Alaamer/medium-articles-posts-with-content

0

null

0

null

2018-07-22

2018-07-22 20:45:03

2018-07-22

2018-07-22 20:38:27

6

false

en

2018-07-23

2018-07-23 06:58:22

15

1c4d4d8f6c9c

4.753774

2

0

When I was a child, I used to love watching movies about Robots. Some of them were good, some of them had gray shades. But all of them…

5

Kambria: The next revolution in Robotics has arrived When I was a child, I used to love watching movies about Robots. Some of them were good, some of them had gray shades. But all of them carried a similar message. The next stage of evolution for Human will be in the field of AI & Robotics and the impact Robotics will bring to your life will be immense. That time I thought, After 20–25 years when I will be a grown-up, Robots will be a part of everyone life, like what we have seen in those movies. Now I am in my mid thirties. That dream is still halfway only. So what went wrong? There are 100’s of reason but we can round-off it into majorly 3 points. Heavy investments lead to highly expensive products which reduce mass affordability. Technology is too complex to develop. You need more money and brains to pool in which create a deterrent for entrance for new players. Since most of the R&D is proprietary in nature, sharing rarely happens which delays the progress. Because of all these above reasons, application of Robots is still confined to niche segments majorly in Industrial applications. Irrespective of all these setbacks there were some who took efforts to make that dream of “ Robot in each home “ a reality. Ohmni labs were one of them. They introduced “ Ohmni “ a home robot for connecting families via video “telepresence.” Have a look Does Ohmni sound interesting? You can read more about the project or order here to own one. So What is Kambria? To realize the dream of “ Robot in each home “, Ohmni has laid the foundation by introducing an ambitious blockchain Project termed Kambria. As per the Whitepaper, Kambria ( named after the Cambrian Explosion ) can be summarized as “Kambria is the first blockchain project to build open innovation platform that enables and incentivizes collaborations in R&D, manufacturing, and commercializing advanced technology with a focus on AI & Robotics applications in the consumer space” In simple terms, Kambria will be building an open source block-chain platform where both developers and manufacturers and can benefit mutually by collaborating with each other. All the contribution will be incentivized via KAT tokens. Over the time due to the continuous knowledge sharing happening to the platform, it will be a one-stop-shop for robotics with a collection of thousands of repositories in any robotics field. With the help of this platform Developers can enjoy and receive compensation for their contribution Companies won’t be needed to invest in heavy R&D as they can take advantage of the knowledge pool. New Startups will be immensely benefited as it removes the high-cost entry barriers. End users will be able to have their own Robot “ low of cost” To further understand let us have a look at Kambria Ecosystem from a layman’s view As you can see the heart and the brain of the eco-system is the “code base” which is written in the modular language, where all the required tools to build a robot are pooled in. Developers can work on a specific component and submit it back to the code base and will be rewarded in Tokens. If a developer wants to get the parts made, they can use the service of Kambria manufacturing alliance and get the job done from any part of the world. Once the part is delivered, the manufacturer will be paid in Tokens as well. Another component of the ecosystem is the Kambria innovation marketplace. Here there will be open competitions and challenges. Whoever offers the best solution will be rewarded back with the handful of tokens. This will accelerate the robotics development as the solutions get mapped back to the platform. So any company who wants to start a robotics revolution just have to visit the platform and utilize the knowledge stored in it rather than to start from the scratch. Roadmap So far Team is in line with the delivery of Roadmap. Currently expecting the platform to be launched by Q4–2018. Team Kambria is lead by Founder and CEO Dr.Thuc Vu. The team is well balanced with rich experience from the field of robotics. Kambria is also well supported by Advisers with rich experience from blockchain world which includes leaders from Tomochain, HASHED, Kyber network etc Kambria Token (KAT) Kambria(KAT) is the native tokens for Kambria ecosystem. Currently an ERC20 based utility token, it will be used to fuel Kambria’s marketplace. Not only that KAT will be used for various other purposes like the reward for bounties, Governance (Voting), payment for licensing etc. Looking at the future of Project, there is a high chance that Kambria will be launching its own mainnet as well. Partnerships Kambria is well supported not only from the industry front but also from the top names in the academical world like Stanford University, Nanyang Technological University, Carnegie Mellon University etc Token Metrics Very limited details of Token metrics have been published so far. The details currently available are Fundraising Goal: 19 Million USD Total Tokens to be issued: 5,000,000,000 Type: ERC-20 based Kambria is not without competitions. There are already projects like Project PAI, Singularitynet etc working in the similar field. But none of them are in Robotics and none of them are working towards building a platform, which will be a “one stop shop” in future for sure which gives Kambria an advantage over the others. If you wish to know more about this Project you can be in touch with the following social channels Website: https://kambria.io/ Whitepaper: https://kambria.io/Kambria_White_Paper_v2.pdf Telegram (ENG): https://t.me/kambriaofficial Telegram (KOR): https://t.me/KambriaKorea Telegram (VIE): https://t.me/KambriaVietnam Telegram Announcement Channel: https://t.me/kambria Kakao Talk: https://open.kakao.com/o/gcUpSEQ Twitter: https://twitter.com/KambriaNetwork Facebook: https://facebook.com/KambriaNetwork Reddit: https://www.reddit.com/r/KambriaOfficial/ Email: info@kambria.io Full Disclaimer: This article is not to be taken as Financial advice. Before Investing DYOR. “This article was created in exchange for a potential token reward through Bounty0x”. Author part of BountyOx ,user name of BountyOx : nibupraju Kambria @ www.kambria.io #blockchian #BountyOx

Kambria: The next revolution in Robotics has arrived

82

kambria-the-next-revolution-in-robotics-has-arrived-1c4d4d8f6c9c

2018-07-23

2018-07-23 06:58:22

https://medium.com/s/story/kambria-the-next-revolution-in-robotics-has-arrived-1c4d4d8f6c9c

false

1,008

null

Robotics

robotics

Robotics

9,103

Crypto Pineapple

Part of Huobi Global I Hacker Noon| The Start Up l Good Audience | Data Driven Investor etc !!!!

3e972b582974

cryptopinapple

711

123

20,181,104

null

0

null

0

null

2017-10-19

2017-10-19 19:00:50

2017-10-19

2017-10-19 19:04:07

1

false

en

2017-10-19

2017-10-19 19:04:07

1

1c4d6d96c152

1.162264

0

Computers do not think the way humans do. For example, the Chinese Room experiment is an imaginary situation on how computers “think.” This…

1

Computers do not Think Photo Credit: ohthehumanityblog Computers do not think the way humans do. For example, the Chinese Room experiment is an imaginary situation on how computers “think.” This experiment shows how a computer does not fully understand what it is supposed to do but follows directions anyway. People do not automatically follow directions like computers. Often times people want to rebel and not follow directions or people ask questions why they have to do a certain task. Another example is, humans have morals that make them think about the decisions they make and the consequences of the decisions whereas computers don’t have a sense of moral that influences their decisions. Computers are programmed to pull up something that relates to whatever is typed in or use content addressable memory, but people can decide if they want to respond to a question or not. A computer’s way of responding has no thought put into it. The last example is, computers do not have a choice in their decisions, and people have free will. People choose what they want to do rather than follow a guideline that one has to follow. A computer does not obtain free will as a person does; it cannot choose for itself and is dependent on the people who program it, there are no decisions made by the computer. All in all, people should not think that computers and humans think alike; computers are programmed to think a certain way and a person is born with the ability to make self made decisions.

Computers do not Think

0

computers-do-not-think-1c4d6d96c152

2018-04-24

2018-04-24 05:40:23

https://medium.com/s/story/computers-do-not-think-1c4d6d96c152

false

255

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Chaewon Park

null

91d6c4fe694

cpark1353

0

2

20,181,104

null

0

null

0

91eb38fc66d0

2018-08-30

2018-08-30 13:32:00

2018-08-30

2018-08-30 13:42:06

1

false

en

2018-08-30

2018-08-30 13:42:06

4

1c4e96e60969

2.184906

0

Introducing the research paper that describes a practical approach to building natural language interfaces for structured data querying.

5

Pragmatic Approach to Structured Data Querying via Natural Language Interface Introducing the research paper that describes a practical approach to building natural language interfaces for structured data querying. As the use of technology increases and data analysis becomes integral in many businesses, the ability to quickly access and interpret data has become more important than ever. Why build natural language interface for structured data querying? Today’s information retrieval technologies utilized by companies claim to democratize data but the reality is that these technologies are very complex and require understanding of query languages, such as SQL, strong analytical skills, extensive training, and knowledge of data structure to formulate a valid query. Business people can barely use these systems without the help of a skilled business analyst. Companies need to employ business analytics teams to help nondata professionals interact with enterprise data. These teams typically have an ever growing reporting backlog, as a result, even a simple question may take days to answer. To reduce some burden on already overstretched data teams, many organizations are looking for self-service tools that allow non-developers to query databases using natural language without needing a data analyst for every report. FriendlyData’s approach to structured data querying via natural language interface At FriendlyData we are building a natural language interface for database querying. Our product translates natural language questions into corresponding SQL queries making data accessible to everyone in a company. Last month we applied our query translation method to WikiSQL dataset (a large crowd-sourced dataset for developing natural language interfaces for relational databases). FriendlyData’s query translation algorithm demonstrated high accuracy, in addition, it doesn’t require training on the massive datasets, which makes it easier and faster to implement compared to machine learning based algorithms. Now we’re pleased to share the research paper “A pragmatic approach to structured data querying via natural language interface”, where we describe our algorithm in detail and discuss a number of factors that can dramatically affect the system architecture and the set of algorithms used to translate NL queries into a structured query representation. Our primary goal is to help companies find the best solution when both high quality query translation and high security standards of architecture are required. Our method is designed for real-life business cases, where such factors as data security, time, scalability, and accuracy are mission critical. By no means will our approach be the best in every case, but our goal is to show what factors really matter for enterprises in real-life scenarios. Find the whole paper on arxiv.org, and also be sure to follow us on Twitter, where we are sharing all the latest thinking in Natural Language Processing along with company news, papers, and other useful resources. Democratize your data FriendlyData helps to respond to one of the key challenges in the world of enterprise data — building the power of data and analytics into day-to-day decision-making. The solution we offer is data democratization. FriendlyData makes data accessible to everyone by providing a user-friendly natural language search interface for databases. Foster data-driven culture in your organization with us! Originally published at www.friendlydata.io.

