nonbinary instead of simply conforming to the prejudices in the world. Subscribe to SheReports

The Future May Be Female, but Does Your Voice-Activated Assistant Need to Be?

1

the-future-may-be-female-but-does-your-voice-activated-assistant-need-to-be-1bd64a9e3f2

2018-07-10

2018-07-10 20:43:29

https://medium.com/s/story/the-future-may-be-female-but-does-your-voice-activated-assistant-need-to-be-1bd64a9e3f2

false

709

Perspectives: At the forefront of media storytelling, we cultivate, illuminate and share narratives that help us better understand the human experience.

null

aenetworks

null

A+E Networks Publication

AENPublishing@aenetworks.com

aenetworks

AENETWORKS,SHEREPORTS,THOUGHT LEADERSHIP,MEDIA

AENetworks

Women In Tech

women-in-tech

Women In Tech

18,577

A+E Networks

The Account #LifeMagnified

ada04340ae00

AENetworks

29

24

20,181,104

null

0

null

0

null

2018-08-17

2018-08-17 03:45:34

2018-08-23

2018-08-23 05:01:01

2

false

en

2018-08-23

2018-08-23 05:01:01

4

1bd762c35c14

2.096541

1

0

Political Analysis selected a paper co-authored by our very own Jason Morgan, Vice President of Behavior Intelligence here at Wiretap and a…

5

Dr. Jason Morgan selected for Editor’s Choice Award by Political Analysis Wiretappers make an impact wherever they go. Political Analysis selected a paper co-authored by our very own Jason Morgan, Vice President of Behavior Intelligence here at Wiretap and a Visiting Scholar at OSU, along with Professors Janet Box-Steffensmeier of The Ohio State University and Dino Christenson of Boston University, for their annual Editor’s Choice Award. An abstract of the paper, titled Modeling Unobserved Heterogeneity in Social Networks with the Frailty Exponential Random Graph Model is shown below. Abstract In the study of social processes, the presence of unobserved heterogeneity is a regular concern. It should be particularly worrisome for the statistical analysis of networks, given the complex dependencies that shape network formation combined with the restrictive assumptions of related models. In this paper, we demonstrate the importance of explicitly accounting for unobserved heterogeneity in exponential random graph models (ERGM) with a Monte Carlo analysis and two applications that have played an important role in the networks literature. Overall, these analyses show that failing to account for unobserved heterogeneity can have a significant impact on inferences about network formation. The proposed frailty extension to the ERGM (FERGM) generally outperforms the ERGM in these cases, and does so by relatively large margins. Moreover, our novel multilevel estimation strategy has the advantage of avoiding the problem of degeneration that plagues the standard MCMC-MLE approach. More About Dr. Jason Morgan Equipped with his robust and rich academic background, Jason Morgan fearlessly leads Wiretap’s data science team every day. Jason leverages artificial intelligence and data science to develop predictive natural language and network models for enterprise collaboration platforms. He explains more in his piece discussing how AI is protecting company culture on collaboration platforms. In addition to solving large enterprise problems, Jason is a true educator and academic. He has taught courses in network modeling and statistical methodology at the undergraduate and graduate level and is currently co-authoring a textbook on social network models titled Inferential Network Analysis, which is under contract with Cambridge University Press. Additionally, Jason has contributed to notable academic journals such as, Annals of Applied Statistics and Political Analysis, and presented at top national conferences, including Black Hat. Continue Reading Learn more about his work at Wiretap in his recent presentation at the Columbus Techstars StartupWeek on building a data science team at a start-up Artificial Intelligence (AI) at a Startup: From 0 to 1 Our own Vice President of Behavioral Intelligence at Wiretap, Dr. Jason Morgan presented on building a data science…www.wiretap.com Liked this? We’d appreciate it if you clicked the 👏 a couple of times to show your love and help others find this article. This first appeared on The Wire, Wiretap’s blog: www.wiretap.com/blog

Dr. Jason Morgan selected for Editor’s Choice Award by Political Analysis

2

dr-jason-morgan-selected-for-editors-choice-award-by-political-analysis-1bd762c35c14

2018-08-23

2018-08-23 05:01:01

https://medium.com/s/story/dr-jason-morgan-selected-for-editors-choice-award-by-political-analysis-1bd762c35c14

false

454

null

Data Science

data-science

Data Science

33,617

Wiretap

💬🤝🛡Control Your Data. Protect Your People. Guard Your Reputation | On @WorkplacebyFB @MicrosoftTeams & @Yammer | @Gartner_inc named us a 2018 Cool Vendor 😎

b3b32d6c1a7

awarebywiretap

4

133

20,181,104

null

0

null

0

null

2018-07-13

2018-07-13 22:29:30

2018-08-03

2018-08-03 18:11:04

1

false

en

2018-08-04

2018-08-04 04:48:11

1

1bd9e5ff8e03

10.815094

1

0

An Open Letter to Dr. Hossenfelder, @sdkh, Including the Purpose of Life.

4

A Treatise on Free Will Greetings Dr. Hossenfelder; Again thank you for writing your book, Lost in Math. Recently you have both written and spoken on the need for further investigation into such metaphysical subjects as free will and the purpose of life, particularly in how these two are related to the physical sciences. I submit here some of my thoughts and hope you may consider them. Before the metaphorical monsters of free will and purpose can be tamed, let us go over some basics of rational thought. Usually, we begin with some assumptions, then utilize various transitions and perhaps reach conclusions. It is most important when reaching conclusions, to objectively examine them, once they are determined, and decide if they are “sensible” or concur with observation. For example, we may arrive at very sound and logical conclusions regarding social behavior which are unethical. You have mentioned in one of your promotional videos that your parents warned you that not taking stock of what is occurring in very objective terms, can result in the most depraved behaviors of large groups of people. Your objective critique of physics is commendable and extremely important. Therefore, if we find our conclusions are “absurd”, immoral or simply do not concur with observation we must examine closely the transitions taken to reach these conclusions. We also examine the data. Obviously, if the data and transitional steps are in order, it becomes clear that our assumptions require examination. And, more than likely, discarded. Dr. Hossenfelder, you are being cut by the razor. The Purpose of Life It can be seen, by inspection, that if life has a purpose, then we have no free will. Having a purpose and having free will are mutually exclusive. This is a good thing. Jean-Paul Sartre points out that the universe is absurd — it has no purpose. Furthermore, neither do we. Some people, as a matter of fact a great many people, believe they have a purpose in their lives; but that is their decision. And that is good. And they subject their free will, voluntarily, to the purpose upon which they have decided. Allow me to elucidate please: Looking at ecological niches, say we have a bunny. The bunny fits into an ecological niche and has a purpose into which it has evolved: for example, to eat certain leaves, to procreate and to end up being food for other life forms. It is driven to fulfill this purpose through instincts and drives. Even if we remove the bunny from its niche and place it in a cage or zoo, it will still do everything it can to fulfill its purpose: to eat various foliage, to procreate, if possible, and end up being food for other life forms. The bunny cannot help this. It was born to fulfill its purpose and that is what it will do or at least try to do. Mind you, it can still make decisions. It can still choose one leaf over another and make many decisions in its life. But is still under the directives of its purpose for the niche into which it has evolved. In this way, regarding its drives and what we call instincts, the bunny has no power to overcome these boundaries to its life. In effect, as a result of its purpose, ecologically speaking, it has no free will and this is not the result of any decision made by the bunny. A bunny does what a bunny does. Applying this to human beings we gain some insight. Biologically, our purpose is to procreate, to keep segments of our DNA “alive”. However, even after having children and raising them (and they leave us), we still have about half of our lives left to live. Why? There is no purpose any more for us and we should be polite and leave. But we did not evolve that way. Furthermore, does the procreative definition of purpose, that we are here only to procreate, mean that women, or men, who decide not to have children have no purpose or that their lives are a complete waste of time? Even biologically speaking, this would be an absurd conclusion. We are at the top of the food chain. Ecologically, we fulfill no niche. Even though this probably will result in the destruction of all planetary life including our own, we are at the top of the food chain and have no purpose as far as the overall ecological system of planet Earth is concerned. As a matter of fact, life would get along a lot better without us. So, let us return to the assumption itself, that there is some purpose to our lives. If we have a purpose which is beyond our control; we have no free will, we are enslaved to that purpose. We may choose to rail against such a purpose, but our lives would still be directed by that purpose, whether we abide to it or rebel against it. So I would like you to entertain the idea that your life has no purpose and neither does the universe. We are free. We are so free that we can decide to have purpose or not. And should we decide to have a purpose, like being of service to others, or being a responsible parent or citizen or human being; then there is nothing stopping us from making such a decision. And we can change our minds. We can stop having a purpose and decide to have none. And then decide to have a completely different purpose later on. As we commit to such decisions, we surrender our free will, of our own free will, to dedicate ourselves to such purposes. These are the things heroes are made of. Free Will What do we mean by free will? Perhaps it has something to do with having the ability to make a decision. But this decision must be free of outside influence for it to denote free will. It may be that this decision could be completely absurd and irrational. Socrates concluded that if we know the truth about things, really knew the cause and consequence of our decisions, that we would have no free will at all. We must decide to do that which is true, (or right, I guess, in the classical Greek philosophical sense). We cannot live a lie. Unfortunately, there are many who live their lives as a lie. The academic world of the physical sciences is filled with such poor and miserable people. It is definitely possible to live a life of falsehood contrary to Socrates’s conclusion. Also, it is possible to live a life of falsehood knowing all you do is absurd and promotes more and more the acceptance of a bald faced lie. And there are people who try to rationalize this kind of a life by claiming it is necessary in order to survive or have greater material reward. And yet they sigh in vain for a life of meaning (see purpose, above). In other words, even though there is no rational, a person can choose to live a life of complete dishonesty, insane though such a life would be. Hannah Arendt writes extensively on this set of circumstances. Are our actions determined through our genetic code or through what we have learned in life? Responding to this question: It doesn’t really matter. Either way, our decisions would be pre-determined and we would have no free will. Agreed, what we are born with, regardless of our genetic grouping, traumas and other events of our lives have a great effect on our behavior. But we can overcome them. We can make decisions to change. And herein lies an insight into what free will is. Consider the computer program Deep Blue written by IBM staff which has the purpose to win chess games. Its algorithm is simply not to make the same mistake twice. It plays a lot of games and loses them and keeps track of every game it played, never making the last move given the pieces’ positions from the last lost game. Should all previous moves be exhausted, it drops a level and never makes the same second from last move given the pieces’ positions from the last lost game. And so on. Eventually it starts winning. Its moves are determined completely by all the previous games it has played. If it is presented with pieces in positions it has not seen before, it just picks a move from an ordered list, and continues on. Even though it has lost a tremendous number of games, it is a computer program which has defeated a grand master. The last human to beat it was Gary Kasparov, who later lost in tournament conditions. Deep Blue, an incredibly inefficient and simple algorithm, is touted as a breakthrough in artificial intelligence. But it has no free will. Its moves are predetermined by the games it has played. In the same way, Kasparov plays chess according to his previous experience and his memory. But Kasparov has free will. He can choose what path he takes in spite of previous experience. Deep Blue cannot. Let us expand this idea of predetermination vs free will. Let us delve into the wonderful world of the Laplace paradox. Actually, it is probably better known as Laplace’s demon. This demon tells us that if the laws of physics are deterministic, (which they are), and initial conditions are known; then all subsequent actions of everything in the universe is determined from the initial conditions. Or, to put it a little more clearly, if you know how everything behaves as a result of what happens to it, and you know where, when and how everything started, then you know how everything is going to be forever after. Even more clearly, there is no free will, everything was put in motion in the beginning and everything from then on is completely determined. This is a truly delightful demon. So, the question, in response to Laplace’s demon, is why does Kasparov have free will and Deep Blue does not? A flat worm has more free will than Deep Blue. So how come? What is the difference between Kasparov and Deep Blue that gives Kasparov free will? And what does this have to do with physics? Deep Blue comes with a set of deterministic instructions over all its parts. And Kasparov comes with a set of deterministic instructions over all his parts. Differences are just a matter of scale. So it is not in the area of determinism that the demon can be exorcised. Here we bring in Green’s Theorem and Emmy Noether’s work regarding applied mathematics. There are three elements to applied math in general: a differential equation, boundary conditions and a solution which we denote as the deterministic instructions. If we know any two of these, we can determine the third. So, if we know the differential equation and boundary conditions, we can find the deterministic instructions. Likewise, if we know the deterministic instructions and the differential equations, we can find the boundary conditions. With Kasparov and Deep Blue, the deterministic instructions and differential equations are set. They obviously are not known by we mere mortals, but that does not matter. They exist. So long as they exist and they are unique, we’re golden. Therefore, we can conclude that the boundary conditions for Kasparov differ from the boundary conditions of Deep Blue. Dr. H., we both know that initial conditions are packaged in with boundary conditions. Using initial conditions: the initial conditions of Kasparov differ from the initial conditions of Deep Blue, which is obvious. However, a more concise question is: What, in particular, is different from the initial conditions of Kasparov as a chess playing machine than the initial conditions of Deep Blue as a chess playing machine? What initial condition(s) gives Kasparov free will and denies it to Deep Blue? I hope this nails down the demon so we can deal with him. The Angel of Eternity Let us first consider the bounds of Deep Blue. Deep Blue is a computer program with a set of instructions, and more importantly, a start date. A team at IBM cut the code, compiled it and set Deep Blue to running with an empty database. As it played, and there may have been many different options as to who Deep Blue played, which was up to its handlers; and could have had many different initial games, so there is some randomness involved; Deep Blue arrived at the state it is in. There could possibly be a Deep Blue Mark I and a Deep Blue Mark II who have played different people and played different games in different orders and the two may possibly have some individuality. But neither would have free will. How they play is predetermined. The key here, is that there is a start date for both programs and finite boundary conditions. There are a finite number of games they have played. Neither Deep Blue Mark I nor Deep Blue Mark II have any choice as to the next move either makes. What about Kasparov? Kasparov also has only played a finite number of games. And Kasparov has a start date … sorta. Um, just a second here. Does he really have a start date in the same way as Deep Blue? No, he does not. At the very beginning, at the initial boot of Kasparov, there were preconditions involved, pre-programs in the RNA of his mother and father. And we can continue this line of thinking, reverting all the way back to the origin of the universe itself. The boundary conditions for Kasparov involve the entire universe. Deep Blue’s universe, with which it interacts, is finite. If the universe of Kasparov is also finite, the demon has won. Now, after all this lead-up, we get to some physics. Is the belief in a finite universe an assumption? We assume the red-shifting of distant galaxies are from the Doppler effect of recession. We both know the difference between a vector and a scalar, but I will repeat it here for emphasis: A vector changes when we change the coordinate system; a scalar does not. If it is true that the universe is accelerating outwardly, then by taking the divergence of the expanding vector field, we are at a preferred place in the universe. We are the source of this outward acceleration. Divergence yields a scalar field, not a vector field. If we transform our measurements of expansion to any other place in the universe, (taking time delay into consideration), we will have a very different vector field. However, if we take the divergence of that vector field, no matter to where it has been transformed, we end up with the same scalar field as before and once again our lonely and insignificant planet is at the center of the universe. That is irrefutable. We are the source of this accelerating expansion no matter from where we have transformed our data. We can conclude that taking the assumption of red shift resulting from only outward Doppler effects, we are in a preferred place in the universe. And that cannot be. Furthermore, every second year astrophysics student is required to do the calculation to determine the final temperature of the cosmic microwave background. It is found to be out by an order of magnitude. Therefore, let us look at the Big Bang theory as though it is an assumption, which it is. Let us instead assume that there was no beginning to the universe and that it is infinite. And that there are non-conservative elements in the realm of physics, such as m-triple-dot within the Einstein field equations, or quality of energy as two examples. Even with these assumptions we still have not exorcised the demon since boundary conditions can be infinite and initial conditions can be in the eternal past. However, they have to be countable. And because of non-conserved elements in the universe, these boundary conditions in the eternal past are not countable. Therefore consider the following statement: the number of stars in the universe is not only infinite, it is uncountable. This is where the boundary enclosing Green’s Theorem breaks. Even though the rules of the game are deterministic and the differential equations known, (or determined), the boundary and initial conditions not only are indeterminate, they don’t exist. And that is the difference between Kasparov and Deep Blue. Deep Blue has finite initial and boundary conditions while Kasparov has no initial or boundary conditions, if we extend the argument to the n-th degree. To conclude: if there was a Big Bang, you have no free will. If the universe is infinite and eternal and contains non-conservative elements, then you have free will. The choice is yours.