Pragmatic Approach to Structured Data Querying via Natural Language Interface

0

pragmatic-approach-to-structured-data-querying-via-natural-language-interface-1c4e96e60969

2018-08-30

2018-08-30 13:42:31

https://medium.com/s/story/pragmatic-approach-to-structured-data-querying-via-natural-language-interface-1c4e96e60969

false

526

Empowering people to make better decisions with data

null

friendlydata

null

Friendly Data

hello@friendlydata.io

friendly-data

ARTIFICIAL INTELLIGENCE,DATA ANALYTICS,BUSINESS INTELLIGENCE

nlpdb

Naturallanguageprocessing

naturallanguageprocessing

Naturallanguageprocessing

848

Irina Peregud

Technology product marketing at FriendlyData. Techno-optimist. Passionate about customer experience, data-driven marketing & innovation.

e0941b145e22

IrenaPeregood

11

23

20,181,104

null

0

null

0

null

2017-09-23

2017-09-23 14:30:25

2017-09-23

2017-09-23 14:31:14

0

false

en

2017-09-23

2017-09-23 14:31:14

1

1c4f4e4b911

0.422642

0

This is starting — thankfully — to become really obvious — advanced robots are going to take the human jobs; every job. But amazingly, most…

4

YOU Have To Make Them Aware. This is starting — thankfully — to become really obvious — advanced robots are going to take the human jobs; every job. But amazingly, most people don’t even think of it, or if they do, it’s just for an instant before they go on to something else. This is one of the two GIANT problems over the next 20 years, and most people are totally unaware of the issue. YOU have to make them aware, while there’s time, so we can all have a meaningful discussion about a decent living income, rather than a starvation UBI. It’s up to YOU. This is a fantastic start. https://www.youtube.com/watch?v=exEj4zsnwj8&feature=youtu.be

YOU Have To Make Them Aware.

0

you-have-to-make-them-aware-1c4f4e4b911

2018-03-29

2018-03-29 13:27:47

https://medium.com/s/story/you-have-to-make-them-aware-1c4f4e4b911

false

112

null

Basic Income

basic-income

Basic Income

2,763

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

76d582de2a0e

2018-06-01

2018-06-01 06:26:57

2018-06-01

2018-06-01 07:42:11

2

false

zh-Hant

2018-06-01

2018-06-01 15:07:33

0

1c50b25e3b19

0.52673

1

0

工作一段時間後，也慢慢體會到職場有許多難處，常常會思考目前公司的產品是否符合消費者的需求，還是只是一群工程師閉門造車的自High產物。

2

深度為王？其實你沒有那麼需要深度學習工作一段時間後，也慢慢體會到職場有許多難處，常常會思考目前公司的產品是否符合消費者的需求，還是只是一群工程師閉門造車的自High產物。這讓我想起以前在參加YEF比賽中，受到的震撼教育，從前不管參與什麼比賽都是管院同學一起組隊，大家的想法都很接近，爭執也很少，而在YEF比賽中，我們隊伍裡面有來自四面八方的菁英，不管是讀管理、工程、語文、設計等等的專業，回想第一次開會，大家針對一個很小的事件就爭執不休，提出的論點、看法也天差地遠，那時的我也才意會到原來以前看事情的角度是那麼狹窄、單一，也了解到要做好一個產品，必須是很多不同領域的人協同合作。那時候業師也講到，常常有很多技術導向的創業，可能就是把在實驗室做出來的結果直接拿出來賣，但往往以悲劇收場，這種團隊裡常常只充斥著工程師，他們做出來一些技術含量高的商品，但使用者可能覺得他們根本沒有這種需求。從以前的大數據，到現在的深度學習熱潮，許多公司無所不盡其極把自己公司的產品跟這些技術沾上邊，看能否分一杯羹，但是你的公司真的需要深度學習嗎？還是只是你的主管為了潮流，急忙跑過來跟你說 Simon Sinek曾在TED介紹黃金圈理論，他在演講中提到 People don’t buy what you do, they buy why you do it 黃金圈其實就是將我們平常傳遞訊息的順序做了一下改變，必須以WHY為主軸去傳遞，感動顧客，進而使他們願意買單。舉例來說，現在很流行的智慧音箱，許多公司都強調自己的音箱有多少功能，接了多少第三方的服務，但顧客往往聽到這些只是不斷覺得頭昏腦脹，因為這只是一直不斷描述功能面，但你還是沒有告訴顧客到底為什麼需要這個音箱？買了這個音箱後生活會發生什麼改變？必須動之以情，使你的顧客進入那個情境，產生共鳴。回到今天的主題其實你沒有那麼需要深度學習在很多企業裡，決策者根本沒有搞清楚自己的服務型態是什麼，就只是一味的說要採用深度學習，以chatbot為例，主要分成聊天型和任務型，前者著重在互動、閒聊，所以你希望你的chatbot可以有舉一反三的功能，而後者則是希望幫助使用者完成一些特定的任務，而這些任務常常是一個口令一個動作，機器也不需要舉一反三的能力，所以在很多時候只要靠rule-based model就可以達到這樣的功能。之前在上一堂推薦系統的課時，講師說要開發一個好的推薦系統產品，有四個關鍵元素: UI 和 UE 數據領域知識演算法而他們的重要程度依次遞減，最重要的一定是UI和UE，如果你開發一個語意系統，但是網站的使用者體驗很差，用戶根本體驗不到你的語意演算法有多deep，早就已經離開服務了；而數據也是相輔相成的，如果你的數據不夠，沒有太多使用者輪廓的資料，你也無法依每個使用者去做個性化回話。筆者認為想要建構一個完整的語意系統並非一蹴可幾，第一步就是做好使用者體驗，留不住使用者後面的東西都是枉然，當你成功吸引到一些使用者使用你的服務時，你才可以開始累積資料，當資料足夠開始可以優化你的模型，產生一個正向的循環。筆者在本篇文章不是否定深度學習，而是希望企業能更仔細的評估你現在所提供的服務是否需要用到深度學習，或者其實rule-based model才是比較適合你的；而你現階段是否應該著重在深度學習演算法開發，還是你連基本的UI跟資料都沒有做好。筆者最後希望大家能回歸需求，以技術解決需求，以需求為本，一步一腳印。

深度為王？其實你沒有那麼需要深度學習

50

深度為王-其實你沒有那麼需要深度學習-1c50b25e3b19

2018-06-01

2018-06-01 15:07:36

https://medium.com/s/story/深度為王-其實你沒有那麼需要深度學習-1c50b25e3b19

false

38

Share Data science, machine learning and deep learning

null

Data Scientists Playground

null

data-scientists-playground

DATA ANALYTICS,DATA SCIENCE,MACHINE LEARNING,DEEP LEARNING,DATA ANALYSIS

null

Data Science

data-science

Data Science

33,617

ricky chang

null

118decbb44c2

ricky94511

5

2

20,181,104

null

0

null

0

null

2017-12-26

2017-12-26 13:25:10

2017-12-26

2017-12-26 13:28:57

1

false

en

2017-12-26

2017-12-26 13:28:57

0

1c511bd7d8a2

0.973585

2

0

INJECTION MOLDING

5

Use case of the day INJECTION MOLDING Injection molding is a manufacturing process for producing parts in large volume . It is typically used in mass production process. Its a production process by injecting material into a mold. Problems faced Overheating, temperature fluctuations, air pressure, manual data collection, unpredictable breakdowns Our smart solution Using the technologies of IoT and AI, TGS can effectively offer a smart solution of the issues faced by everyone in the industry regularly i.e. overheating, high temperature, etc. Our smart device using a smart integration of sensors will collect all the data (data is gathered from the sensors directly via a PLC) and send it to the cloud while making it remotely accessible with its mobile applications. It will help in analyzing the over-usage and tell the exact amount of production and waste produced in that particular time giving complete transparency to make the operations efficient. Moreover, it gives you the flexibility to set alerts and notifications to know if the machine exceeds the limit of heat, temp etc in real time. Get the benefits of efficient predictive maintenance thereby reducing the manual efforts and errors while automating the management. Get your smart solution today, contact us for more details.

Use case of the day

15

use-case-of-the-day-1c511bd7d8a2

2017-12-26

2017-12-26 19:42:58

https://medium.com/s/story/use-case-of-the-day-1c511bd7d8a2

false

205

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Things Go Social

Your interaction with machines will change when your machines will talk to you. Find out what happens when things go social!?

d6d7546b0773

ThingsGoSocial

18

155

20,181,104

null

0

null

0

null

2017-12-14

2017-12-14 22:43:09

2017-12-15

2017-12-15 01:12:51

1

false

en

2017-12-17

2017-12-17 08:15:23

13

1c52ed44439c

3

7

1

0

With a lot of talk about data science in Nigerian business circles in recent times, quite a number of practitioners may be losing sight of…

5

Saving Millions in Naira with Simple Decision Tree Classifier With a lot of talk about data science in Nigerian business circles in recent times, quite a number of practitioners may be losing sight of its true essence. According to a Cloudera training material on data science, data science is ultimately meant to uncover business value for an organization and then communicating it as actionable insights to stakeholders. At a very high level, data science is meant to solve business problems. Either reducing cost or increasing revenue. This approach to solving business problems is very scientific, highly methodical and research-based. Hence, we are not just talking about slapping your data with some Python/R libraries. We are involved in solving real life problems to create real value. Use Case Problem Statement E-commerce industry faces quite a number challenges peculiar to the Nigerian economy ranging from market maturity, trust in online payment systems, harsh economic realities to human behavioral tendencies. In bid to wring out profit from slim margins in running an e-commerce platform in Nigeria, improving customer retention (e.g. item recommendation engines) and lifetime value or reducing operational overhead are key. One major contributor to high operational overhead is Customer Induced Order Cancelation. In general, orders may be canceled due to various reasons e.g. payment gateway failures, technical glitches, customer dissatisfaction and many more. Our use case is concerned with orders that were canceled by customers after a successful process on the e-commerce platform. Our problem was a classification problem {Class A: CompleteOrder, Class B: CancelDelivery} and our goal was to predict the probability of a customer’s order in either class. Business Goal Reduce customer induced order cancelation by X%. Achieving the aforementioned business goal will impact the organization’s bottomline positively by reducing revenue lost due to canceled orders as well as operational expenses incurred during item return process. Hypothesis The approach to solving the above defined problem is scientific rather than head knowledge, we had to define an hypothesis: If we the know order history of a product and the history of a customer’s behaviour, then we should be able to predict the probability that a specific order placed by a customer would end up in canceled delivery. Analytic Pipeline Stages Maintaining simplicity while ensuring we don’t deviate from our scientific approach, we highlighted to following analytic pipeline stages: Identify patterns in the data set (AKA Exploratory Data Analysis — EDA). Extract inherent features from existing features to uncover underlying patterns (Feature Extraction and Engineering). Communicate findings to stakeholders to agree on features for model formulation (Domain Expertise) Model formulation Production deployment Ensure reproducible research in analysis for portability to other use cases within the organization. Models, Algorithms and Libraries Since the defined hypothesis is a binary classification problem, we deployed classification techniques such as Logistic Regression, Decision Trees, Random Forests (ensemble algorithm) and Support Vector Machines (SVM). Performance comparison in terms of prediction accuracy, speed, model interpretation and ease of deployment to production were used to determine the algorithm of choice. For the first iteration, Decision Tree Classifier was used and analysis was done in Python, however the analysis can be reproduced in R or Scala. The libraries used include: Numpy Pandas Matplotlib for visualizations Seaborn for visualizations Sklearn for modeling and evaluation Patsy for feature preparation Pickle for Model Persistence Results and Evaluation Our approach to training the model was simple: Train set Test set Cross Validation Accuracy on test set and cross validation were 77% and 74% respectively. This may not be as high but the beauty of Decision Tree Classifier is in its interpretability. In production, we achieved about 70% accuracy in predicting Class A and about 40% in predicting Class B. The overall result of this was about 26% MoM reduction in overhead costs incurred due to the identified problem statement over a 3 month period. Conclusion What we have discussed here is a real life use case where we approached a business problem with Data Science Techniques to create real and measurable business value. Much of the work done was a team effort which leads to the need for a data science team and not a data scientist. I hope this drives businesses to challenge their data teams not discuss the fanciness of the newest machine learning tools but create immediate business value with them. Feel free to drop me your comments, connect via LinkedIn or via email to discuss your data opportunities.

Saving Millions in Naira with Simple Decision Tree Classifier

60

saving-millions-in-naira-with-simple-decision-tree-classifier-1c52ed44439c

2018-06-03

2018-06-03 17:47:11

https://medium.com/s/story/saving-millions-in-naira-with-simple-decision-tree-classifier-1c52ed44439c

false

742

null

Machine Learning

machine-learning

Machine Learning

51,320

Olalekan Elesin

Data Engineer, AWS Certified SA, Associate. Lover of scientific experiments and love to get my hands dirty with code…

98b05140c1ad

elesin.olalekan

26

14

20,181,104

null

0

null

0

null

2017-12-14

2017-12-14 15:15:39

2017-12-18

2017-12-18 03:59:24

6

false

en

2017-12-18

2017-12-18 04:28:08

14

1c54344e3328

5.066981

5

1

0

Did you ever think we would have the technology to transform a modern bird into a prehistoric dinosaur? To, theoretically speaking, give a…

5

Hacking the genetic code for fun and profit: The role of A.I.-enabled radiologists in the CRISPR era Did you ever think we would have the technology to transform a modern bird into a prehistoric dinosaur? To, theoretically speaking, give a pig wings and make it fly? To evolve a mere cat into the far superior dog? Well, this is about to be you right now: Soon, there may even be no cats. Enter CRISPR, hailed by Forbes as the “next step” in human evolution. This technology represents an extremely robust tool for editing the genome, in a dirt cheap fashion and in any biological system, however we want — from snipping away an error in cystic fibrosis to removing and replacing that gene that makes cilantro taste like soap. For a summary, enjoy this cute animated YouTube video: To review the system diagrammatically: Cas9 is an enzyme that functions as molecular scissors, snipping DNA at any arbitrary location of interest using “guide RNA” as a molecular address. We can thus delete any DNA sequence of interest. After the old DNA is destroyed, we have the option of replacing it with some arbitrary new piece of DNA. This allows us to, for example, destroy the variant making cilantro taste like soap and replace it with the normal gene, facilitating a fuller and more complete appreciation of awesome tacos. Will CRISPR allow you to better appreciate tacos? This technology is projected to generate ~$25 billion in revenue by 2030, and might cure all genetic disease, replace antibiotics, and solve the shortage of organs for transplantation. In this article we will discuss how radiology will be part of this revolution, increasing our imaging and procedural volume and offering immense benefit to our patients. In addition to medical stuff, these techniques can create a delightful particularly tiny breed of micropig. In addition to making cute tiny pigs, these techniques can create hilarious super-muscular dogs. CRISPR in diagnostic radiology: Traditional radiology will play an important role in many aspects of CRISPR therapy, from monitoring/validating treatment response to diagnosing genetic syndromes amenable to treatment. For this discussion, however, we will focus on a specialized branch of radiology that will offer unique value in the CRISPR era: radiogenomics. In radiogenomics, imaging is quantitatively data-mined to detect correlations with genomic patterns — that is, we extract imaging features that tell us something about the patient’s genome. Take this fascinating work from RSNA 2017: They were able to use machine learning to analyze MRI and predict, with 95% accuracy, the KRAS mutation status of a tumor. They accomplished this using fancy imaging features that humans poorly understand, including coarseness, eccentricity, shade, and gray-level co-occurrence matrix (GLCM) standard deviation (which they swear is a real thing). This is a mind-blowingly cool aspect of machine learning in application to radiology: we can use these computational techniques to extract features that are imperceptible to the human eye, deriving value from images that would otherwise be invisible. In effect, machine learning gives us image interpretation superpowers: The sorts of things we will do when radiogenomic techniques become commonplace. In addition to being so cool I freak out and tell my mom about it, the ability to use machine learning to derive genetic information from an image has significant implications for CRISPR therapy. We can use CRISPR to kill tumors by deleting certain genes or inserting genes making them self destruct, potentially curing cancer. To do so, however, we need information about the cancer genotype such that we can give our CRISPR the correct “guide RNA” to make it target and destroy the tumor. As radiologists perform additional research and develop robust radiogenomic techniques, one can imagine a future in which cancer is imaged, analyzed with an array of machine learning algorithms, and genetically profiled by radiologists in a way that optimizes CRISPR treatment. Radiologists will thus, armed with machine learning, offer patients a powerful tool that is instrumental in curing cancer. CRISPR in interventional radiology: Although diagnostic radiology will allow us to put the right CRISPR in the syringe, we need interventional radiology to properly inject it. While some CRISPR therapies might work through standard injection into a vein — in fact, this is already being done in humans in China — there will be significant value in a more targeted approach. Targeted delivery will be important largely because CRISPR can display an off-target effect, snipping DNA in locations that are distant from the site of interest that share a similar molecular address. This has the potential to cause serious problems: if we inject CRISPR optimized to destroy a kidney tumor into the blood, it will travel to all tissues. It may, for example, travel to the liver, snip an arbitrary benign gene, and cause deregulation of a genetic control mechanism that results in the development of cancer. Interventional radiologists can minimize this risk by delivering CRISPR directly to a tumor, either through a needle within the tumor or by releasing it into a blood vessel feeding the lesion. In addition to treating cancer, IR physicians can use CRISPR to help patients in a number of other ways: Fibroids. Gene therapy has shown significant potential in treating fibroids in animal models. The treatment, which triggers apoptosis through a complex genetic mechanism, was injected directly into fibroids. One can imagine such a technique supplementing uterine fibroid embolization, with interventional radiologists eradicating fibroids through both ischemic and genetic mechanisms. Infection. CRISPR can act as a powerful antibiotic that overcomes drug resistance, and IR physicians could potentially enhance abscess drainage or treat other forms of infection with targeted CRISPR administration. We can use a bone drill for a bone biopsy — why not drill into a bone for the targeted CRISPR treatment of osteomyelitis? Fluid. Radiologists often drain cysts and seromas that repeatedly re-accumulate, with sclerotherapy being of questionable and inconsistent utility. One can imagine a CRISPR super-sclerosant that deactivates genes that facilitate fluid production. Voila, cyst cured, permanently. We spend a lot of time pondering the next “big thing” in radiology that will both fundamentally improve patient care and increase our volume of business. In the not too distant future, that thing might be CRISPR. Dr. Kevin Frederick Seals is a big-time important physician-scientist and *totally* not writing this about himself. He is currently a resident physician in diagnostic radiology at UCLA, and will begin fellowship in interventional radiology at UCSF in July 2018. His research focuses on applications of machine learning to medical imaging, including work in both machine vision and natural language processing, using the 2–3 things he still remembers from engineering school. Feel free to contact the esteemed Dr. Seals on Twitter or at kseals@mednet.ucla.edu if you would like to discuss healthcare, technology, the best hamburger in a particular geographic region, or cute animals such as otters and koala bears.

Hacking the genetic code for fun and profit: The role of A.I.-enabled

64

transforming-biology-for-fun-and-profit-1c54344e3328

2018-03-11

2018-03-11 19:31:36

https://medium.com/s/story/transforming-biology-for-fun-and-profit-1c54344e3328

false

1,091

null

Tech

tech

Tech

142,368

Kevin Seals

Physician/engineer in San Francisco, focused on using technology to improve healthcare. Corgi dad.

c5e52d8f6dbb

kevinsealsmd

210

20

20,181,104

null

0

null

0

null

2018-01-26

2018-01-26 06:49:54

2018-01-26

2018-01-26 07:02:44

1

false

en

2018-01-26

2018-01-26 07:02:44

0

1c55380b25fb

0.603774

1

0

Infermedica improves the diagnostic process using the most advanced reasoning technology for preliminary medical diagnosis. Infermedica…

1

AI Diagnostics Infermedica improves the diagnostic process using the most advanced reasoning technology for preliminary medical diagnosis. Infermedica collects, analyzes and uses medical knowledge to ask diagnostic questions to unmask your patients’ conditions. Infermedica’s medical knowledge is stored in a proprietary database of probabilistic relationships between such various diagnostic variables as conditions and symptoms. Our platform becomes smarter with use. We apply machine learning and data extraction algorithms to expand our diagnostic capabilities. Our inference engine utilizes unique and the scalable data modeling approach based on Bayesian Networks. It is designed to resemble the reasoning of expert level human diagnosticians and guarantees an unmatched query response time.

AI Diagnostics

6

ai-diagnostics-1c55380b25fb

2018-06-15

2018-06-15 23:50:53

https://medium.com/s/story/ai-diagnostics-1c55380b25fb

false

107

null

Machine Learning

machine-learning

Machine Learning

51,320

dHealthNetwork

dHealthNetwork is a decentralised social healthcare platform that will change lives on a global scale.