A Treatise on Free Will

5

a-treatise-on-free-will-1bd9e5ff8e03

2018-08-04

2018-08-04 09:15:59

https://medium.com/s/story/a-treatise-on-free-will-1bd9e5ff8e03

false

2,813

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Bruce Rout

Gleanings from the back of an envelope.

3d97f267e906

bbrout

38

81

20,181,104

null

0

null

0

null

2018-03-15

2018-03-15 14:31:18

2018-03-15

2018-03-15 14:39:24

1

false

en

2018-03-15

2018-03-15 15:10:02

6

1bda21aef5a6

2.071698

0

Official recognition of RevLifter’s disruptive qualities arrives courtesy of a nomination at the 2018 UK Performance Marketing Awards…

5

It’s Official: RevLifter is a ‘Disruptor’ On February 28, RevLifter received news of its nomination for the inaugural Industry Disruptor prize at the 2018 UK Performance Marketing Awards. We will place up against some of the biggest and most exciting names in the industry, all of whom should be expecting a result when the ceremony rolls into London on April 24. As one of the founders of RevLifter, I’d like to firstly thank the PMA judging panel for recognising true innovation within the performance marketing space and giving our technology a platform ahead of its global launch, which is scheduled for later this month. While our solution does not compete directly with some of the companies on the shortlist, we’re confident that we can match these big names in terms of creativity, innovation and, of course, disruption. The competition is certainly fierce, but I’d imagine the same thoughts were echoing through the heads of fellow disruptors like Uber and Amazon when assessing what they’d be up against. We too believe in our product and its ability to revolutionise a core practice within e-commerce: incentivisation. The performance marketing space has represented something of a hotbed of innovation over the years. Lateral thinking is the driving force in an area where every improvement counts. Yet. in all my years on the frontline of performance marketing, it has been strange to see a lack of trailblazing strategies and solutions connected to the distribution of vouchers and other incentives. Countless studies have outlined an expectancy among consumers for personalised offers and promotions. Research from PwC shows that marketers are using personalisation to improve customer loyalty, up-selling and cross-selling opportunities, along with the general effectiveness of their campaigns. So why, then, are customers repeatedly left to fend for themselves in the search for incentives that will edge them closer to the checkout? RevLifter seeks to answer this issue through the use of an AI-powered deal library which, after consulting with first-party data gleaned from the retailer’s owned properties, can identify and serve personalised incentives to the customer in real time. It is inspired by our conversations with retailers — groups who were demanding more intelligent ways of incentivising their audiences — and we are set to go live with a number of global household brand names. Winning the 2018 PMA for Industry Disruptor would represent a huge achievement for the team at RevLifter, for whom performance marketing remains a passion. However, our nomination is simply the start of what will be an exciting journey to behold. To be among the first to hear of our launch and future updates, please follow @RevLifter on Twitter and our official Linkedin page, where we’ll also be sharing content geared around e-commerce and performance marketing trends. Until then, we’ll be putting the final touches on our global debut while keeping our fingers crossed for a positive result come April. Simon Bird Co-Founder, RevLifter Contact: General: hello@revlifter.com Press: PR@revlifter.com

It’s Official: RevLifter is a ‘Disruptor’

0

on-february-28-revlifter-received-news-of-its-nomination-for-the-inaugural-industry-disruptor-1bda21aef5a6

2018-03-15

2018-03-15 15:10:02

https://medium.com/s/story/on-february-28-revlifter-received-news-of-its-nomination-for-the-inaugural-industry-disruptor-1bda21aef5a6

false

496

null

Marketing

marketing

Marketing

170,910

RevLifter

RevLifter is the future of voucher code technology. Head to revlifter.com for the lowdown on our AI-driven solution.

751983f0ac0

revlifterHQ

1

2

20,181,104

null

0

null

0

c333fe01f6b8

2017-10-20

2017-10-20 01:39:06

2017-11-03

2017-11-03 21:26:47

1

false

en

2017-11-03

2017-11-03 21:26:47

2

1bda9693f4cd

1.815094

3

0

Interview by Breanne Thomas

5

Founder and CEO Nancy Shenker talks about the shift in marketing and introduction of AI Interview by Breanne Thomas Hi Nancy! Can you tell us a little about what you do and how you ended up there? I run my own marketing consulting company called theONswitch and am a content strategist/writer and professional speaker. I spent the early part of my career as a marketing executive for big brands — the classic “executive woman of the 1980s.” After reaching the C-level I decided to start my own business, providing start-ups and small businesses with the same quality of marketing that big brands enjoy. I’ve always been fearless in terms of technology innovation and how tech can enhance and streamline both business and life. When I decided to launch my next book/website, AI, machine learning, and robotics seemed like a natural — they will change every aspect of our lives over the next decade — much the way social/digital media transformed marketing. You’ve been in marketing for a long time and have seen many of its iterations. What has been the most significant shift you’ve noticed, whether in the industry at large or on a smaller, more personal scale? A.I. will be the most significant shift we’ve ever seen. Marketing will change with A.I. in ways that we must embrace — from skills needed, to hiring, customer experience, content, social media and more. But prior to that, the biggest change was the “democratization” of technology. As devices, user-friendly tech, and worldwide Internet proliferation put the power into the hands of consumers and business leaders outside of the tech industry at a rapid pace. What advice would you give to other Tech Ladies in marketing who want to embrace emerging technologies like AI or AR/VR in their own work? Stay current. Every day, new developments emerge. But also learn to differentiate between “AI-washing” and real artificial intelligence. Lots of people are slapping the term “AI” or “robot” on a product or service to increase its appeal. Beware the posers. The future of AI also depends on the intelligence and engagement of the humans behind it. Like all technologies, people need to look at each invention/development and ask the critical questions, like “Will this improve the quality of work? Save money? Make more money?” Knowing how to communicate the value and benefit of technology is as important (if not more important) than the technology itself. Tech Ladies connects women with the best jobs and opportunities in tech. Join the group, or submit a job posting at: www.hiretechladies.com

Founder and CEO Nancy Shenker talks about the shift in marketing and introduction of AI

41

founder-and-ceo-nancy-shenker-talks-about-the-shift-in-marketing-and-introduction-of-ai-1bda9693f4cd

2018-04-03

2018-04-03 01:34:49

https://medium.com/s/story/founder-and-ceo-nancy-shenker-talks-about-the-shift-in-marketing-and-introduction-of-ai-1bda9693f4cd

false

428

We connect women with the best opportunities in tech. We connect companies with the best women techmakers. www.hiretechladies.com