8d2cabf063a8

marketing_61318

11

2

20,181,104

null

0

null

0

null

2018-09-19

2018-09-19 08:20:55

2018-09-19

2018-09-19 08:22:20

1

false

en

2018-09-19

2018-09-19 08:22:20

1

1c58ddab7529

5.566038

0

Recently, it became known that Nasdaq is going to launch a new tool that is based on machine learning for its analytical hub. This tool…

4

Neural Networks In Trading: Goldman Sachs Has Fired 99% of Traders Replacing Them With Robots Recently, it became known that Nasdaq is going to launch a new tool that is based on machine learning for its analytical hub. This tool will process the users data from the social networks, providing institutional investors with a new market analysis tool. However, Nasdaq specialists didn’t give any direct answer for Bitnewstoday.com but only said noncommittally about the “interesting experience” and then referred to a subscription about nondisclosure. Apparently, this can only mean that Nasdaq AI is being tested now. Earlier in March this year, Thomson Reuters media company launched the updated MarketPsych Indices service, which makes the market forecasts by tracking for more than 2000 news sources and 800 social networks. This tool which is based on AI in its analysis uses not only quantitative indicators, but also emotional metrics of traders, like behavioral economics. Thus, the service promises its customers to create more accurate forecasts and choose the best strategy. Traders tools and strategies Traders usually use several technical indicators for the analysis: from resistance and support level, moving average, total trading volume to relative strength index and stochastic indicator. Among the technical indicators, trend lines are also used, which display the information about the characteristic direction of the crypto asset movement to predict the further trend. However, due to the high volatility of the cryptocurrency, these trends are extremely difficult to determine, which is why traders rarely take this indicator into account. But in general, nowadays the technical market analysis is no longer sufficient for the forecast, as it does not take into account the political and socio-economic situation in the world. Mark LYND, a Top 15 ranked influencer and practitioner for Digital Transformation, DLT and Crypto, the influencer for companies like IBM in an exclusive interview for Bitnewstoday.com confirmed that since 2017, the use of AI for database processing has become the main trend in the race for large companies profit: “More companies are looking to see if these AI technologies can help them realize more value from their huge stores of data. Algorithms based on biology, more specifically Artificial Neural Networks (ANNs) and Genetic Algorithms are considered the primary types used for trading analysis, risk measurement and price predictions”. Whether all traders use the neural networks Up today the financial and socio-economic system can not be predicted by a human because of too many introductory factors. And here the AI comes to the arena: it is capable to analyze the variety of life spheres and conduct an independent analysis, which is not worse than any financier’s one. Besides it has one unique advantage — the lack of emotions. But then the question arises: why all the traders do not use AI in their forecasts, and whether the neural network will become a universal tool for traders in the near future? According to Mark LYND ”the tech industry itself is pushing how these technologies can be used to gain a competitive advantage and create new revenue streams, which has the big vendors like IBM with Watson push out more sophisticated integrated tools to help their clients realize these benefits”. Nevertheless, Mr. LYND believes that neural networks at this stage of development are still not ideal and human efforts are being required: “Ultimately the results are still largely approximations for predictions and still require some human judgement or monitoring before utilizing for trades” We conducted a small survey among private traders and made sure that the vast majority are very indirectly familiar in their analysis with AI systems. Nevertheless some noted that they use AI-based services in their forecasts, and Dr. Gordon JONES, the founder of the DLT-project in South Carolina, said: “We focus on the machine learning aspect of all NN and AI where the system is designed to track actions and whether those actions lead to efficiencies or deficiencies and ultimate a positive or negative ROI”. Traders are afraid to be replaced soon Top hedge funds managers earned $1 billion in 2015, that creates a great motivation to replace them with neural networks and reduce the cost of employees who earn $500 per hour. And the company Goldman Sachs went this way by reducing the number of advisers from 600 traders to 2. Now the rest of the work is done by robots using machine learning that give out invaluable algorithms for the company’s customers. It is easy to imagine that many large companies are rapidly following this example to reduce their costs. In an exclusive interview for Bitnewstoday.com Amardeep SINGH, PhD candidate in AI use in Ethereum and at Nasdaq analyst, who has been developing strategies for the introduction of AI in the economy for many years, said that traders are well aware that their place will soon be taken by artificial intelligence and desperately resist to the future changes.”Economic analysis is a complete picture that traders look at in the wider world: events, announcements from the governments, that influence the price of what they are analysing. In trading you have something called technical analysis. They look at things like moving averages, they look at hyperparameter optimization, volatility forecasting, multivariate time series — all this, basically, really really complicated models, that are superimposed upon a moving graph, that they can then kind of make predictions. And technical analysis is also based on human emotion. So, it’s based on how you interpret what the market has done previously. This is why AI is incredibly more powerful, than a human”. How neural networks operate and get trained Thanks to machine learning neural networks allow much more efficient construction of nonlinear relationships in comparison with linear methods of statistics, such as linear regression, autoregression and linear discriminant. Any analyst that uses the technical analysis will make more successful forecast that based on the preliminary work of AI systems. Nonlinear mapping and data visualization by neural networks in the space of fewer nonlinear principal components optimize their processing. The main advantage of neural networks is their ability to simulate the behavior of the economy on the basis of social overtones comparing to existing algorithms that are based on the specified and determined parameters. However, in this regard, not all the experts agree that neural networks have to be totally autonomous in the decisions that are based on their forecasts. Thus, the co-founder of the DLT platform in India Bibin BABU believes that the neural network is not very successful in forecasting the digital currency. “The problem is that most NN (neural networks) only has access to few types of data from within the market, and not much from outside. Thus, the forces of crowd psychology, consumer confidence, good or bad headlines, political or regulatory decisions, the size of the Cryptocurrency network, the number of users, and merchant adoption are not completely visible to it. It may simply be the case that internal market data alone is not sufficient to make any kind of long term predictions, or foresee short term price fluctuations. Mark LYND believes that the ability to use neural networks in trading is a kind of art, as the determining factor for qualitative analysis is the ability to determine the context and volume of data and their relevance to the type of results required. “Simply at a high-level it depends on whether you need a searching or sorting and then determining the context and the quality of the data. It is actually more complicated than that, but at a high-level it brings some clarity. You use mathematical functions called neurons that takes several numbers as inputs and then use a linear combination formula to multiply by the corresponding weights and then sum it up. Essentially, these networks transform data until they can classify it as an output. This is done by a bunch of connected neurons that create outputs that are then used iteratively as inputs for other neurons, hence the network. Then a loss function is used to determine how good the neural network is for solving/working out a certain task or problem. Initially, it will not provide strong results, but over time and use practices it will provide higher accuracy. In the next article we will talk about predictions that experts make about the future of trading with the neural networks: whether the AI will manage the market on its own or the final decisions will still remain for traders. Read about it and other exciting issues in the next article. READ MORE

Neural Networks In Trading: Goldman Sachs Has Fired 99% of Traders Replacing Them With Robots

0

neural-networks-in-trading-goldman-sachs-has-fired-99-of-traders-replacing-them-with-robots-1c58ddab7529

2018-09-21

2018-09-21 08:23:28

https://medium.com/s/story/neural-networks-in-trading-goldman-sachs-has-fired-99-of-traders-replacing-them-with-robots-1c58ddab7529

false

1,422

null

Machine Learning

machine-learning

Machine Learning

51,320

BitnewsToday

Everything that you need to know about cryptocurrencies

2f3952862f47

bitnewstoday

9

1

20,181,104

null

0

null

0

59ef541b9a39

2018-08-03

2018-08-03 10:19:48

2018-08-11

2018-08-11 05:12:18

10

false

th

2018-10-12

2018-10-12 16:21:22

7

1c58e516bfcd

3.208491

4

0

วันนี้จะมาพูดถึงเรื่องที่กำลังร้อนแรงในขณะนี้ (ซึ่งผมก็ตามเทรนนี้ด้วย)นั้นคือ

5

เริ่มเรียน Machine/Deep Learning 0–100 (Introduction) About Machine Learning วันนี้จะมาพูดถึงเรื่องที่กำลังร้อนแรงในขณะนี้ (ซึ่งผมก็ตามเทรนนี้ด้วย)นั้นคือ Data Science ตัวอย่างเช่น Dota 2 OpenAI VS Human https://youtu.be/eaBYhLttETw รถยนต์โดย tesla https://www.youtube.com/watch?v=0NtdZNWUBik แต่ Data Science เนี่ยมันกว้างมาก(ก ไก่ อีก10ตัว) แบ่งงานย่อยออกมาได้อีกหลายอย่าง เช่น Data engineer , Data Analysis , Research https://www.youtube.com/watch?v=xC-c7E5PK0Y ในวิดีโอจะพูดถึงว่า Data sci ความจริงแล้วคืออะไรและทำอะไรกันแน่ ? ในวิดีโอนี้จะเห็นได้ว่า Data sci ในบริษัทเล็กๆ วิเคราะห์ทุกอย่าง ให้บริษัทเติบโตขึ้น (ในไทยส่วนมากเป็นแบบนี้) Data sci ในบริษัทกลางๆ วิเคราะห์ข้อมูล สร้าง Model (Wongnai,Line) Data sci ในบริษัทใหญ่ๆ ทำงานวิจัย (งานวิจัย ป.เอก ป.โท งานวิจัยสำหรับบริษัท) ถ้าให้สิ่งที่เกี่ยวกับ Data Sci เป็นปิรามิด Machine Learning จะอยู่กลางๆค่อนไปทางด้านบน เพราะมันจะไม่ค่อยเกี่ยวกับ Deep Learning ไม่ค่อยเกี่ยวกับ Analysis หรือ Visualization แต่ทั้ง 3 อย่างจะเกี่ยวกันและต่อยอดกันได้ (ถ้าคุณสนใจนะ) แต่สำหรับคอร์สนี้เน้นไปที่ Machine Learning อย่างเดียวเท่านั้น (OpenAI ไม่เกี่ยวแน่นอนจ้า) จะเน้นไปที่การคาดเดา ทายผล ทำนาย คำนวณ ไม่มีการสร้างไอ้หมอนี้แน่นอนจ้าาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาาา Requirement สำหรับการเรียน Machine Learning : ความรู้ภาษา Python Library ที่ใช้ Numpy , Matplotlib , Sklearn (ติดตั้ง Anacoda ก็จะมีทั้งหมดนี้ให้เลย) 2. ใช้ Jupyter Notebook เป็น (ก็อยู่ใน Anacoda อีกนั้นเหละ) หรือจะใช้ Spyder,Vscode ก็ได้ Vscode ก็สามารถใช้แทนได้ แต่ jupyter เอาไปขึ้น github แล้วจะสวยกว่า 3. ความรู้พื้นฐานด้านคณิตศาสตร์ สถิติ (ความน่าจะเป็น บวก ลบ คูณหาร ดูสมการ) 4. หนังสือช่วยในการเรียนที่ผมใช้คือ ….. OReilly Hands On Machine Learning with Scikit Learn and TensorFlow 1491962291 5. SQL ผมจะพยายามใช้ data set พวก .csv และมีการใช้ฐานข้อมูล SQL สิ่งที่ผมจะสอน Introduction 1 บท Supervised Classification(k-NN+IRIS 1 บท) Under Fitting / Over Fitting ปัญหาที่มองไม่เห็นแต่สัมผัสได้ว่ามี….. Evaluate Model (Precision,Recall,F1 score) Supervised Classification(Linear Regression+GridSearchCV) Supervised Classification(Polynomial,Gradient Decent) Supervised Classification(Logistic Regression+IRIS) Supervised Classification(SVM ) Supervised Classification(Decision tree) Supervised Classification(Random Forest) Neural Network [Perceptron-Activation-Backpropagation] TensorFlow -> CNN (Image Classification) TensorFlow -> RNN (NLP) ยังไงก็ฝากติดตามด้วยนะครับ เอาละ เข้าเรื่องกันเลย ! Machine Learning คืออะไร ? แปลตรงๆตัวเลยว่า “การเรียนรู้ของเครื่องจักร” โดยเครื่องจักรในที่นี้หมายถึง “Program Computer” นั้นเอง โดยเราต้องเลือกอัลกอริทึมให้คอมได้เรียนร ู้โดยแต่ละอันจะมีข้อดี ข้อเสีย แตกต่างกัน ใช้งานต่างกัน ผลลัพธ์ต่างกัน….. โดยเราจะนำปกติถ้าเราจะทำโปรแกรมอะไรสักอย่าง เช่น คำนวณจำนวนประชากันในอีก 5 ปี เราอาจจะเขียนสมการ ใส่ input แล้วคำนวณออกมาก็จะได้ประชากรในอีก 5 ปี ข้างหน้า แล้วถ้าเป็น Machine Learning ละ ? เราใส่ input (จำนวนประชากร 15 ปีที่ผ่านมา) และ output ว่าเพิ่มขึ้นกี่เปอร์เซ็น เจ้า computer ของเรา ก็จะคำนวณออกมาให้เราเลย ! (เลือกอัลกอริทึมให้ถูกนะครับ ถ้าไปใช้ classification ยังไงก็ไม่ถูก) Machine Learning จะมีคำศัพท์ที่คุณต้องเข้าใจอยู่นิดหน่อย Machine Learning แบบ Supervised Training Set : ชุดข้อมูลต้นฉบับสำหรับ Machine Learning โดยฝึกให้ผลลัพธ์ออกมาเป็นไปตามชุดข้อมูลต้นฉบับ (หากชุดข้อมูลตั้งต้นผิด ผลลัพธ์ก็จะผิด) เช่น ข้อมูลคนไข้ , ข้อมูลคนที่ชื้อของผ่าน Lazada (เขาไม่ได้จ่ายให้ผมพูด…..) Label,Non Label : เป็นตัวบ่งบอกว่าข้อมูลที่ให้ฝึกเป็นอะไร โดย Label ใช้กับ ML แบบมีผู้ช่วยสอน กับ Non Label ใช้กับแบบไม่มีผู้ช่วยสอน เช่น ข้อมูลคนไข้โดยมี Label บอกว่าคนไข้คนนี้ป่วยหรือไม่ , ข้อมูลคนที่ชื้อของผ่าน Lazada โดย Label บอกว่าเขาชื้อของใช้หรือของกิน Feature : ลักษณะที่เด่นๆของ Training set หรือ Test set เช่น ข้อมูลคนไข้ที่ป่วยเป็นไข้หวัดใหญ่ โดยลักษณะสำคัญคือ ไข้ขึ้นสูงกว่า 35 องศา , คนที่ชื้อของผ่าน Lazada โดยเขาชื้อของใช้ โดยลักษณะสำคัญคือ ของชิ้นที่ชื้อกำลังลดราคา Machine Learning Algorithm : จุดสำคัญสำหรับ ML คอร์สนี้คืออันนี้เหละครับมันคือตัวที่เอา Training Set , Label Non Label , Feature เอามาคำนวณ ประมวลผล (เอาไปเทรน)โดยมีหลายตัวมากครับ โดยมี2แบบคือ Regression กับ Classification เล่นมีอัลกอริทึมอีกมากมายครับ เช่น k-NN , Linear Regression , Logistic Regression , Polynomial ฯลฯ Predict Model : หลังจากได้ ML Algo มาแล้ว เราจะได้สิ่งที่เรียกว่า Predict Model ขึ้นมา โดย Model นี้จะเอาไว้มาทดสอบ/ใช้งาน โดยเราหวังว่า ผลลัพธ์ที่ออกมาจะเป็นไปตามที่เราคาดหวังไว้ (หวังว่านะ…..) Test Set : ชุดข้อมูลเอาไว้ทดสอบ โดย Test Set ไม่ควรเอาไปใช้ร่วมกับ Training Setเพราะถ้าทำแบบนั้น มันคือ เฉลย !! (เหมือนถามว่า 3+3 ได้เท่าไร ? แต่ในนั้นมีคำตอบให้แล้ว แล้วจะสร้าง ML ขึ้นมาทำไมฟะ !)และจะเกิดในเรื่องของ Model Over fitting กับข้อมูลชุดนั้นๆด้วย [อธิบายเพิ่มในบท2นะครับ] Expected Label or Value : Output ของการทำนาย โดยเราหวังว่ามันจะออกมาถูก โดยบางครั้งเราสามารถมองเห็นผลลัพธ์ด้วยตาเปล่า เช่น ทำนายว่าเลขนี้คือเลขอะไร ? แต่บางอย่างเราไม่สามารถทำการตรวจสอบผลลัพธ์ด้วยตาเปล่า ทำให้เราต้องมีการทดสอบขึ้นมา โดยเรียกว่า E-Test โดยเอา Output มาเทียบ Training Set หรือ Good global โดยค่าที่ผมใช้ประจำคือ Precision,Recall,F1 score [สอนในบท3] ในที่นี้ ผมจะสอนอยู่ 2 แบบใหญ่ๆคือ แบบมีผู้ช่วย(Supervised)และแบบไม่มีผู้ช่วย(Unsupervised) Machine Learning Type แบบที่ 1 แบบมีผู้ช่วยสอน (Supervised) Supervised โดยในแบบนี้ เราจะมีสิ่งที่เรียกว่า Label มาช่วยบอกว่าข้อมูลชนิดนี้คืออะไร ? หรือจำแนกประเภทของมัน โดยจะแบ่งออกมาได้อีก 2 ประเภทคือ Classification และ Regression Classification คือการจำแนกประเภท เช่น มีสัตว์อยู่1ตัว อยากจะแยกมันออกมาว่ามันคือตัวอะไร กระทิง หรือ หมี Regression คือ การทำนายออกมาเป็นตัวเลข เช่น จำนวนประชากรจำนวน n คน จะมีรายได้ประมาณ x บาท แบบที่ 2 แบบไม่มีผู้ช่วยสอน (Unsupervised) คิดเองเอ่อเอง โดยในแบบนี้ผมเรียกได้ว่า คิดเองเอ่อเอง (แบบนี้ก็มี) แต่มันมีหลักการครับ เราทำเหมือน Supervised ทุกอย่างเลย แต่ไม่ให้ Label กับมัน ซึ่งเดียวมันจะจับ pattern ได้ด้วยตัวมันเอง สำหรับ Inro คอร์สนี้ก็จบเพียงเท่านี้ครับ เป็นการรู้จักคำศัพท์ที่ใช้ในงาน Data sci และสอนเกี่ยวกับประเภทต่างๆ ในบทความหน้า ผมจะพูดคำศัพท์เทคนิคมากขึ้น ขอให้จดจำคำศัพท์พวกนี้ด้วยนะครับ data set,training set,test set,model,ML บทความถัดไป ผมจะสอนเรื่อง k-NN กับ sklearn ขอให้ติดตั้ง Anacoda ด้วยนะ ผมเพราะผมจะใช้ jupyter notebook (จะใช้อย่างอื่นก็ได้ครับ) เจอกันเมื่อผมมีไฟในการเขียบทความถัดไป แล้วเจอกันใหม่ครับ See you .

เริ่มเรียน Machine/Deep Learning 0–100 (Introduction)

8

เริ่มเรียน-machine-learning-0-100-introduction-1c58e516bfcd

2018-10-12

2018-10-12 16:21:22

https://medium.com/s/story/เริ่มเรียน-machine-learning-0-100-introduction-1c58e516bfcd

false

519

Life style of Programmer , Web App , Machine Learning and new trend of technology

null

nping.loaf

null

mmp-li

null

mmp-li

PROGRAMMER,LIFESTYLE,MACHINE LEARNING,TECHNOLOGY,EDUCATION

null

Machine Learning

machine-learning

Machine Learning

51,320

Mr.P L

Life Style & IoT & Web App และ Data Sci(ML,NN) บทความเกี่ยวกับโปรแกรมเมอร์ ,Life Style , Data Science ,IoT(Node-RED) and Web App(C#)

324f6d3f124f

pingloaf

78

1

20,181,104

null

0

null

0

null

2018-07-16

2018-07-16 14:57:39

2018-07-16

2018-07-16 17:00:58

2

false

en

2018-07-16

2018-07-16 17:00:58

5

1c5901644287

2.009748

3

0

For the past 300 years, the art market has been heavily controlled by just a wealthy few, which has invited constant manipulation and shady…

5

Image from ARTSTAQ. Value Protocol’s partnership and application with ARTSTAQ — a decentralized market for art and collectibles For the past 300 years, the art market has been heavily controlled by just a wealthy few, which has invited constant manipulation and shady behavior. ARTSTAQ replaces this old and ineffective art market model with one from the 21st century — empowering creativity by allowing anyone to invest in art with transparency and a fair market value. Value Protocol and ARTSTAQ’s partnership paves the way for art to be accessed in a completely new, trusted and decentralized manner by democratizing the access to the $2.7 trillion market of art, wine, old cars, watches, stamps, coins and other collectibles. With Value Protocol’s technology and robost infrastructure for marketplaces with physical assets, ARTSTAQ can rely on any physical item’s provenance record on the blockchain — being guaranteed that every item is authentic upon creation or registration and verifiable at any time with just a smartphone. About ARTSTAQ: ARTSTAQ is the new global NASDAQ-like Art Exchange model that is entirely based on the principles of capital markets. The art exchange model allows one to trade art like stocks and fully ensures that traders are able to base their investment decisions on the most up-to-date information. ARTSTAQ is considered the new standard of transparency in the art market because of its standardized trading principles, independent rating, live market valuation and open market data. They are developing ARTSTAQ’s Rating System by using cutting-edge technology, mathematical models and analytical tools to predict the art market’s behavior. They are also developing a shares trading platform where anyone can own a piece of Picasso and other artists from just $1 while protecting the item’s authenticity, ownership and security. About Value Protocol: Value Protocol is the authenticity protocol and the first blockchain-based infrastructure as a service to connect physical assets to the blockchain with A.I at all times — a real use case which allows businesses to build their own decentralized markets for physical assets. Value Protocol sets key market principles which are immutable & predetermined by smart contracts. It features platforms and modules that businesses can use for shares trading, authenticity, protection, provenance, insurance, lending, shipping and more. By utilizing artificial intelligence and computer vision for fingerprinting any surface’s unique material structure, Value Protocol can hash it into the blockchain with just a smartphone, making it affordable, convenient and finally tamper-proof. For more information, read our white-paper or our one-pager. To stay updated, follow us on Facebook. For partnerships or questions, please contact us at team@valueprotocol.org.