null

therealtechladies

null

Tech Ladies

allison@hiretechladies.com

tech-ladies

WOMEN IN TECH,TECH,DIVERSITY,STARTUPS,HIRING

hiretechladies

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Breanne Thomas

sometimes social media person 📱 sometimes filmmaker 🎥 sometimes both ✨

6a8ccd2ff5ab

breannethomas

276

190

20,181,104

null

0

null

0

null

2018-03-23

2018-03-23 15:16:56

2018-03-23

2018-03-23 22:01:01

5

false

en

2018-03-23

2018-03-23 22:01:01

10

1bdb5ef6ba4

3.893711

12

2

0

One day remaining until Token Sale

5

Effect.AI Token Sale Tutorial One day remaining until Token Sale With less than 12 hours to go until the Effect.AI Token Sale, now is the time to make sure you are ready for the big event. The Token Sale starts at 11am sharp tomorrow (CET). By now you have prepared the NEO/GAS you want to contribute and you’re up-to-speed on how the two Stages will work. All that remains is the Token Sale itself. This step-by-step guide will show you exactly how to participate in the Effect.AI Token Sale. This is the written version of the tutorial. Watch the video version here. Step 1: logging in Start by going to the official Effect.AI Token Sale URL: https://token.effect.ai This is the only correct URL. Visually confirm this is the address you are visiting. To start participating in the Effect.AI Token Sale, log into the account you used to sign up for the Whitelist. This has to be the same account. Enter your email address in the field, accept the Terms & Conditions, complete the captcha and click ‘request access.’ Go to your email account and click on the personal login link provided. If you can’t find the email, check spam folders. Clicking the link will direct you to your personal contribution page for the Token Sale. Step 2: accepting Terms & Conditions After logging in, you are greeted by the introduction screen. It shows you the minimum and maximum contribution amounts and how much time is left in this Stage of the Token Sale. The Effect.AI Token Sale is divided in two Stages. Stage 1, Fair Share Distribution, starts tomorrow at 11am. In this Stage, everyone has the opportunity to purchase an equal amount of the Effect.AI hard cap. The EFX tokens that remain are sold in Stage 2 on a first come, first serve basis. Stage 2 starts after 24 hours. Once you have checked the boxes confirming that you have read the Whitepaper and agree to the Terms & Conditions, click ‘next’ to proceed. Step 3: calculate your contribution amount To make it easy, this screen shows you exactly how much time is left and how many NEO/GAS you can contribute in this Stage. Enter the amount of NEO/GAS you wish to contribute. The built-in calculator shows you how many EFX (and bonus) tokens your contribution will yield. If you are satisfied with the amount shown, click ‘next’ to proceed. Step 4: sending your NEO/GAS This is the most important step. On this screen, you can see the amount of NEO/GAS you have chosen to contribute, as well as a NEO address. This is your unique deposit address. Send your NEO/GAS to this address only. Please note: the NEO address shown in these screenshots is for demonstration purposes only. Find your unique deposit address by logging into the Effect.AI Token Sale tomorrow. Open your NEO wallet. Most likely, this is the NEON wallet, but you can use other wallets as well. Just remember that you will receive your EFX Tokens on the public address you provided during the Whitelist registration. In the NEON wallet, make sure you are in the MainNet mode (shown in top right corner). Click ‘send.’ Here you can paste the unique deposit address you’ve just copied. Enter the chosen amount of NEO/GAS. Make sure it matches the chosen amount and does not go over the personal cap. If you want to send both NEO and GAS. Click ‘add recipient’ and select the second amount and type of currency you wish to send. When you’re done, check the box to agree and click ‘Send Assets.’ Congratulations, you have now successfully participated in the Token Sale! Important: you can follow the same steps for Stage 2 The steps explained above are exactly the same for Stage 1 and Stage 2. After the first 24 hours have elapsed, you can visit the same URL using your email address to login. Stage 2 will allow you to contribute again, except your personal cap is now 25000 (twenty-five thousand) euro. Tokens in Stage 2 are sold on a first come, first serve basis. Good to know A 10% bonus applies to both Stages of the Token Sale. For Stage 2, the bonus applies only to the first 2% of EFX tokens sold. If you have already contributed in Stage 2, but wish to increase your contribution, simply login again and follow the steps a second time. This is possible in both stages. Safety and security We ask that you remain vigilant with regards to possible scams. We will not ask you to transfer funds via means other than the Token Sale itself. If you do experience problems during the Token sale, contact support@effect.ai for assistance. You can also monitor the Effect.AI Telegram Announcement channel for important updates. The Effect.AI Team Effect.ai T.me/effectai Facebook.com/effectai Twitter.com/effectaix Github.com/effectai

Effect.AI Token Sale Tutorial

132

effect-ai-token-sale-tutorial-1bdb5ef6ba4

2018-04-12

2018-04-12 01:43:13

https://medium.com/s/story/effect-ai-token-sale-tutorial-1bdb5ef6ba4

false

811

null

Blockchain

blockchain

Blockchain

265,164

Effect.AI

null

470bf7e8d4d3

effectai

2,323

1

20,181,104

null

0

aws kinesis create-stream --stream-name "VPCFlowLogs" --shard-count 1 aws kinesis describe-stream --stream-name "VPCFlowLogs" cat > allowCloudWatchAccesstoKinesis.json { "Statement": { "Effect": "Allow", "Principal": { "Service": "logs.us-east-1.amazonaws.com" }, "Action": "sts:AssumeRole" } } aws iam create-role --role-name CloudWatchToKinesisRole --assume-role-policy-document file:///home/igor.kantor/bash/allowCloudWatchAccesstoKinesis.json { "Role": { "AssumeRolePolicyDocument": { "Statement": { "Action": "sts:AssumeRole", "Effect": "Allow", "Principal": { "Service": "logs.us-east-1.amazonaws.com" } } }, "RoleId": "AROAJLKEBEZGORPSXGMIA", "CreateDate": "2018-02-06T16:24:31.670Z", "RoleName": "CloudWatchToKinesisRole", "Path": "/", "Arn": "arn:aws:iam::31415926:role/CloudWatchToKinesisRole" } } cat > cloudWatchPermissions.json { "Statement": [ { "Effect": "Allow", "Action": "kinesis:PutRecord", "Resource": "arn:aws:kinesis:us-east-1:31415926:stream/VPCFlowLogs" }, { "Effect": "Allow", "Action": "iam:PassRole", "Resource": "arn:aws:iam::31415926:role/CloudWatchToKinesisRole" } ] } aws iam put-role-policy --role-name CloudWatchToKinesisRole --policy-name Permissions-Policy-For-CWL --policy-document file:///home/igor.kantor/bash/cloudWatchPermissions.json aws logs put-subscription-filter \ --log-group-name "VPCFlowLogs" \ --filter-name "VPCFlowLogsAllFilter" \ --filter-pattern "[version, account_id, interface_id, srcaddr != "-", dstaddr != "-", srcport != "-", dstport != "-", protocol, packets, bytes, start, end, action, log_status]" \ --destination-arn "arn:aws:kinesis:us-east-1:31415926:stream/VPCFlowLogs" \ --role-arn "arn:aws:iam::31415926:role/CloudWatchToKinesisRole" aws kinesis get-records --limit 10 --shard-iterator $(aws kinesis get-shard-iterator --stream-name VPCFlowLogs --shard-id shardId-000000000000 --shard-iterator-type TRIM_HORIZON | jq -r ."ShardIterator") | jq -r .Records[].Data | base64 -d | zcat aws kinesis get-records --limit 10 --shard-iterator $(aws kinesis get-shard-iterator --stream-name VPCFlowLogs --shard-id shardId-000000000000 --shard-iterator-type TRIM_HORIZON | jq -r ."ShardIterator") | jq -r .Records[].Data | base64 -d | zcat ... account_id":"31415926","interface_id":"eni-22222","log_status":"OK","bytes":"212","srcport":"58237","action":"ACCEPT","end":"1517935827"}},{"id":"33851098767603148713170041907970295510985708680010596642","timestamp":1517935767000,"message":"2 31415926 eni-8175917b 10.64.34.7 10.64.32.54 51143 2370 6 4 216 1517935767 1517935827 ACCEPT OK","extractedFields":{"srcaddr":"10.64.34.7","dstport":"2370","start":"1517935767","dstaddr":"10.64.32.54","version":"2","packets":"4","protocol":"6","account_id":"31415926","interface_id":"eni-22222","log_status":"OK","bytes":"216","srcport":"51143" ...

13

null

2018-02-06

2018-02-06 16:15:13

2018-02-12

2018-02-12 15:51:25

2

false

en

2018-02-13

2018-02-13 15:08:58

6

1bdd8a9426f1

3.681447

2

0

Kinesis Stream

5

Real-time Anomaly Detection in VPC Flow Logs, Part 3: Kinesis Stream Photo by Nathan Anderson on Unsplash To recap Part 2, we are setting up a pipeline to capture and stream all of our VPC Flow Logs, for long-term archival and immediate anomaly detection. In other words, we are building the stuff in red: a pipeline that accepts our VPC Flow Logs as input, analyses the stream for anomalies and produces an output that is an anomaly score. Let’s do this! First, we need to setup our Kinesis Stream. We use a shard count of 1. Shard count is a well-covered topic in AWS documentation. In essence, it’s a unit of streaming capacity: The data capacity of your stream is a function of the number of shards that you specify for the stream. The total capacity of the stream is the sum of the capacities of its shards. Make sure the Kinesis stream was successfully created: You should see something like this: A sample Kinesis describe output Create a JSON file for the IAM role. Note the us-east-1 in bold, substitute your region as needed. Ctrl+D to save the file. Next, let’s create a role with the JSON file from above: You should get this back (instead of 31415926 you’ll see your account number). Next, we need a policy to attach the role to (swap out 31415926 for your account number): Again, Ctrl+D to save. Now, let’s attach the role to the policy: That is all for the permissions. Next, we need to create a subscription filter to send the VPC Flow Log entries to our Kinesis stream. And we are done! Let’s check out our Kinesis stream to make sure data is actually flowing. Get ready for some mad jq-fu! Here’s a command to get you started: Let’s break this down. First, aws kinesis get-records command simply gets data records from a Kinesis data stream’s shard. It accepts a shard-iterator as a parameter. How do we get that? With the aws kinesis get-shard-iterator command, of course! That’s the command in parenthesis above. If you are not familiar with bash, the syntax of $(doStuff) means, run the doStuff command and return its value. However, the value we get back is actually JSON. So, we pipe the output to jq — a very powerful Linux JSON processor. Here, we are only interested in the ShardIterator property of the JSON payload. Therefore, we are telling jq to return the raw value (-r parameter) and grab the ShardIterator only and feed it back to the aws kinesis get-records command. So, at this point, we are here: And, surprise — that command also returns JSON. See the pattern here? Once again, we need to extract the Data property from the Records array. This jq does the trick: But the data we get back is base64 encoded and compressed to boot! No problem, we pipe the whole thing to these two handy utilities: The first one decodes and the second one uncompresses. Done! If it all works, you will get a massive output dump that will look something like this: Needless to say, this is not easy on the eyes and is quite obviously not meant for direct human consumption. However, the fields are well described and can be seen in the filter subscription above. You can also read the official Amazon documentation, if you so desire. That’s all for the Kinesis setup. Next, we create a Kinesis Analytics application that will read the records from the stream and (hopefully!) automatically detect anomalies in the stream. Machine learning for the win! Read on for Part 4 .

Real-time Anomaly Detection in VPC Flow Logs, Part 3: Kinesis Stream

7

real-time-anomaly-detection-in-vpc-flow-logs-part-3-kinesis-stream-1bdd8a9426f1

2018-06-01

2018-06-01 12:45:09

https://medium.com/s/story/real-time-anomaly-detection-in-vpc-flow-logs-part-3-kinesis-stream-1bdd8a9426f1

false

874

null

Anomaly Detection

anomaly-detection

Anomaly Detection

148

Igor Kantor

Part-time coder, full-time dad.

76057da89ce9

devfire

1,435

16

20,181,104

null

0

null

0

null

2018-05-22

2018-05-22 15:56:07

2018-05-22

2018-05-22 16:32:31

1

false

en

2018-05-22

2018-05-22 16:53:18

5

1bdfce1a125d

2.822642

2

0

It’s no big secret. Computers help us to optimize, transform and improve our overall processes. They make humans more efficient, so why not…

5

Reasonable Search & Seizure vs. Computer Vision It’s no big secret. Computers help us to optimize, transform and improve our overall processes. They make humans more efficient, so why not police? One of the most loved and influential things about America is the idea that every human should have rights. NPR posted an article today about Orlando police testing Amazon’s real time facial recognition. According to the 4th Amendment to the Constitute of the United States, searching for people in crowds using cameras could possibly violate this right. Why is this a big deal? We all want to find sex offenders, murderers, violent exes, and shooters, right? In theory, yes, this would be beneficial but would it be beneficial always, and always the right thing to do? A few things to explore. America’s “systems” are not set-up to make sure that every person has afair judgement against them which would include but not be limited to a fair chance to not commit crimes, have the proper levels of support indicated to be necessary to move people out of poverty (in schools and community centers) and to help them obtain and keep stable jobs (if you haven’t read Carl Hart Jr.’s High Price yet, please read it before responding). That all laws were ethical. In a mobile society, and where the 4th Amendment was set-up to protect someone so they could retreat into their own ‘home’, and where homes are considered cars, or wearable architecture, it isn’t clear whether or not our body or clothing should be considered a ‘home’ in which we have the right to privacy (I certainly think wearing clothing to protect my privacy is a strong case for this). In a mobile society where most of us ‘meet’ or have business meetings over our phones while walking and in the case of Silverman v. United States, “the Court unanimously held that a federal officer may not, without warrant, physically place themselves into the space of a person’s office or home to secretly observe or listen and relate at the man’s subsequent criminal trial what was seen or heard.” Suitaloon, considered wearable architecture. It would make sense that reasonable search and seizure would include someone who had recently committed a crime and was an immediate threat to others, so cameras were searched to find that person to prevent others from being hurt, but in cities like Ferguson, where arrest warrants are out for minor crimes, and people are removed from coffee shops because of the color of their skin are getting in trouble it does not seem it should be appropriate that cameras be used to find people to hold them accountable for any and all crimes. How should computer vision be used to improve society? Be used to find persons exhibiting unusual behavior (such as pulling out a gun or acting erratically). Be able to find people who are known to be and are in immediate danger. Be available to help prove the location of a person if a crime was under investigation in which they could be jailed, or lose significant rights as a result of the outcome. You have a choice in the future. Many of you who read this will be among the privileged who have the time and luxury to do so and will be less likely to be affected by these rulings. Would you speak for others so that we all can have a chance to work together to build a better world? What is the right thing to do? I’d love to hear your stance and if you believe constant public monitoring is ethical and reasonable. Disclosure: We first deeply explored this area when asked to investigate the use of Lumenora’s AR capable glasses as a tool for the police force. These are reflective of our findings but not those of a lawyer or of Lumenora, Inc. officially. We have made it known that any use of our glasses in such a situation would need to follow all applicable laws, including but not limited to there needing to be a reason to investigate for a crime when using this technology. This is subject to change in the future.