Value Protocol’s partnership and application with ARTSTAQ — a decentralized market for art and…

115

value-protocols-partnership-and-application-with-artstaq-a-decentralized-market-for-art-and-1c5901644287

2018-07-16

2018-07-16 17:00:59

https://medium.com/s/story/value-protocols-partnership-and-application-with-artstaq-a-decentralized-market-for-art-and-1c5901644287

false

431

null

Blockchain

blockchain

Blockchain

265,164

Veracity Protocol

The decentralized infrastructure for the lifecycle of things (LoT) to secure end-to-end traceability, data veracity and resource efficiency in supply chains.

ffa0d24088f8

veracityprotocl

18

5

20,181,104

null

0

null

0

d8ca8bdc1cd3

2018-01-17

2018-01-17 06:03:25

2018-01-17

2018-01-17 13:46:01

1

false

en

2018-01-17

2018-01-17 19:49:35

2

1c5b04c744bf

4.10566

14

0

Two words were on my mind. As I handed back my badge, let the double-pane glass door lock behind me and walked out into the Bay’s fog…

5

Building Technology from the Inside Out “You will either step forward into growth or you will step back into safety” — A. Maslow // Photo: Philippe Ramette Two words were on my mind. As I handed back my badge, let the double-pane glass door lock behind me and walked out into the Bay’s fog. Those two pesky words described why I’d joined Google, and why I left. “Search on.” My name is Cristina. I’m an average American millennial. As such, I interact with technology over 2,500 times a day. That’s not good or bad, it’s just a fact. These glances, taps and swipes absorb 10 hours of each day, 152 days each year, or roughly 1/3 of our lives. Intuitively, we know we use technology often, regardless of whether you personally buy that 2,500 number or not. As an ex-Googler, ex-yoga teacher, and Yale sociologist, I often overthink this topic: How does technology influence our mental state? Our use of smartphones has been shown to neurologically change our brains. As a designer of user engagement at Google, I can informally attest to this. By tweaking some prompts and colors in Google’s mobile search app, we influenced many to try speaking their questions out loud to Google, instead of typing in that eponymous search bar. BJ Fogg, a Behavioral Psychologist at the Stanford Persuasive Tech Lab, has more formally proven the persuasiveness of technology in academic contexts. The summary is this: Triggers, like the notifications you get from Instagram or Facebook, themselves don’t release hits of pleasure-inducing dopamine, but they can trigger us to start new behavior routines down the road that will. That new notification you received won’t necessarily give you a hit of dopamine today, but if you take action on it, it will likely influence you to create a new habit down the road. A habit, which by its very nature, is a behavior hard-wired by the chemicals of your brain circuitry. The smart product designers of today are designing your habits of tomorrow, through simple prompts, designs, and well-timed messages. It’s on a neurological level that technologists, scientists, and psychologists have begun to operate, myself included. Technologists can help you to establish new habits — or cause you to — depending on how you want to look at it. Call this process habitual, addictive, influential — what if we settled on the word choice? Once you know how technology works on you psychologically, it becomes your choice in how you want to use it. Or if you’re a technologist yourself, your choice in how you want to make it. Yet sometimes we forget. Technology is — a choice. We choose to use it. And we choose to continue to make it. I’m not here to judge your use of technology, or any technology for that matter. As humans, we’ve made technologies for the last 200,000 years, many of which have been extremely powerful tools. Technologies have always stirred controversy and been used for both good and bad. That is historical fact, and likely not going to change. Sure, our technological tools are more complex today than in the fire-and-flint days, but the premise hasn’t changed. We continue to design new tools to help us with what we need. So — what is it we all need help with today? If you suspected that we named our company “Maslo” in a nod to Abraham Maslow, you were right. Maslow’s hierarchy of needs, his most well-known contribution to psychology, lays out the range of “needs” we have as humans to live fully satisfying lives. Physiological and Safety needs make up the base — things like food, water, and shelter. These are required for us to reach a state of generally stable “aliveness”. But what then? We’re alive, but we’re not necessarily happy. To grow into happiness, we need things that aren’t basic to our survival — needs for Belonging, for Esteem, and for Self-Actualization. It’s not necessarily enough to have the basics covered to have a happy life, as much as the zen yogi within me wants to try. Our motivation pushes us further. We’re motivated to make friends, get to know people intimately, identify with certain groups and use these allegiances to describe ourselves with symbolic labels. We do things that help us feel accomplished, help us feel that we’ve contributed to society, to give us that all around feeling of being content and satisfied. Social media today plays off our higher order need for belonging and esteem. So does millennial slang. FOMO is a fear of missing out, of being estranged, of not belonging. But fulfillment is ultimately an inside job, and we all must start shutting out the chorus in order to find those things we so crave. Here at Maslo, we believe in helping each other grow. We believe in helping people be more than just alive and breathing. We believe in defining meaning, choosing belonging, and growing into identities that are bigger versions of who we are now. These things aren’t stagnant: they take work, and they will shift over time. What likely isn’t going anywhere is our deep reliance on technology. So if we’re going to share 2,500 moments with it a day — arguably the closest physical relationship we have — let’s use it to help us in our searches for meaning and fulfillment. Searches that lie way beyond those for takeout, Uber, or the occasional Airbnb. Search on, indeed. It’s time for technology to mature. When it engages with us on deeper human topics, we’ll know it has. That’s what we’re building at Maslo: technology that grapples with the existential and that understands the psychological, because our generation will mature hand-in-hand with technology. If it doesn’t relate with us on these levels, we won’t either. So instead of stigmatizing those philosophical, psychological, and existential questions as cliche, our technology dives right in to help us answer the meatiest and most perplexing questions for ourselves. What do we want to do with our lives? How do we identify? Where do we belong? What makes us happy? These are questions that no app, search engine or virtual assistant will ever be able to answer for us. But they are questions that a companion could ask of us.

Building Technology from the Inside Out

269

building-technology-from-the-inside-out-1c5b04c744bf

2018-05-11

2018-05-11 18:37:15

https://medium.com/s/story/building-technology-from-the-inside-out-1c5b04c744bf

false

1,035

You want to be your best self. We build technology to help you get there. Tips from friends at Maslo.

null

heymaslo

null

Your Virtual Self

founders@maslo.ai

maslo

ARTIFICIAL INTELLIGENCE,PSYCHOLOGY,FUTURE TECHNOLOGY,STARTUP

heymaslo

Psychology

psychology

Psychology

49,261

Cristina Poindexter

Cofounder @ Maslo. Technologist. Human.

245156559600

virtualcristina

201

211

20,181,104

null

0

null

0

null

2017-10-13

2017-10-13 12:26:59

2017-10-13

2017-10-13 13:07:04

1

false

en

2017-10-13

2017-10-13 21:59:00

6

1c5bab8a1a68

4.509434

36

5

0

In recent years, Artificial Intelligence (AI) researchers have finally cracked problems that they have worked on for decades, from the…

4

How Artificial Intelligence can support blockchain applications like Telcoin In recent years, Artificial Intelligence (AI) researchers have finally cracked problems that they have worked on for decades, from the ancient game of Go to human-level speech recognition. A key piece was the ability to gather and learn from mountains of data, which pulled error rates past the success line. AI is now applied to almost every possible domain, ranging from healthcare to robotics. However, there’s one domain where AI has not made a significant breakthrough yet. With the recent interest in cryptocurrencies, the blockchain has emerged as a possible alternative to the existing banking system. AI has been used successfully in finance up to now. So why not combine AI with cryptocurrencies, such as Telcoin, and add value for end users? Blockchain technology by itself is just a big decentralized and immutable distributed ledger. This shared control architecture encourages data sharing. Indeed, every user can access and read the records of this public ledger. If many transactions are processed by the network, data will accumulate over time. This means big data. And if we have big data, AI can kick in. For the case of Telcoin, we listed at least four applications where AI can add value. Keep in mind that there are probably a lot more. Here they are: Deal with anti money laundering and fraud management Mitigate the forex risk for Telcoin, as a liquidity provider Manage volatility in a cryptocurrency world Automatically tune the remittance threshold for which customers have to go through to complete the transfer Deal with anti money laundering and fraud management Let’s consider a malicious trader doing price manipulation by sending a lot of resting orders in the order book, then canceling all of them when the market shifts. By pulling the market down, he can buy coins at a better price than other people. This is reported as market abuse and is severely prosecuted in most countries. AI is great at pattern detection and this trader can be easily detected with advanced AI systems. Let’s consider another case where an unlucky customer has lost his phone while traveling abroad. A malicious character finds it and starts transferring coins. Because blockchain is here, we have access to large amounts of data and AI can understand the standard behaviors of the customers which can therefore detect anomalies. For example, this is the first time that money is sent at night from another country. Mitigate the forex risk for Telcoin, as a liquidity provider In the case when a mobile operator decides to buy or sell coins to/from Telcoin, an appropriate basic hedging forex strategy has to be undertaken in order to provide liquidity. Because of the inherent risk, Telcoin has to increase the spread between the bid and the ask in order to slightly mitigate this risk. With the recent advances of AI in quantitative finance, efficient hedging strategies have been proposed and successfully applied to the forex market. One direct advantage of using AI is a lower forex risk and a lower spread, benefiting directly the mobile operators willing to buy and sell Telcoins. Manage Cryptocurrency Volatility Cryptocurrency is a very young and promising area. People using cryptocurrencies can be referred as the pioneers of this new and disrupting technology. Cryptocurrencies are still in their infancy and suffer from price volatility. We all know that any currency needs to be stable in order to be used as a trusted medium of exchange. The more prices fluctuate, the more ordinary people will shy away from using the coins for everyday transactions. We also know that, with a truly stable currency, on the other hand, you can have currency conversion, remittance, ATM withdrawals, and other financial services with lower fees than fiat systems. In other words, it can be used as intended — as money. At Telcoin, we are committed to build a trusted environment around the coins we are going to issue. We are aware that our users will be concerned by the volatility of the prices. Patience is a virtue and we are utterly sure that cryptocurrency volatility will be dramatically reduced as blockchain gets more popular. However, sometimes technology can help. As a matter of fact, AI can actually help reduce the inherent volatility of cryptocurrencies. Some papers related to stock volatility modeling have been published recently. AI was shown to perform much better than traditional models, when massive data sets were available. Cryptocurrencies have been around for quite a while now and such data sets are at hand. By analyzing prices, volumes and news on different exchanges, AI can price future contracts fairly. The purpose of those contracts is to freeze the price of the currency for a fixed amount of a time. Users are proposed to buy this contract to cancel the volatility of the currency for a certain period of time, for a small premium. Automatically tune the remittance threshold Lastly, let’s consider the case when a customer A wants to send a large amount of money. Is authentication required? Traditional systems are based on rules and are therefore static by definition. For example, if this customer A wants to initiate a wire, equal to 200 percent of the average amount he usually sends without authentication, then additional checks are required. Those rules are of course easy to understand but can sometimes lead to undesired behaviors, such as having to authenticate when transferring only one dollar. To this archaic system, we propose the dynamic threshold idea based on the fact that every person is different and should be treated differently. Telco operators know their customers better than their bankers, though KYC compliance. Coupled with telecom service usage patterns, AI can greatly enhance the user experience while maintaining a very high level of safety. Because Telcoin is based on the Ethereum blockchain, most of the AI services exposed here will require an Oracle. In the context of blockchains and smart contracts, an Oracle is an agent that finds and verifies real-world occurrences and submits this information to a blockchain to be used by smart contracts. In our AI context, the oracle can provide external data feeds like price volatility forecasts, dynamic remittance thresholds per customer. Throughout this blog post, we have described how AI can help cryptocurrency, especially Telcoin. We are convinced that the combination of AI and blockchain is explosive and both domains can really take advantage of one another. Coupling those two high profile technologies with telecom data can help realize the long-standing promise of a global cryptocurrency that is both usable and safe. Telcoin website: http://www.telco.in/ Sources: Lee, Yan Nee. “This company wants to grow A.I. by using blockchain.” CNBC. Sept. 17, 2017. https://www.cnbc.com/2017/09/17/hanson-robotics-singularitynet-integrate-blockchain-and-artificial-intelligence.html McConaghy, Trent. “Blockchains for Artificial Intelligence.” BigchainDB. Jan 3, 2017. https://blog.bigchaindb.com/blockchains-for-artificial-intelligence-ec63b0284984

How Artificial Intelligence can support blockchain applications like Telcoin

203

how-artificial-intelligence-can-support-blockchain-applications-like-telcoin-1c5bab8a1a68

2018-05-23

2018-05-23 02:39:10

https://medium.com/s/story/how-artificial-intelligence-can-support-blockchain-applications-like-telcoin-1c5bab8a1a68

false

1,142

null

Telcoin

telcoin

Telcoin

9

Telcoin

Financial inclusion for a mobile world

9517e1c9dc5e

telcoin

770

51

20,181,104

null

0

null

0

a00e37810120

2017-11-07

2017-11-07 01:57:43

2017-11-07

2017-11-07 01:58:30

1

false

en

2017-11-07

2017-11-07 02:02:13

7

1c5be58252bc

3.818868

0

As part of the NaNoWriMo, I’m going to write a post every day of November for the rest of the month. Some of these will be technical and…

4

Hayek and the A.I. Dream Frederich August von Hayek As part of the NaNoWriMo, I’m going to write a post every day of November for the rest of the month. Some of these will be technical and others more philosophical. Hope you all enjoy! In 1945, there was a proposal to institute the Central Pricing Board. The idea was sound: why not try and smooth out the irregularities involved in commodities pricing by allowing one centralized board to decide prices? The problem, as F.A. Hayek noted in his seminal essay “The Use of Knowledge in a Free Society”, was that this is fundamentally impossible because it assumes that one person, or a small group of people, can know everything about about every aspect of society at all times. The essay has since become quite influential, influencing everything from the modern Libertarian movement to Wikipedia. A lot has been said about this work that I won’t say here (books have been written about this small essay), but one point that always stood out to me is that people can know a lot about a very small specific topic, but they can never have a high depth of knowledge about all topics. This explains why Ph.D.s specialize; they are out to learn a lot about something very tiny and specific. The essay itself mentions how pricing is best determined by those closest to production because they understand the methods best. I know many who never read the news on the basis of a journalist can never know as much as a physcisist about physics (or whatever other topic is particularly relevant). The implications of Hayek’s essay can be extended to the data science/machine learning/artificial intelligence community (note: I use them interchangeably in this post). There’s a strong movement amongst AI enthusiasts to bring about General Intelligence. The idea is that we can build a machine learning algorithm or agent, that is as smart as humans or is at least generalizable to any situation (the specifics depend on who you talk to). This idea is popular with figures like Ray Kurzweil of the Singularity University, who believes we will one day upload our conscience to a computer network and live forever, and Nick Bostrom, who thinks we’re doomed to make paperclips forever. Its found strong support from major entrepreneurs like Elon Musk and Larry Page as well. But it suffers from the same issues that a central planing board would suffer. We can’t ever know the all of everything. That’s period, point-blank ever. There will always be some unknown, some portion of our reality we just don’t get. To say that we can fully understand everything is to say that we can fit all of reality in our heads. The AI god believers would have you think this is either possible of us or that we can build a machine that does this. I don’t believe that the former for the reasons laid out in Hayek’s essay and because its self-evident, to me at least. You really should read it as I’m not as smart as Hayek was and can’t explain it as comprehensively as he can. As for being self-evident, I think the whole internet is evidence of this. Arm chair stock analysts who live in their parent’s basements but claim to know more than billionaire hedge fund managers abound in the comments section of most financial publications. On the possibility of building a machine that can do this, we as human beings have always succeed in building tools when the problem space as specified very clearly and limited in some way. Let me use an example from the Industrial era to show you what I mean by that. When you wanted to build a machine, say a grain thresher, you had to know how to fully define the problem. In this case, you had a specific purpose for that machine to accomplish, that of separate grain from the stalk. This was a way of limiting your focus, similar to how human beings work with heuristics by ignoring unimportant events. By narrowing down, you could start to think about all the particularities involved, which are often much greater than is seen at first glance. In fact, anybody whose been involved with the hard sciences know this to be true. You simplify the problem space by either taking assumptions or by narrowing the field of possible conditions. The former helps when you’re in a new area, but the latter is ultimately necesary as topics grow in their complexity. By contrast, General Intelligence does nothing to simplify the problem space. When you do start to limit the problem space, the results are indeed extraordinary. This is one of the reasons why machine learning works so well in games: we can very specifically set down rules to play by and define the space of all possible moves. Its also why computer vision works so well if we have labeled data, as we can specify the exact set of examples to work against. Thus we can build AlphaGo to steam roll the greatest players of our era, but that same AI can’t make me breakfast in the morning or even comprehend how to do that. This all has implications in how we approach problems. We try to think of a fits-all solution when fundmentally this is impossible. Each problem, even when using the same dataset, must be approached as if its solution was separate and unique. This cannot be understated, as sometimes meaningful predictions might only require a linear regresion while other times it might require a complex ensemble. This AI-as-messiah that seems increasingly prevalent misses this problem and leads us down a path of wasted effort. I hope you all avoid this.

Hayek and the A.I. Dream

0

hayek-and-ai-dream-1c5be58252bc

2018-02-07

2018-02-07 10:36:27

https://medium.com/s/story/hayek-and-ai-dream-1c5be58252bc

false

959

Seeking to understand the new era of Computing being explored via Data Science, Machine Learning, & Artificial Intelligence

null

Singular Distillation

null

singular-distillation

MACHINE LEARNING,DATA SCIENCE,BIG DATA

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Vincent Alexander Saulys

Machine Learning & Data Science Extraordinaire. Senior Data Scientist at The Bank of New York Mellon. My views do not reflect my employer and are solely my own.

bb115807b085

vasaulys

164

528

20,181,104

null

0

null

0

null

2017-10-12

2017-10-12 13:29:45

2017-10-14

2017-10-14 03:32:14

3

false

en

2017-10-14

2017-10-14 05:59:38

3

1c5dcd9bbf86

3.40283

21

0

Today I had a really interesting conversation with Dr. Rachana Bhat of All India Institute of Medical Sciences (AIIMS) on my way to Sochi…