Reasonable Search & Seizure vs. Computer Vision

51

reasonable-search-seizure-vs-computer-vision-1bdfce1a125d

2018-07-06

2018-07-06 10:47:24

https://medium.com/s/story/reasonable-search-seizure-vs-computer-vision-1bdfce1a125d

false

695

null

Privacy

privacy

Privacy

23,226

Rosie Haft

Founder of Lumenora. Humanitarian, technologist, futurist, nowist, dancer, art lover and other things.

6b275c37e957

haftrm

195

23

20,181,104

null

0

null

0

c24a79900e57

2018-07-04

2018-07-04 14:20:08

2018-07-04

2018-07-04 14:23:36

1

false

th

2018-07-04

2018-07-04 14:23:36

0

1be2280ea27b

0.324528

0

พวกเรามีความตื้นตันใจที่จะประกาศว่า บริษัท Hearti ได้ร่วมมือทางการค้ากับบริษัท OPGtech เป็นที่เรียบร้อย ซึ่งบริษัทของเรา ทั้ง Hearti และ…

5

HEARTI ผนึกกำลัง OPGTECH ร่วมสร้างนวัตกรรมเทคโนโลยีการเกษตร พวกเรามีความตื้นตันใจที่จะประกาศว่า บริษัท Hearti ได้ร่วมมือทางการค้ากับบริษัท OPGtech เป็นที่เรียบร้อย ซึ่งบริษัทของเรา ทั้ง Hearti และ Opg Tech มีความรักและต้องการที่จะปรับกระบวนการทางธุรกิจให้เข้าสู่ระบบดิจิทัลทั้งในประเทศไทยและลาวให้มากยิ่งขึ้น พวกเรา Hearti พร้อมทั้ง บริษัทประกันที่ร่วมมือทางธุรกิจกับเรา จะขยาย micro-insurance ให้ครอบคลุมผู้ประกอบการทางการเกษตร และ ลูกจ้าง รวมถึงครอบครัว ซึ่งมีนับหลายพันคนของบริษัท OPGtech ในประเทศไทยและลาว ผ่านทาง SURETY.AI

HEARTI ผนึกกำลัง OPGTECH ร่วมสร้างนวัตกรรมเทคโนโลยีการเกษตร

0

hearti-ผนึกกำลัง-opgtech-ร่วมสร้างนวัตกรรมเทคโนโลยีการเกษตร-1be2280ea27b

2018-07-04

2018-07-04 14:23:36

https://medium.com/s/story/hearti-ผนึกกำลัง-opgtech-ร่วมสร้างนวัตกรรมเทคโนโลยีการเกษตร-1be2280ea27b

false

33

Your Next Generation Insurance Partner

null

HeartiThailand

null

Hearti Thailand

pakorn@hearti.io

heartithailand

FINANCIAL INCLUSION,INSURTECH,BLOCKCHAIN,CRYPTOCURRENCY,INITIAL COIN OFFERING

HeartiThailand

Financial Inclusion

financial-inclusion

Financial Inclusion

1,260

Pakorn Thienthong

null

7ed267de004b

porpakorn11

5

2

20,181,104

null

0

null

0

c5237204e9bb

2018-08-14

2018-08-14 13:24:23

2018-08-20

2018-08-20 14:14:32

3

false

en

2018-08-20

2018-08-20 20:38:19

55

1be359dac33c

9.546226

16

0

The first Pangeo Developers Meeting was held at NCAR in Boulder, Colorado from Aug. 13–15. The purpose of this meeting was to allow our…

5

The 2018 Pangeo Developers Workshop The first Pangeo Developers Meeting was held at NCAR in Boulder, Colorado from Aug. 13–15. The purpose of this meeting was to allow our highly distributed group of contributors to talk face to face, brainstorm ideas for the future of Pangeo, and address pressing technical challenges head on. Although the workshop was deliberately small (30 people; see full list below), we had attendees from as far as the UK, France, and Australia! The workshop consisted of a few talks by various Pangeo contributors as well as plenty of free, unstructured time for code sprints and discussion. Below I sum up some of the most exciting outcomes of the meeting. This is a large dump of information — we plan to have some more specialized blog posts on some of the topics mentioned below in the near future. Pangeo + Jupyter Pangeo leans heavily on Jupyter for both its user interface (Jupyter Lab and Notebook) and its cloud architecture (based on zero2jupyterhub-k8s). So we were thrilled to have three members of the Berkeley Jupyter team attend the workshop: Ian Rose, a geophysicist and UI expert on the Jupyter Lab project; Yuvi Panda, devops guru and mastermind of all things cloud; and Fernando Perez himself, the creator and BDFL of the Jupyter project. Over the course of the workshop, the three people lent their unique knowledge experience to various ongoing efforts within Pangeo (detailed below). Overall we were excited to learn that Fernando sees Earth Science as a high priority for his own future research and development efforts; we are looking forward to collaborating even more closely with the Project Jupyter! Analyzing Climate Model Data on Cheyenne Being at NCAR, we were excited to advance Pangeo’s capabilities for analyzing climate model data living on Cheyenne, NCAR’s flagship supercomputer. Matt Long, Mike Levy, and Gustavo Marques all worked on various aspects of this problem. Matt and Mike worked on making it easy to launch Jupyter notebooks on Cheyenne by developing some custom launch scripts. (There was much discussion over how we could get CISL to provide more official support for these tools, which would make the process a lot smoother. The excellent Jupyter support at NERSC looks like a good goal to strive for.) They developed a simple example notebook for loading and analyzing data from the CESM Large Ensemble Project in parallel using dask jobqueue. Gustavo instead focused on exporting data from the new MOM6 ocean model into zarr format (using dask jobqueue on Cheyenne) and uploading it to Google Cloud Storage. We were all impressed by how quickly Gustavo spun up on the Pangeo stack. We now have a new example notebook for MOM6 which runs on pangeo.pydata.org! Just log on and try it out. Enhancing the Pangeo Cloud Experience In addition to our work on traditional HPCs like Cheyenne, Pangeo has been increasingly focused on using the commercial cloud for large-scale scientific data analysis. The cloud offers many technical and social advantages over traditional HPC, including easy user access, access to scalable object storage, and the ability to scale quickly to large numbers of compute nodes. Pangeo’s experimental cloud service — pangeo.pydata.org — has been running since March. Although the service is public, it is not exactly “production ready” — our goal has been to learn about how to work with the cloud, and we have learned a ton so far! Number of daily logins on pangeo.pydata.org. This workshop gave us an opportunity to start upgrading the Pangeo cloud experience. A central goal is to provide more diagnostic information about the cluster to users. Building on prototypes he developed for the UK Met Office Pangeo cluster, Jacob Tomlinson worked on expanding Grafana-based real-time monitoring of cluster usage statistics. He has developed amazing dashboard which provides all sorts of useful information (see pangeo-data/pangeo#359). The next step is to integrate this directly into Jupyter Lab as an extension. Pangeo Grafana dashboard by Jacob Tomlinson Speaking of extensions, Ian Rose worked on a killer feature that users clearly want: integration of the dask dashboard directly into Jupyter Lab. We already have the ability to launch dask clusters interactively from notebooks thanks to tools like dask-jobqueue and dask-kubernetes, but the dashboards from these clusters appear in another browser window. This extension would be a game changer in terms of creating interactive users experiences within Jupyter lab itself. Ian’s preliminary work is on GitHub. Finally, Yuvi Panda continued his heroic efforts to democratize access to cloud computing for scientists, from which Pangeo has already benefitted immensely. Yuvi showed off The Littlest JupyterHub, a project which makes it dead simple to deploy a simple Jupyter Hub in a range of different circumstances. We brainstormed someideas for what the “Littlest Pangeo” might look like. He also worked on HubPloy, a tool to simplify the deployment of Kubernetes-based Jupyter Hubs. HubPloy would solve lots of the headaches Pangeo is facing in managing cloud clusters. In our discussions, Yuvi shared his thoughts about how projects like Pangeo can maintain freedom and independence from the cloud computing giants by designing their platform in a way that avoid vendor lock in. He summarized his advice in a blog post. (See also a related post by Matthew Rocklin.) This post wouldn’t be complete without an acknowledgement of the important role played by Matt, who generally spent his time at the workshop bouncing around and lending his expertise wherever it was needed. Beyond his deep technical knowledge (Matt is the creator of Dask, a key part of the Pangeo platform), he is great at communicating with people of all backgrounds and keeping technical meetings focused and productive. Thanks Matt! Pangeo + Binder The Binder Project allows you to turn any Jupyter notebook stored in a GitHub repo into an actually running notebook (with all its dependencies) in the cloud; this technology is a quantum leap for scientific reproducibility. Binder 2.0, a Tech Guide Authors: The Binder project is comprised of many individuals within and outside of the core Jupyter team. A list of…blog.jupyter.org By default, binder launches these notebooks into a dedicated Jupyter Hub running on Google Cloud, but we would like the ability to launch directly into a Pangeo cloud deployment. This would enable the binder notebook to take advantage of dask-based parallelism (via dask kubernetes) and access any cloud-based data stores associated with that cluster. Joe Hamman is pushing hard to make this a reality. With help from the Jupyter folks, he has developed a prototype solution up on GitHub as a customized helm chart. We should soon expect to see a test Pangeo binder service up and running! We hope this efforts will benefit the broader scientific community by expanding the flexibility of the binder service. A longer blog post on this will be up very soon! Pangeo + Cloud Optimized Geotiff The GIS community has done a great job defining a format for “analysis ready data”: the Cloud Optimized Geotiff (COG). Pangeo aims to make it easy for scientists to use cloud computing for interactive, parallel analysis. So this is potentially a match made in heaven. Dan Rothenberg and Scott Henderson worked hard on applying Pangeo cloud deployments to process COGs stored in Google Cloud Storage and Amazon S3. Dan focused on improving xarray-COG integration by developing automatic chunk-size detection for geotiffs read by rasterio (see pull requests pydata/xarray#2255 and dask/dask#3878). Meanwhile, Scott developed a very cool example notebook showing off how quickly and efficiently xarray + dask + rasterio can pull and analyze data from AWS COGs in parallel. There is lots of potential here. Expect a more detailed blog post on this work in the near future! Cloud Ready Data In a recent post, I laid out a (slightly tongue-in-cheek) vision for cloud-native data repositories: Step-by-Step Guide to Building a Big Data Portal The volume of scientific datasets is growing at an exponential rate, and scientists are struggling to keep up. This Big…medium.com As described clearly in a blog post by Matthew Rocklin, the science community currently lacks a good standard for cloud-optimized NetCDF / HDF type data. At the developers workshop, I summed up our positive experience so far with zarr, a promising new format storage format for chunked, compressed multidimensional numeric arrays. But I noted that zarr was still far from a community standard. So I was pleasantly surprised to learn from Ward Fisher, NetCDF Team Lead, Lead Developer at Unidata, that the NetCDF group is considering making zarr the back-end for it’s next generation NetCDF library. (The other contender is TileDB, and it’s possible we will see support for both libraries within NetCDF.) This would truly be a game changer, allowing data providers to host NetCDF data in the cloud in a way that is simultaneously archive-quality and analysis-ready. We had a useful discussion of the pros and cons of zarr vs. TileDB. Overall I’m satisfied to see that Pangeo has managed to influence this important discussion. Our platform aims to unlock the untapped scientific value contained in large datasets, so making it easy to discover and load data is a crucial need. During the code sprints, several of us took on the problem of how to build data catalogs for Pangeo. To this end, Rich Signell of USGS and Andrew Pawloski of Element84 worked on connecting Pangeo cloud deployments to OGC metadata services. Their example notebooks shows how it’s possible to search for data from existing catalogs, connect to remote opendap or buckets with zarr datasets, then load and process the data in a streaming fashion using dask. Focusing instead on our newly created zarr cloud datastores, I worked a bit on integrating Martin Durant’s new Intake data catalog tool with Pangeo. Intake makes it easy to expose data sources to users from within python and could help remove some friction from the data discovery / loading process. We are especially excited about the possibility of integrating these tools with Jupyter Lab, allowing users to browse and load data visually. To this end, Ian Rose helped us get started on developing custom Jupyter Lab Extensions. Finally, Nick Mortimer of CSIRO worked on transcoding ARGO float data into zarr and uploading to cloud storage. His goal is to make it easier to process this extremely rich dataset en masse in order to train machine learning models. Connecting with Legacy Geoscience Libraries Pangeo aims to build a platform for the future of data-driven geoscience research, based on the modern scientific python ecosystem. But there is a huge amount of knowledge embodied by legacy geoscience software that is not easily accessible to this ecosystem. To this end, we were thrilled to have the participation of Ben Koziol of NOAA ESRL — an expert in the Earth System Modeling Framework — and Mary Haley of NCAR — leader of the NCL team. Ben is developing a proof of concept notebook for wrapping ESMF routines with dask.delayed. The success of Jiawei Zhuang’s xesmf library shows there is great appetite for ESMF-based tools that integrate well with the python ecosystem. Mary spent a long time talking to Fernando Perez about how to best integrate NCL’s powerful computational and visualization routines with python. We are excited to see where this leads. Pangeo Governance Pangeo started out very organically, but as the project grows, we see a need for a more formal governance structure. Having some formal governance will help us develop the project in a more focused way, coordinate our relationships with other organizations, pursue funding more effectively, and ensure wider participation and diversity. Borrowing heavily from the Project Jupyter, we developed an official Governance Repo. The basic idea is that, like many other open source projects, Pangeo will be governed by a steering council consisting of active project contributors. The details are still under discussion, and we welcome feedback from the broader community. A particular priority for the steering council will be to figure out how to make Pangeo a more diverse community, which our organic process so far has not managed to achieve. A starting point for this is our new official Code of Conduct, which aims to make Pangeo as welcoming as possible to people of all backgrounds. Partial group photo from the 2018 Pangeo Developers Workshop. Workshop Talks: Day 1 Getting Started 9:00–9:15: Welcome and Logistics — Kevin Paul 9:15–9:30: Introductions 9:30–10:30: The evolution of Pangeo — Ryan Abernathey 10:30–11:00: The Pangeo Principles — Jacob Tomlinson Data Proximate Science Gateways 11:25–11:30: Intro — Kevin Paul 11:30–11:55: Jupyter Team — Yuvi Panda and Ian Rose 11:55–12:30: Discussion / more talks Science Use Cases 2:00–2:35: Why I’m interested in Pangeo as a scientist — Matt Long 2:35–2:55: Moving satellite radar processing and analysis to the Cloud — Scott Henderson 2:55–3:15: Notebook Examples / Discussion — Rich Signell et al. Analysis-Ready Data Formats 3:35–3:55: My Experience with Storing Xarray Datasets in the Cloud using Zarr — Ryan Abernathey 3:55–4:15: NetCDF’s plans for cloud-based data — Ward Fisher Workshop Talks: Day 2 Software Ecosystems 9:00–9:25: Summary of relevant Dask updates and challenges today — Matt Rocklin 9:25–9:45: Updates to MetPy based on Xarray — Ryan May 9:45–10:05: ESMF, OpenClimateGIS, and the Birdhouse WPS Stack: Connecting to Pangeo — Ben Koziol 10:05–10:25: NCL & Pangeo — Mary Haley Federation and Sustainability 11:20–11:40: Pangeo and NASA’s Cloud Hosted Earth Observing System Data — Joe Hamman 11:40–12:00: Discussion around federation across multiple cloud / hpc platforms 1:30–2:15: Discussion of outreach efforts, training (tutorials), and diversity efforts — Kevin Paul Attendee List Adekunle Ajayi, Institut des Géosciences de l’Environnement, Université Grenoble Alpes, France Andrew Pawloski, Element 84 Aurélie Albert, Université Grenoble Alpes- IGE Ben Koziol, NESII/CIRES/NOAA-ESRL Bill Ladwig, NCAR Chiara Lepore, Lamont Doherty Earth Observatory of Columbia University Daniel Rothenberg,ClimaCell Fernando Perez, UC Berkeley, Gustavo Marques, NCAR Ian Rose, UC Berkeley (Project Jupyter/Earth and Planetary Science) Jacob Tomlinson, Met Office Jeff de La Beaujardiere, NCAR/CISL/ISD John Allen, Central Michigan University John Exby, Jupiter Joseph Hamman, National Center for Atmospheric Research Kevin Hallock, NCAR Kevin Paul, NCAR Luke Madaus, Jupiter Matthew Long, NCAR Matthew Rocklin, Anaconda Inc Michael Levy, NCAR Niall Robinson, Met Office Informatics Lab Nick Mortimer, CSIRO Rich Signell, USGS Rick Brownrigg, NCAR Ryan Abernathey, Columbia University / Lamont Doherty Earth Observatory Ryan May, UCAR/Unidata Scott Henderson, University of Washington Yuvi Panda, UC Berkeley / Project Jupyter

The 2018 Pangeo Developers Workshop

34

the-2018-pangeo-developers-workshop-1be359dac33c

2018-08-20

2018-08-20 20:38:19

https://medium.com/s/story/the-2018-pangeo-developers-workshop-1be359dac33c

false

2,384

A community platform for Big Data geoscience

null

pangeo

pangeo@googlegroups.com

pangeo

DATA SCIENCE,BIG DATA,EARTH SCIENCE

pangeo_data

Data Science

data-science

Data Science

33,617

Ryan Abernathey

Assistant Professor, Earth & Environmental Sciences, Columbia University

ae28b16eab24

rabernat

87

98

20,181,104

null

0

null

0

427737f7fbc1

2018-03-29

2018-03-29 21:57:13

2018-03-29

2018-03-29 22:12:24

3

false

en

2018-03-29

2018-03-29 22:12:24

6

1be431a4658

2.183962

0

Google’s weapon-of-choice in assimilating each industry and its revenues is the “Knowledge Graph.” How is it built? Why it is so powerful?

5

Two Words: Knowledge Graph [2/6] Google’s weapon-of-choice in assimilating each industry and its revenues is the “Knowledge Graph.” How is it built? Why it is so powerful? We explain using the Hotel and Apartment Knowledge Graph. How does Google organize industry data? And, how does it shift revenues from the industry ecosystem to itself? The answer is in the deceptively simple, but awesomely powerful Knowledge Graph. Let’s look at the Hotel and the Apartment Knowledge Graph to unpack this technology. (This is the 2nd post in the Google vs Multifamily ILS series. In the 1st post, we discussed how Google organized travel information and captured more revenue than the largest player in the travel industry.) The Hotel Knowledge Graph Here is Google’s knowledge panel for a hotel visible on the right of a search results page: What did it take to organize the Hotel Knowledge Graph? First, Google collected information from many different sources: Some of the basics come from spidering the hotel’s website Other basics come from information volunteered by the hotel’s owner using Google My Business Photos come from customers that have visited the hotel Photos could also come directly from the owners themselves Finally, reviews have come from third party sites And, then, Google linked all of the above to create the knowledge graph for a single hotel. Why is the Knowledge Graph so much more powerful than other search results? First, the knowledge graph is not a link to other sites. Second, its goal is to directly answer the most important questions without having to click through to multiple sites. As Google’s Amit Singhal said: If people are searching for “2+2”, why shouldn’t Google give a direct answer to that versus sending searchers to a site? Third, it has a built-in call-to-action that allows the user to initiate a transaction from Google. In this case, booking a room. As a result, users are able to do all their research and make their decisions without leaving Google. In contrast, the online travel agency (OTA) is left out of the loop. At best, it is left to make the actual reservation. At worst, the reservation is made directly on the hotel’s site itself. The Apartment Knowledge Graph How close is Google to organizing the multifamily apartment knowledge graph? Look at it for yourself: Pretty close to the Hotel Knowledge Graph, right? In the next post, we will compare the Apartment Knowledge Graph and the Hotel Knowledge graphs in detail. Liked this? Get more multifamily technology explainers by signing up for Multifamily Minute! Originally published at hy.ly.

Two Words: Knowledge Graph [2/6]

0

two-words-knowledge-graph-2-6-1be431a4658

2018-03-29

2018-03-29 22:12:25

https://medium.com/s/story/two-words-knowledge-graph-2-6-1be431a4658

false

433

Technology is changing Multifamily Marketing & Operations. Stay informed of the latest in bite-sized updates.

null

Multifamily Minute

marketing@hy.ly

multifamily-minute

MULTIFAMILY,APARTMENTS,TECHNO,MARKETING

hellohyly

Google

google

Google

35,754

Munish Gandhi

Founder & CEO, Hy.ly

23021255f4a

munishgandhi

224

11

20,181,104

null

0

null

0

dc3468e1f99e

2018-03-20

2018-03-20 20:41:12

2018-03-20

2018-03-20 20:41:47

1

false

en

2018-03-20

2018-03-20 20:41:47

0

1be55a0b147

1.245283

0

Speaking in the most general sense artificial intelligence is a technology for creating algorithms and programs, their subsequent training…

1

How will our AI work? Speaking in the most general sense artificial intelligence is a technology for creating algorithms and programs, their subsequent training and use for certain purposes. A computer receives information, processes it, reveals certain patterns and applies the knowledge. One of the main differences between artificial intelligence and conventional algorithms is the ability to learn. There are different technologies for creating artificial intelligence. One of the most known is the creation of AI based on principle of human brain neural network. Neurons are brain cells. Connections between neurons form as a result of information received from the outside. For example, a person sees an object and receives information “this is a house”. The next time a person sees an object like this and connection between neurons will define an object as a house. In artificial intelligence computational elements work in the same way. How does artificial intelligence learn? Learning AI can be compared with the way a small child learns to read. Child is shown a sign and he is told “This is letter A”. Then a baby is taught to compose words from letters and sentences from words. An artificial brain receives a certain amount of data such as an image database. AI studies images, reveals patterns in shape, color, position of elements. Being learned in this way AI subsequently is able to recognize images shown to it. Similarly, AI can recognize text, voice. But amount of data is very important. The more the better. A computer brain will have more opportunities for learning and correcting data processing algorithms. With sufficient data amount the result produced by artificial intelligence will be more correct.

How will our AI work?

0

how-will-our-ai-work-1be55a0b147

2018-03-20

2018-03-20 20:41:48

https://medium.com/s/story/how-will-our-ai-work-1be55a0b147

false

277

TFH AI Ratings: innovation in data audit based on artificial intelligence

null

TFH-AI-Ratings-218299148797072

null

TFH AI Ratings

db@tfh-ai.com

tfh-ratings

null

tfhairatings

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

TFH AI Ratings

null

48d03229ace4

1983irina.vish

0

1

20,181,104

null

0

null

0

null

2018-06-07

2018-06-07 10:09:02

2018-06-10

2018-06-10 17:51:43

1

false

en

2018-06-10

2018-06-10 17:51:43

11

1be5f11378bf

5.132075

1

0

In my previous article I likened the launch of Google Duplex with that of the first iPhone back in 2007. The wave that had followed the…

4

How will Google Duplex change our world? Part 1: Shifted Goal Posts. In my previous article I likened the launch of Google Duplex with that of the first iPhone back in 2007. The wave that had followed the iPhone launch ended up drastically changing our lives, with smartphones now as common as pets (or maybe more so). If Google Duplex is indeed as revolutionary as it seems, what are we to expect in the coming future? Photo by Thomas Serer on Unsplash Siri, Alexa, Cortana may well be human-like names, but adding the name of a fourth colleague, with the uncryptic name of Google Assistant, unveils the true character behind them: these are technology giants’ agents that we have adopted so serenely in our homes and pockets. The list of assistants is longer than most are aware of. In China, technology giants Baidu and Alibaba did not shy out from their western counterparts and have developed their own Duer and Tmall Genie respectively. Also from Asia we have Samsung’s Bixby; and Sony have started baking their own assistant too. After the Google Duplex teaser, I started to become curious as to what will happen to this space. Building on top of the link I drew between Google Duplex as the “new” iPhone for virtual assistants, the question becomes simpler: Who will be the Nokias and Blackberrys of virtual assistants? Just a refresher for those of you who missed the past decade, Nokia and Blackberry were two companies with major market shares in the mobile phone space. Both Nokia and Blackberry were in the forefront to produce mobile phones with Internet and email capabilities, which when paired up with their phone’s physical reliability (back then mobile phone batteries did last longer than 2 days) made them the natural option for businessmen and common people alike. But soon after the launch of the iPhone, both companies struggled to the point of losing practically all of their market share, for very particular reasons. Blackberry was extremely proud of its iconic mobile design, and resisted to the change iPhone brought, that is, keyless and button less mobiles. Nokia, on the other hand, got cold feet when it came to embracing leading operating systems and pursued with its Symbian to a point when it ran out of options and finally partnered with Microsoft. This marriage was far from fruitful, and nowadays both Microsoft and Nokia are rare species in the world of mobile phones. The mistakes by Blackberry and Nokia can be easily repeated by any company in any competitive market. Blackberry’s mistake was resistance to change, embracing tradition over progress, which kept the company as an icon of the past (which unfortunately doesn’t generate any revenue, unless for museum tourism). Nokia’s mistakes were two: first choosing build in the build-vs-buy dilemma, and ignoring its real forte which was hardware; and second was a wrong partnership decision. Back to the virtual assistant world, the giants we know today need to be very careful if they want to survive the test of history. Here is what I believe are thoughts which may help companies remain relevant in the space for at least another decade. Real World Applications For virtual assistants to survive they need to start doing more than telling you the time or read the weather forecast for the next day, week or month. That was one of the reasons Google Duplex was a breakthrough: with Google Duplex, Google Assistant can become a personal secretary. Many people already rely on Google applications in their day to day activities: Google Drive for collaboration and keeping records, Google Calendar for scheduling appointments, Google Maps for navigation, and so on. Google users with some imagination can already start seeing practical applications for the Duplex technology. What about Apple, Amazon, and the rest? The giants need to move away from their strongholds. We already know that most of the big players have vested interest in healthcare. Apple and Amazon have already shown movements in the sector. While diagnosing the healthcare market segment as a launching pad for AI disruption, there is a good news and a bad news. The good news is that healthcare is an evergreen market. People will always need better healthcare, no matter how good it already is, and technology has proved itself in the past that it can really make a difference there. The bad news is that the space tends to get crowded pretty quickly. As the linked article shows, Google are already extremely active in healthcare and generally the space tends to be susceptible to hype. So for any company to succeed it needs to look beyond healthcare to make real money. So, one would ask, where to go? The space for innovation is huge, we just need to read the right signals. In China, Alibaba have started looking at the car market. This is no groundbreaking innovation, true, as in the West we already got used to car makers including virtual assistants in the new models. What makes Alibaba’s move interesting, however, are two things which make most sense when evaluated together. First of all, as I’ll discuss in deeper detail in a dedicated future article, Chinese companies sit on the biggest data set in the whole world. Data, as many will know, is the fuel for better AI. Alibaba’s service through their virtual assistants, therefore, might eventually start offering a vaster (or at least better) range of applications and assistance. Secondly is Alibaba’s wise strategic move. Despite residing behind the Great Wall, the Chinese technology giant has chosen to partner with some of the West’s best car makers: Mercedes-Benz, Audi, Volvo are well known in China but also extremely respected in the West. If Alibaba proves itself to be a superior AI service provider for these carmakers, there is nothing to stop them (the car makers) to bring them over to the West. Revolutionary Mindset If there is anything we learned from the smartphone market it is the increased sense of openness which became the de facto mindset for the technology world. Open APIs, free apps, customizable OS, were key to foster interest among the technology lovers to innovate, contribute and adopt. Developer platforms and code libraries stopped being exclusive property of the technology company, and open source (or at least, the philosophy) made the ground fertile for innovation. Just like data is the fuel for AI, developers are what make any new technology relevant. For any initiative to be successful it needs to be developer-centric, and perhaps this is what was missing from the Google Duplex launch. Following the grandiose launch, I for one went to search for any available, even in alpha-mode, APIs, to see what can be done or how can I play around with it. To my disappointment, and I believe to the disappointment of many, we discovered that this was just a demonstration which shall be substantiated further more (perhaps) later this year. “Why all this hurry then?”, one would ask. But that’s a different argument at a tangent. A more relevant question for this article will be: “What kind of revolutionary mindset should we expect?”. Companies, for one, need to start to cooperate and share more. All the technology giants mentioned in this article already possess a tremendous (and scary) amount of data, which gives them plenty of “capital” to build effective AI applications. No matter how big their databases are, however, they will never hold all of the world’s information. Data held in a secure fortress is doomed to always be incomplete, and its owner will be held as hostage from breaking through the market. This will mean that companies, not only need to give access to their data but doing so by setting their infrastructures more open for different integrations. Empowering the developer community will be crucial for their survival.

How will Google Duplex change our world? Part 1: Shifted Goal Posts.

1

how-will-google-duplex-change-our-world-part-1-shifted-goal-posts-1be5f11378bf

2018-06-11

2018-06-11 11:54:49

https://medium.com/s/story/how-will-google-duplex-change-our-world-part-1-shifted-goal-posts-1be5f11378bf

false

1,307

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Matthew Croker

A blockchain of Athlete-Team Lead-Software Developer-Creative Strategist-…

1f527f58f042

matthewcroker

2

4

20,181,104

null

0

null

0

null

2018-07-04

2018-07-04 15:49:20

2018-09-12

2018-09-12 05:50:45

7

false

en

2018-09-12

2018-09-12 18:31:43

21

1be70cf9737f

7.034906

13

0

Which One Should I Choose?

5

Deep Learning MOOCs With deep learning becoming more accessible to developers, new courses and programs are being launched at an unprecedented pace. There are abundant resources for studying deep learning. Deep Learning MOOCs: Udacity, Coursera, Stanford, MIT, Fast.ai, MLCC I get this question a lot: “Which online deep learning program should I choose?” My answer is: “ It depends.” There are so many choices nowadays, which one to choose depends on your learning style, time or money you can afford, and your career goals. I’m going to share my perspective on a few of the well-known deep learning MOOCs: Udacity — Deep Learning Nano degree. Coursera — Deep Learning Specialization. Stanford —CS20 (TensorFlow for Deep Learning Research), CS230 (Deep Learning), CS231n (CNN for Visual Recognition) & CS224n (NLP with Deep Learning). MIT — Introduction to Deep Learning. Fast.ai — Deep Learning for Coders. Google — Machine Learning Crash Course with TensorFlow APIs. Udacity Deep Learning Nano degree Image source: Udacity.com Link to program | Cost: $999 or $84 per month Pace: assigned to a cohort with other students. This program is fairly challenging so be prepared to spend at least 10 to 15 hours per week. You are expected to complete the program within 4 months with 1 month extension. If you don’t finish by the deadline then you will be dropped out. Then you will need to pay the full price again to re-enroll. Where to go for help? Slack group and channels for each topic. Office hours available as well. Mentor — each student used to get assigned a mentor. Now instead you use study group. Study Group — a chat system inside of the Udacity portal. replaced the mentor system. Discussion forums—where students can help each other out with their questions. Course Structure: the nanodegree is organized into 6 parts, each with a few lessons: Intro to Deep Learning Neural Networks Convolutional Networks Recurrent Networks Generative Adversarial Networks Deep Reinforcement Learning Lectures were taught with both TensorFlow and Keras. There are several projects written in low level TensorFlow. Students are not expected to do the data pre-recessing and will focus on model architecture. Students used to need to create conda environments themselves and train on GPU on either AWS or FloydHub. Now the nanodegree provides environments with enough free GPU hours to complete the projects. Students who graduated will have guaranteed admission into the Self-Driving Car and Flying Car nanodegree programs. Other notes: The Udacity lectures are taught by several top notch instructors with quality content. The nano degree is full of cutting edge information in deep learning. It also offers career assistance. The deep learning is one of several nanodegrees from Udacity School of AI. You can choose to continue with another nanodegree such as Artificial Intelligence, Robotics, or Reinforcement Learning after finishing this one. Coursera Deep Learning Specialization Image source: coursera.org Link to program | Cost: $49/month Pace: courses are organized by weeks so you don’t need to spend extra time planning. If you fall behind, just switch to the next month co-hort. It’s a very flexible program and fairly affordable. Where to go for help? Discussion forums are the only way to find help if you get stuck. This program is very well structured. Many challenges you encountered are perhaps tools issues or some concepts which you can get answers from the forum discussions. Program structure There are 5 courses in this specialization: Intro to neural networks Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization How to structure your deep learning project Convolutional Neural Networks (CNN) Sequence Models Each course is organized by weeks, and each week has video lectures with transcripts, followed by a quiz then one or more programming exercises. Each programming exercise is well written with detailed guidance on what the student is supposed to do. There are also guest speakers for each week so you get perspective from the industry. Other notes: Dr. Andrew Ng is an excellent professor and he is great at explaining complex concept in a way that is easy to understand. The program provides a solid foundation to deep learning. Stanford Deep Learning Coureses Image source: stanford.edu There are several courses from Stanford on deep learning so I will not cover each of them in details. Rather I’m listing out the ones that I know of: CS20 (TensorFlow for Deep Learning Research) — great for learning TensorFlow, as well as an intro to deep learning. You will learn topics such as TensorFlow, Keras, CNN, RNN, GANs and Reinforcement learning. CS230 (Deep Learning) — this course is pretty much the same as the Coursera one above, but without the videos. CS231n (CNN for Visual Recognition)- great introduction to computer vision with deep learning. CS224n (NLP with Deep Learning)- focuses on natural language processing with deep learning. All of the slides and notes are free and accessibly to the public with videos from previous years available. MIT Intro to Deep Learning Image source: introtodeeplearning.com Link to program | Cost: free Pace: self-paced. Program structure: There are five sessions with two videos per session: Two videos: Intro to Deep Learning & Deep Sequence Modeling. Lab: Intro to TensorFlow + Music Generation with RNNs. Two videos: Deep Computer Vision & Deep Generative Models. Lab: Disease Detection from Human X-Ray Scans. Two videos: Deep Reinforcement Learning & Limitations and New Frontiers. Guest lectures from Google and NVIDIA. Guest lectures from IBM and Tencent. The lectures are well organized and in top quality. This is a good fit if you would like to go through various deep learning topics in a short period of time. I also find guest lectures from industry very interesting. Fast.ai Deep Learning for Coders Image source: fast.ai Link to program | Cost: free Pace: self paced. Be prepared to spend 1 to 3 weeks per lesson depending on how comfortable you are with the topics covered in the lesson, assuming you study 10 to 15 hours per week. Where to go for help? Wiki page — one of the things that I really like about fast.ai are the wiki pages for each lesson. I included link to lesson 1 wiki as an example. As you see there are tons of learning resources with links to articles, research papers and student notes which are amazing and can help you to better understand the videos. Discussion forum — head over to http://forums.fast.ai/ if you get stuck or have questions. The forum has students worldwide actively participating. Course Structure: The program is divided into two parts with 7 lessons in each. Each lesson consists of a 2-hour video and a wiki page. In each video Jeremy goes over code in Jupyter Notebooks and deep learning concepts. Since the lectures were recorded for in-class lessons then later on converted to online MOOC, the Jupyter notebook numbers may not match up with the lesson number. Also depending on your background knowledge, you may need supplement learning materials in order to fully understand the videos. Other notes: Fast.ai is unique and unconventional. It’s very different from the other programs in these ways: It adopts a teaching style of top-down approach focusing on hands-on experience instead of theories. The lectures jump back and forth a bit between CNN and RNN, unlike the other programs where the students learn CNN and then RNN. Sample code uses its own fast.ai library written on top of PyTorch. It has its own blog post page with great information. You may need to figure out your environment setup for training on GPU which resembles the challenges you may face in the real world. (Note: I was able to run the Part 1 notebooks in Colab with additional code for installing packages & downloading datasets). The program focuses on optimization techniques so the accuracy of the models are much higher than what you will see from the other MOOCs. You will find lots of fast.ai students on Kaggle, showcasing the state of art techniques they learned from the program. Protip: if you lack the basics on ML or deep learning, you may find this curriculum a bit challenging to keep up with. Fast.ai does have intro to ML videos. Also since this MOOC is so different from the others, you may benefit from taking another deep learning MOOC before or after completing the fast.ai one. Google ML Crash Course with TensorFlow API Image source: Google MLCC website Link to program | Cost: free Pace: self paced. Where to go for help? On meetup.com search for a Google Developer Group (GDG) near you and see if they are running a Study Jam. You can join the Study Jam to attend in person sessions and to study with others online together. Other notes: This program was used internally to train thousands of Google engineers and then open sourced early 2018 to the public. All concepts are well explained with visualization to help students better understand. For example, take a look at the explanation of what is learning rate and the exercise on how to optimize learning rate. It’s excellent at teaching the machine learning basics before moving on to neural networks and deep learning. It also has a very helpful pre-requisites page and a ML glossy page. Protip: this MOOC uses low level TensorFlow which is great for studying fundamentals. You should also check out the Keras guide and tutorials on TensorFlow.org to get familiar with the modern TensorFlow high level APIs. All the tutorials can be opened and run in Google Colab with free GPU. There are many other excellent online learning resources such as Siraji Raval’s School or AI, or Kaggle’s learn.

Deep Learning MOOCs

34

deep-learning-moocs-1be70cf9737f

2018-09-12

2018-09-12 18:31:43

https://medium.com/s/story/deep-learning-moocs-1be70cf9737f

false

1,586

null

Machine Learning

machine-learning

Machine Learning

51,320

Margaret Maynard-Reid

Google Developer Expert for ML | TensorFlow & Android

ed2f34822130

margaretmz

1,101

168

20,181,104

null

0

null

0

null

2018-07-23

2018-07-23 12:26:37

2018-07-23

2018-07-23 14:01:08

2

true

en

2018-07-23

2018-07-23 14:35:10

17

1be7322886ab

8.288994

5

2

0

Summary of Reinforcement Learning at ICLR 2018

5

A Digest of Reinforcement Learning Papers from ICLR 2018 Summary of Reinforcement Learning at ICLR 2018 The Sixth International Conference on Learning Representations (ICLR) 2018 was held at Vancouver from Monday April 30 — Thursday May 03, 2018. The Sixth International Conference on Learning Representations (ICLR) 2018 There were 337 out of 935 paper submissions accepted. Best paper awards went to On the convergence of Adam and Beyond, Spherical CNNs and Continuous Adaptation via Meta-learning in Nonstationary and Competitive Environments. Statistics RL Categories Conclusion Deep Reinforcement Learning (RL) is one of the hottest topics at ICLR this year. To get a quick overview, I categorize all RL papers accepted into following topics: RL Theory RL Algorithms RL Network Architecture RL Optimization RL Exploration RL Reward Distributed RL Hierarchical RL Multi-Agent RL Meta-learning, Transfer, Continuing Learning RL Applications Statistics The number of papers in each category is illustrated below: Next is the detailed list. The format is title (review rating) -> TL;DR. Oral presentations are marked with (oral) in addition. RL Theory RESIDUAL LOSS PREDICTION: REINFORCEMENT LEARNING WITH NO INCREMENTAL FEEDBACK (top 22%) -> a novel algorithm for solving reinforcement learning and bandit structured prediction problems with very sparse loss feedback. Learning Parametric Closed-Loop Policies for Markov Potential Games (top 23%) -> general closed loop analysis for Markov potential games and show that deep reinforcement learning can be used for learning approximate closed-loop Nash equilibrium. Leave no Trace: Learning to Reset for Safe and Autonomous Reinforcement Learning (top 26%) -> an autonomous method for safe and efficient reinforcement learning that simultaneously learns a forward and backward policy, with the backward policy resetting the environment for a subsequent attempt. Reinforcement Learning Algorithm Selection (top 31%) -> online algorithm selection in the context of Reinforcement Learning. Divide-and-Conquer Reinforcement Learning (top 33%) -> optimizes an ensemble of policies, each on a different slice of the state space, and gradually unifies them into a single policy that can succeed on the whole state space. RL Algorithms Variance Reduction for Policy Gradient with Action-Dependent Factorized Baselines (oral) (top 12%) -> Action-dependent baselines can be bias-free and yield greater variance reduction than state-only dependent baselines for policy gradient methods. The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning (top 6%) -> Reactor combines multiple algorithmic and architectural contributions to produce an agent with higher sample-efficiency than Prioritized Dueling DQN while giving better run-time performance than A3C. Action-dependent Control Variates for Policy Optimization via Stein Identity (8%) -> a control variate method to effectively reduce variance for policy gradient methods, motivated by the Stein’s identity. TD or not TD: Analyzing the Role of Temporal Differencing in Deep Reinforcement Learning (top 9%) -> re-examine the role of TD in modern deep RL, using specially designed environments that each control for a specific factor that affects performance, such as reward sparsity, reward delay or the perceptual complexity of the task. Boosting the Actor with Dual Critic (top 32%) -> from the Lagrangian dual form of the Bellman optimality equation. The algorithm achieves the state-of-the-art performances across several benchmarks. Temporal Difference Models: Model-Free Deep RL for Model-Based Control (top 35%) -> a special goal-condition value function trained with model free methods can be used within model-based control, resulting in substantially better sample efficiency and performance. Guide Actor-Critic for Continuous Control (top 45%) -> propose a novel actor-critic method that uses Hessians of a critic to update an actor. Trust-PCL: An Off-Policy Trust Region Method for Continuous Control (top 52%)-> extend recent insights related to softmax consistency to achieve state-of-the-art results in continuous control. TreeQN and ATreeC: Differentiable Tree-Structured Models for Deep Reinforcement Learning (top 40%) -> We present TreeQN and ATreeC, new architectures for deep reinforcement learning in discrete-action domains that integrate differentiable on-line tree planning into the action-value function or policy. RL Network Architecture Emergence of grid-like representations by training recurrent neural networks to perform spatial localization (top 1%) -> how neural representations of space, including grid-like cells and border cells as observed in the brain, could emerge from training a recurrent neural network to perform navigation tasks. Neural Map: Structured Memory for Deep Reinforcement Learning (top 5%) -> use the successor representation to discover eigenoptions in stochastic domains, from raw pixels. Eigenoptions are options learned to navigate the latent dimensions of a learned representation. Active Neural Localization (top 10%) -> a fully differentiable neural network that learns to localize efficiently using deep reinforcement learning. NerveNet: Learning Structured Policy with Graph Neural Networks (top 12%) -> using graph neural network to model structural information of the agents to improve policy and transferability. Semi-parametric topological memory for navigation (top 43%) -> a new memory architecture for navigation in previously unseen environments, inspired by landmark-based navigation in animals. RL Optimization Backpropagation through the Void: Optimizing control variates for black-box gradient estimation (top 8%) -> present a general method for unbiased estimation of gradients of black-box functions of random variables. We apply this method to discrete variational inference and reinforcement learning. Model-Ensemble Trust-Region Policy Optimization (top 16%) -> show that the learned policy tends to exploit regions where insufficient data is available for the model to be learned, causing instability in training. propose to use an ensemble of models to maintain the model uncertainty and regularize the learning process, and use of likelihood ratio derivatives yields much more stable learning. Maximum a Posteriori Policy Optimisation (top 33%) -> based on coordinate ascent on a relative-entropy objective. for continuous control, our method outperforms existing methods with respect to sample efficiency, premature convergence and robustness to hyperparameter settings. Policy Optimization by Genetic Distillation (45%) -> present Genetic Policy Optimization (GPO) that uses imitation learning for policy crossover in the state space and applies policy gradient methods for mutation. RL Exploration Parameter Space Noise for Exploration (top 17%) -> Combining parameter noise with traditional RL benefits off- and on-policy methods. DORA The Explorer: Directed Outreaching Reinforcement Action-Selection Memory Augmented Control Networks (top 20%) -> a generalization of visit-counters that evaluate the propagating exploratory value over trajectories, enabling efficient exploration for model-free RL. Deep Bayesian Bandits Showdown: An Empirical Comparison of Bayesian Deep Networks for Thompson Sampling (top 27%) -> An Empirical Comparison of Bayesian Deep Networks for Thompson Sampling. Noisy Networks For Exploration (top 27%) -> A deep reinforcement learning agent with parametric noise added to its weights can be used to aid efficient exploration. RL Reward Intrinsic Motivation and Automatic Curricula via Asymmetric Self-Play (top 11%) -> Unsupervised learning for reinforcement learning using an automatic curriculum of self-play. Unsupervised Learning of Goal Spaces for Intrinsically Motivated Goal Exploration (top 19%) -> We propose a novel Intrinsically Motivated Goal Exploration architecture with unsupervised learning of a space where goals can be sampled, and compare systematically various representation learning algorithms in this context. Emergent Complexity via Multi-Agent Competition (top 23%) -> competitive multi-agent environment trained with self-play can produce behaviors that are far more complex than the environment itself. Learning Robust Rewards with Adverserial Inverse Reinforcement Learning (top 25%) -> propose an adversarial inverse reinforcement learning algorithm capable of learning reward functions which can transfer to new, unseen environments. Truncated Horizon Policy Search: Combining Reinforcement Learning & Imitation Learning (top 53%) -> theoretically-motivated method combines imitation and reinforcement learning via the idea of reward shaping using an oracle. The IL part is in the form of an oracle that returns a value function, which is an approximation of the optimal value function. Distributed RL Distributed Distributional Deterministic Policy Gradients (top 14%) -> We develop an agent that we call the Distributional Deterministic Deep Policy Gradient algorithm, which achieves state of the art performance on a number of challenging continuous control problems. Distributed Prioritized Experience Replay (top 5%) -> A distributed architecture for deep reinforcement learning at scale, using parallel data-generation to improve the state of the art on the Arcade Learning Environment benchmark in a fraction of the wall-clock training time of previous approaches. Hierarchical RL Meta Leanring Shared Hierarchies (top 39%) -> learn hierarchal sub-policies through end-to-end training over a distribution of tasks. Hierarchical and Interpretable Skill Acquisition in Multi-task Reinforcement Learning (top 31%) -> a novel framework for efficient multi-task reinforcement learning. Our framework trains agents to employ hierarchical policies that decide when to use a previously learned policy and when to learn a new skill. Multi-Agent Learning Deep Mean Field Games for Modeling Large Population Behavior (oral) (top 1%) -> Inference of a mean field game (MFG) model of large population behavior via a synthesis of MFG and Markov decision processes. Emergence of Linguistic Communication from Referential Games with Symbolic and Pixel Input (oral) (top 8%) -> A controlled study of the role of environments with respect to properties in emergent communication protocols. Emergent Translation in Multi-Agent Communication (top 16%) -> a communication game where two agents, native speakers of their own respective languages, jointly learn to solve a visual referential task. We find that the ability to understand and translate a foreign language emerges as a means to achieve shared goals. Emergent Communication through Negotiation (top 35%) -> We teach agents to negotiate using only reinforcement learning; selfish agents can do so, but only using a trustworthy communication channel, and prosocial agents can negotiate using cheap talk. RL Meta-learning, Transfer, Continuing Learning Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments (oral) (top 1%) -> develop a simple gradient-based meta-learning algorithm suitable for adaptation in dynamically changing and adversarial scenarios. Zero-Shot Visual Imitation (oral) (top 2%) -> Agents can learn to imitate solely visual demonstrations (without actions) at test time after learning from their own experience without any form of supervision at training time. Modular Continual Learning in a Unified Visual Environment (top 5%) -> a neural module approach to continual learning using a unified visual environment with a large action space. Learning an Embedding Space for Transferable Robot Skills (top 7%) -> a method for reinforcement learning of closely related skills that are parameterized via a skill embedding space, taking advantage of latent variables and exploiting a connection between reinforcement learning and variational inference. The main contribution is an entropy-regularized policy gradient formulation for hierarchical policies, and an associated, data-efficient and robust off-policy gradient algorithm based on stochastic value gradients. Learning to Multi-Task by Active Sampling (top 21%) -> Letting a meta-learner decide the task to train on for an agent in a multi-task setting improves multi-tasking ability substantially A Simple Neural Attentive Meta-Learner (top 22%) -> a simple RNN-based meta-learner that achieves SOTA performance on popular benchmarks. Progressive Reinforcement Learning with Distillation for Multi-Skilled Motion Control (top 23%) -> A continual learning method that uses distillation to combine expert policies and transfer learning to accelerate learning new skills. Meta-Learning and Universality: Deep Representations and Gradient Descent can Approximate any Learning Algorithm (top 24%) -> Deep representations combined with gradient descent can approximate any learning algorithm. RL Applications Ask the Right Questions: Active Question Reformulation with Reinforcement Learning (oral) (top 7%) -> propose an agent that sits between the user and a black box question-answering system and which learns to reformulate questions to elicit the best possible answers. Reinforcement Learning on Web Interfaces using Workflow-Guided Exploration (top 13%) -> solve the sparse rewards problem on web UI tasks using exploration guided by demonstrations. Go for a Walk and Arrive at the Answer: Reasoning Over Paths in Knowledge Bases using Reinforcement Learning ( top 27%) -> learns to walk on a knowledge graph and answer queries. Leveraging Grammar and Reinforcement Learning for Neural Program Synthesis (top 27%) -> Using the DSL grammar and reinforcement learning to improve synthesis of programs with complex control flow. N2N learning: Network to Network Compression via Policy Gradient Reinforcement Learning (top 28%) -> A novel reinforcement learning based approach to compress deep neural networks with knowledge distillation. Conclusion Deep Reinforcement Learning is one the of biggest and hottest topics. In addition to theory and algorithms, meta-learning, continual learning, credit assignment (reward), exploration, hierarchical, multi-agent and network architectures are popular sub-directions for RL. There a big less-explored space for RL on network architectures, considering the amount of papers on network architectures for vision problems.

A Digest of Reinforcement Learning Papers from ICLR 2018

8

a-digest-of-reinforcement-learning-papers-from-iclr-2018-1be7322886ab

2018-07-23

2018-07-23 21:58:02

https://medium.com/s/story/a-digest-of-reinforcement-learning-papers-from-iclr-2018-1be7322886ab

false

2,095

null

Machine Learning

machine-learning

Machine Learning

51,320

Jian Zhang

Autonomous System at Apple Inc.

953e3126c9a4

jianzhang_23841

42

6

20,181,104

null

0

null

0

634d4b270054

2018-06-07

2018-06-07 09:25:23

2018-06-07

2018-06-07 09:26:36

1

false

en

2018-06-07

2018-06-07 09:26:36

3

1bea1ec0026c

1.184906

1

0

Drones have become a hot topic these days and cities, companies, organizations are deploying it for various purposes. Companies like real…

5

New Drone Rules Coming For Canada Drones have become a hot topic these days and cities, companies, organizations are deploying it for various purposes. Companies like real estate agencies, wedding photographers, and farmers are using drones, commercially. With the growing use of drones, a major concern also rises. For the safe and secure flights in Canada, the state has come up with new drone rules. Earlier the rules included, drones weighing more than 250 gms, would operate 9 km away from airports, no more than 90-meter altitude, no more than 500 meters from the operator and should remain within the sight. Now, new Transport Canada rules are being proposed that includes dividing drones into weight classes and permits and liability insurance for heavier drones. The drones would be marked with owner names and contact information. The distance from the airport would be reduced to 5.5km and it’s not to be operated in city parks. With the changes in the regulations, more drone courses could also be seen. Source: https://bit.ly/2J3LJBa About DEEPAERO DEEP AERO is a global leader in drone technology innovation. At DEEP AERO, we are building an autonomous drone economy powered by AI & Blockchain. DEEP AERO’s DRONE-UTM is an AI-driven, autonomous, self-governing, intelligent drone/unmanned aircraft system (UAS) traffic management (UTM) platform on the Blockchain. DEEP AERO’s DRONE-MP is a decentralized marketplace. It will be one stop shop for all products and services for drones. These platforms will be the foundation of the drone economy and will be powered by the DEEP AERO (DRONE) token.

New Drone Rules Coming For Canada

35

new-drone-rules-coming-for-canada-1bea1ec0026c

2018-06-16

2018-06-16 15:58:58

https://medium.com/s/story/new-drone-rules-coming-for-canada-1bea1ec0026c

false

261

AI Driven Drone Economy on the Blockchain

null

DeepAeroDrones

null

DEEPAERODRONES

null

deepaerodrones

DEEPAERO,AI,BLOCKCHAIN,DRONE,ICO

DeepAeroDrones

Deepaero

deepaeros

Deepaero

0

DEEP AERO DRONES

null

dcef5da6c7fa

deepaerodrones

277

0

20,181,104

null

0

null

0

null

2018-08-17

2018-08-17 11:09:05

2018-08-17

2018-08-17 11:09:51

2

false

tr

2018-08-17

2018-08-17 11:09:51

6

1bec31b405f5

2.088994

0

Heroic, Güneydoğu Asya’da etiketsiz ve asılsız tüketicilere verilen garantilerle kredi sağlayan bankaların geleceğidir. Hero Token adında…

3

Heroic ICO: Token Bilgisi Heroic, Güneydoğu Asya’da etiketsiz ve asılsız tüketicilere verilen garantilerle kredi sağlayan bankaların geleceğidir. Hero Token adında bir kripto ünitesinin lansmanı ile, HEROIC bir blok zincirine dayalı, yüksüz kredi olarak genişlemeyi amaçlamaktadır. Softbank, Alibaba ve 500 Startups gibi risk sermayesi kapitalistleri tarafından desteklenen organizasyon, 2015 yılında Filipinler’de operasyonlara başlamış ve binlerce Filipinli’nin uygun fiyatlı krediye erişmesine yardımcı olmuştur. MISSIONHero’nun misyonu, bankacılık sektörünü daha erişilebilir ve uygun maliyetli hale getirmek için devrim yapmaktır. Güneydoğu Asya’da başlayarak, banka dışı veya az bağlı olarak. Heroic Token nedir? Kahraman olumlu finansal ve sosyal faydalar yaratmak için yaratılmıştır. Heroic Token, kazanılan faiz gelirinin bir parçası olan ödülü alma hakkını temsil eder. Şirkette eşitliği temsil etmez, ne de içsel değeri yoktur. Tüm toplam jetonlar% 20'ye kadar faiz geliri elde etme hakkına sahiptir. ORTAKLIKLARIMIZ Aslında bu projenin alibaba.com tarafından desteklendiğinden çıktım! Herkes, alibaba’nın ne olduğunu ve ne kadar büyük olduklarını bilir. Buradaki alibaba’nın ne olduğunu bilmeyenler için açıklama “Nisan 2016 itibariyle dünyanın en büyük perakendeci olan Alibaba, 200'ün üzerinde ülkede ve en büyük internet şirketlerinden biri olan Walmart’ı geride bırakıyor. Çevrimiçi satışları ve karları, 2015'ten beri bütün ABD perakendecilerini (Walmart, Amazon ve eBay dahil) aştı. Medya ve eğlence endüstrisine yayıldı ve gelirleri yılda 3 haneli yüzde arttı Onların piyasa değeri Ağustos 2017'de 442 milyar ABD Doları’na yükseldi ve kişisel olarak hizmetlerini kullandım ve HERO’yu destekledikleri için gerçekten çok mutluyum. Bir jetonun faydaları Kahraman olumlu finansal ve sosyal etkiler yaratmak için yaratılmıştır. Heroic Token, kazanılan faiz gelirinin bir parçası olan ödülü alma hakkını temsil eder. Tüm toplam jetonlar% 20'ye kadar faiz geliri elde etme hakkına sahiptir. Kanıt satış sırasında sağlanan fonlar için kanıtlar değerlendirilecektir. Dağıtılabilen faiz gelirlerinin% 20'ye kadarı, her üç ayda bir belirli Ethereal (ETH) cüzdanlarına aktarılır. ETH daha sonra, akıllı sözleşmenin koşullarına (yani, sahip olunan simgenin payına oranla kazanılan payın payı) uygun olarak Kahraman sembolünün tüm sahiplerine dağıtılır. Uzun vadeli kahraman, ödülü daha kısa sürede dağıtmaya çalışacak. Aylık ödül dağıtımı hedefiyle birlikte, şirket hakim piyasa oranlarında açık pazardan Heroic jetonu satın almak için kâr yüzdesi kullanabilir, bu nedenle jeton değerinin Kahramanın başarısı ile pozitif bir şekilde ilişkilendirilmesi gerekir. Heroic Token Dağıtımı Jetonlar, satış sözleşmesine verdikleri anda, alıcılara yaratılır ve verilir. Ticaret ve aktarılabilirlik sadece jeton satışının sona ermesinden sonra başlar. Pratik bir bakış açısından bu, satışa kadar jeton vermemekle aynı şeyi başarır.Ancak teknik açıdan bakıldığında, satışın sonunda belirteç kullanım probleminin gaz limitine girmesini engelliyoruz.1 HEROIC jetonu 0,005 ETH’ye eşittir (1 ETH 200 Heroic jetonu satın alacak).Kabul Edilen KriptocurrenlerEthereum (ETH), ETH Klasik, BitCoin (BTC), Dalgalanma, LiteCoin, Dalgalar. Sonuç olarak, projenin değerli ve yararlı olduğunu belirtmek isterim. Bağlantıları aşağıdaki resmi kaynaklara bırakacağım, böylece her biriniz projeyi şahsen tanıyabiliyorsunuz. İLETİŞİM WEBSITE : https://heroic.com ANN BITCOINTALK :https://bitcointalk.org/index.php?topic=3992644.0 FACEBOOK : https://www.facebook.com/heroiccybersecurity/ TWITTER : https://twitter.com/heroiccyber TELEGRAM : https://t.me/heroicdotcom Medium: https://medium.com/heroic-com

Heroic ICO: Token Bilgisi

0

heroic-ico-token-bilgisi-1bec31b405f5

2018-08-17

2018-08-17 11:09:52

https://medium.com/s/story/heroic-ico-token-bilgisi-1bec31b405f5

false

452

null

Cybersecurity

cybersecurity

Cybersecurity

24,500

Coinveteran

null

561d704c77d

coinsveterans

263

602

20,181,104

null

0

null

0

null

2017-09-03

2017-09-03 18:11:22

2017-09-03

2017-09-03 18:14:43

0

false

en

2017-09-03

2017-09-03 18:58:48

4

1becd4d5cf64

0.362264

0

He really believes this insane thing about himself and his birthright:

5

Michael O. Church’s Excuse He really believes this insane thing about himself and his birthright: “I’m almost 30 and my IQ is over 150. I should be an EIR or doing cutting-edge machine learning work, and I’m not, because I was robbed by those fuckers. Yes, they’ll get their comeuppances. I’m sure they’ll fall into the wrong fight. But I suffer every day from what those pieces of shit stole from me. I’m years behind where a person at my level of talent should be.” — Michael O. Church, May 17th, 2013 Archive.org WebCitation.org Archive.fo

Michael O. Church’s Excuse

0

michael-o-churchs-excuse-1becd4d5cf64

2018-04-06

2018-04-06 05:05:30

https://medium.com/s/story/michael-o-churchs-excuse-1becd4d5cf64

false

96

null

Entrepreneurship

entrepreneurship

Entrepreneurship

226,400

Michael O. Church Quotes

Bizarre and awful quotes from Michael O. Church. Trigger warning: bigotry and advocacy of violence.

373929999227

michael_o_church_quotes

4

2

20,181,104

null

0

null

0

null

2017-09-25

2017-09-25 18:16:32

2017-09-25

2017-09-25 19:11:49

0

false

en

2017-09-25

2017-09-25 19:11:49

0

1bed244ed102

0.988679

1

0

Time series models are self-explanatory. These models are designed for understanding the data with time components.

3

Time series — simple ARIMA explanation Time series models are self-explanatory. These models are designed for understanding the data with time components. There could be various reasons why you would like to learn time-series model, but I personally think that ‘forecasting’ or ‘predicting’ is one of the interesting area. Let’s assume that ‘Alice’ is a data-driven doner kebab restaurant owner. Since she has data-driven mindset, she always record daily sales of kebab. She would like to predict tomorrow chicken kebab selling to decide how much chicken would need for tomorrow. Alice read various articles on forecasting, predicting models, and found out ARIMA is one of the most commonly used time-sereis model. She felt intimidated after reading all those scary formal mathematical terms explaining what is ARIMA model is. If Alice sounds like you, this story is for you. Let’s break down the word ‘ARIMA’. AR + I + MA is this. AR — Auto Regressive I — Integration MA — Moving Average Auto Regressive model uses past data to explain current data. For instance, number of chicken sale today was influenced by last 2 days of chicken. Specifically, Number of Kebab sale = X Today X = (yesterday X * (0.8)) + (the day before yesterday X * (0.2)) Moving Average model also uses past data to explain current data. However, this kid is a little bit more tricky. It uses errors. TODO Then, how can you utilize this model? I’d suggest you to use R or Python packages, and try implement ARIMA package with simple dataset.

Time series — simple ARIMA explanation

14

time-series-simple-arima-explanation-1bed244ed102

2018-01-04

2018-01-04 02:34:46

https://medium.com/s/story/time-series-simple-arima-explanation-1bed244ed102

false

262

null

Data Science

data-science

Data Science

33,617

Jaeseok An

CS @ Oxford University | Developer @ Indorse.io

cfdeb5241ee7

jaeseokn

6

38

20,181,104

null