4

Unbiased Inferencing: Tips from a Doctor for AI Scientists Invested in Healthcare Today I had a really interesting conversation with Dr. Rachana Bhat of All India Institute of Medical Sciences (AIIMS) on my way to Sochi, Russia from New Delhi. I was describing how we Machine Learning engineers correlate data from different modalities to arrive at a final diagnosis when she asked me if we follow appropriate blinding protocols. When asked to elaborate, she said, blinding protocols are a set of best practices that must be followed in medical research for the removal of bias. A physician studying one test (modality) should not be knowing the results of the other tests. There is high chance that prior knowledge of the result of a different test can bias his/her diagnosis. Therefore, if a patient is undergoing two different tests, the best practice is to let two different doctors diagnose these different modalities. Finally, a third doctor with enough experience in the domain of the disease under consideration should look into their diagnoses and come to a decision regarding the disease. I found her words interesting and started thinking about a statistical explanation to it. Suppose D ∈ {0,1} is a binary random variable with D=1 denoting presence of a disease. The goal is to determine P(D=1). Let T¹ and T² be the results of two diagnostic tests that the patient undergoes. Let us first try and explain the first step of the bias removal problem: “a physician studying one test (modality) should not be knowing the results of the other tests”. Let us consider the physician who looks at T¹. His task is to estimate P(D=1 | T¹). However, if he knows the results of the second test i.e. T², then instead, he would end up estimating P(D=1 | T¹, T²). By Bayes’ rule we have: Thus the physician’s estimate of P(D=1 | T¹) is biased by a factor of: A closer look at the bias term reveals there can be two cases: P(T² | D=1, T¹) = P(T² | T¹): That is, prior knowledge about the disease does not cause any change in the physician’s interpretation of T² in the light of T¹. In that case, the bias term is equal to 1 which corresponds to the best case that is, the physician’s estimate being unbiased. P(T² | D=1, T¹) ≠P(T² | T¹): This corresponds to the more common case in which prior knowledge about the disease does affect the physician’s interpretation of T² in the light of T¹. For example, if a doctor knows apriori that the patient’s father had a heart attack, he will tend to find signs of mental stress in his EEG signal with more confidence than otherwise. If the value of this term is greater than 1, the physician would over-estimate P(T² | T¹). If it is lesser than 1, (s)he ends up underestimating. Hence, for an unbiased diagnosis, a physician studying one test (modality) should not be knowing the results of the other tests. Now let us consider the second step: “a third doctor with enough experience in the domain of the disease under consideration should look into their diagnoses and come to a decision regarding the disease”. We have: The task of the doctor giving the final estimate of P(D=1) is to work out the above equation with the likelihoods P(D=1 | T¹) and P(D=1 | T²) determined in the previous step and of the priors of the individual modalities: P(T¹) and P(T²). In the context of medical diagnosis, prior probability of a modality corresponds to a score of relevance of the modality for the diagnosis of the disease in question. Assignment of these scores calls for domain expertise. Also this domain expert must not be one of the physicians who diagnosed a modality before. This is because if the physician who diagnosed T² estimates P(T¹) then (s)he would instead end up estimating P(T¹ | T²) due to the prior knowledge of T² which may not necessarily be equal to P(T¹). Hence proved. These insights are extremely important for designing multi-modal prediction systems which are most ubiquitous in Healthcare. Simple things like this tend to be overlooked and this can lead to severe consequences after deployment. Rachana and me are a part of the Indian Delegation to the 19th World Festival of Youth and Students (WFYS) 2017 at Sochi, Russia. This is the first of a series of blogs that I plan to write during this trip which will contain cross-domain insights on Machine Learning practices.

Unbiased Inferencing: Tips from a Doctor for AI Scientists Invested in Healthcare

34

making-unbiased-diagnosis-tips-from-a-doctor-for-ai-scientists-invested-in-healthcare-1c5dcd9bbf86

2018-05-15

2018-05-15 07:17:10

https://medium.com/s/story/making-unbiased-diagnosis-tips-from-a-doctor-for-ai-scientists-invested-in-healthcare-1c5dcd9bbf86

false

756

null

Medicine

medicine

Medicine

14,481

Anirban Santara

Google Ph.D. Fellow in Machine Learning at Indian Institute of Technology Kharagpur, India. Read more about him at http://santara.github.io

b09987ef048e

nrbnsntr

694

104

20,181,104

null

0

null

0

a72e580f87ae

2018-04-27

2018-04-27 01:46:55

2018-05-04

2018-05-04 22:34:41

5

false

en

2018-05-04

2018-05-04 22:34:41

4

1c5fc82f6ffc

6.116352

2

0

Risk management is pivotal when trading on the crypto-currency exchanges. The way you approach this vital task determines the lifespan of…

5

Light Paper: Artificial Intelligence in HyperQuant Ecosystem Risk management is pivotal when trading on the crypto-currency exchanges. The way you approach this vital task determines the lifespan of the utilized trading strategies. The use of Artificial Intelligence (AI) is the most prospective among the possible solutions. That is why we in HyperQuant heavily focus on revolutionary AI approaches. The combined decision is thus taken based on the results of the studies listed below. Let’s go through the main AI directions and methods that HyperQuant favours. Methods that incorporate expert knowledge and are taught on open and proprietary datasets in HyperQuant’s possession. Bayesian estimation of risk distribution. This is an estimator that minimizes the posterior expected value of a loss function. It also maximizes the posterior expectation of a utility function. An alternative way of formulating an estimator within Bayesian statistics is maximum a posteriori estimation. Full stack of actuarial methodology. Actuarial methods are used in insurance. This is a mix of economics-mathematical methods of salary calculation. These are based on the law of large numbers. The methods reflect the mechanism of insurance fund creation and utility is long-term insurance operations that are connected to the population’s lifespan. The goal is to determine the participation share of each insurant in the creation of the insurance fund, i.e. — the salary size. The methodology is based on using the theory of probability, demographics and long-term financial calculations. Insurance operations include the equivalence principle visible through the equality of financial liabilities between an insurer and an insurant. Graphical models. For example, a decision tree. This method of decision making support is based on the use of a tree-like graph. This “tree” is a model of decision making that incorporates potential consequences (the probability of or another event taking place), effectiveness and resource-consumption. In a business process this tree is comprised of a minimal amount of questions that should be answered with a “yes” or a “no”. By answering these questions in sequence — we arrive to the right choice. The methodological advantages of the decision tree lie in it structuring and systemizing the problem with the final verdict based on the logical conclusions. Gaussian systems for processes, multivariate functions and the objects of a sophisticated structure like a set of qualitative indicators. In probability theory and statistics, a Gaussian process is a stochastic process. The stochastic process is a collection of random variables indexed by time or space. In Gaussian process every finite collection of those random variables has a multivariate normal distribution. Thus every finite linear combination of them is normally distributed. The distribution of a Gaussian process is the joint distribution of all those (infinitely many) random variables, and as such, it is a distribution over functions with a continuous domain, e.g. time or space. A machine-learning algorithm that involves a Gaussian process uses “lazy” learning and a measure of the similarity between points (the kernel function) to predict the value for an unseen point from training data. The prediction is not just an estimate for that point, but also has uncertainty information — it is a one-dimensional Gaussian distribution (which is the marginal distribution at that point). For some kernel functions, matrix algebra can be used to calculate the predictions using the technique of kriging. Point processed kernel methods. In machine learning kernel methods are an algorithm class of model analysis. The mots known from them is the support vector method (SVM), It is a whole set of algorithms required for solving the tasks of classification and regression analysis. Based on the fact that object located in N-dimensional space belongs to one of the two classes, the SVM creates a hyperplane with N-1 dimension for all the objects to be in one of the two groups. T he general aim of pattern analysis is in the search and study of general relationship types (for example, clusters, ranging, components, correlations and qualifications) in data sets. For many algorithms solving these tasks — the data in its raw form needs to be changed to the vector interpretation of objects with the help of user’s object map, At the same time — the cored methods only demand the core specified by the user — thus the similarity function for the couples of data points in the unprocessed form. Methods of deep networks learning for the work with categorical objects. Long Short-Term Memory based systems (LSTM). Long short-term memory is a variety of recurrent neural networks architecture.. LSTM-network is unique in the sense that with enough of network elements — it can complete any calculation that a regular computer can take on. For this a corresponding weight matrix is required — it can be considered a programme. Contrary to the traditional recurrent neural networks, TSTM is well-equipped for learning on the tasks of classification, processing and forecast of temporary lines in cases, when important events are divided with time lags of unidentified length and borders. A relative immunity to lengthy time lags gives LSTM and advantage in front of alternative recurrent neural networks, closed mark models and other sequential systems of this kind. Seq2seq for sets models. One of the most popular architectures in machine translations are the sequence to sequence (Seq2Seq) models. The models consist of two recurrent networks: the coder and the decoder. The coder creates the input of the entry word sequence. The gained input (the last exit and the value of network cell) are copied to the decoder. Through the received input the decoder tries to reassemble the initial word sequence. In the tasks of machine translation the input and output sequences are the sentences in different languages. In question and answer as well as dialogue systems — question and answer. Reinforcement learning methods for situations of unclear cost function. Learning with reinforcement — is one of methods of machine learning during which a test system is taught by interacting with a certain environment. From the cybernetics point of view — this one of the types of cybernetic environment. The environment response (and not a special system of reinforcement management as it happens when studying with a teacher) on the taken decision are the reinforcement signals. So this learning is a subtype of learning with a teacher, but the teacher is the environment or its model. It is also important to note that certain reinforcement rules are based on non-clear teachers, for example in cases of artificial neural environment, on the simultaneous activity of formal neurons —which is way this can be counted as teacher-less learning. Generative models for simulations and data hungry subtasks. The process of teaching a generative model is as following: a large amount of data from certain area (millions of images, sentences or sounds, etc.) is amassed and then the model is taught to generate this data by itself. Neural networks that are used for generative models possess much less parameters compared to the assembled data that they rely on. So to generalize the data — the models need to locate and effectively analyze the samples. It is important to consider the following factors. The decision on utilizing s certain method is taken only after the detailed analysis of its effectiveness. The models can further self-educate with the new market data arriving. There is a possibility of increasing model sensitivity by increasing the computing performance and analysis depth in case of shifting to the emergency mode when an attack is being suspected. The models have place for the experts to monitor and indicate insides regarding a specified market situation at the current moment. The methodology of direction data is unified for all types of modules and passed through the A/B tests. The architecture of the system: distributed nodes in the cloud — this allows to choose between the system reaction speed, the depth of the analysis and the cluster prize. HyperQuant’s frame of mind. We in HyperQuant have outstanding experience with the development and use of such complex systems. We keep on trying new methods, expanding the horizons with the cutting-edge ideas. We build our work on the methods of the IT giants, yet advance them even further for your merit. We will make sure to tell you more about how our technologies work and profit you in the next instalments of this article series. HyperQuant Social Media

Light Paper: Artificial Intelligence in HyperQuant Ecosystem

90

light-paper-artificial-intelligence-in-hyperquant-ecosystem-1c5fc82f6ffc

2018-06-12

2018-06-12 03:59:24

https://medium.com/s/story/light-paper-artificial-intelligence-in-hyperquant-ecosystem-1c5fc82f6ffc

false

1,400

Automatic Trading Revolution: https://hyperquant.net/

null

hyperquant.net

null

hyperquant

info@hyperquant.net

hyperquant

null

HyperQuant_net

Machine Learning

machine-learning

Machine Learning

51,320

HyperQuant

Automatic Trading Revolution https://hyperquant.net/

da4c15da74be

hyperquant

185

87

20,181,104

null

0

null

0

null

2018-03-13

2018-03-13 12:03:17

2018-03-13

2018-03-13 12:05:16

1

false

en

2018-03-13

2018-03-13 12:05:16

3

1c608147e68

2.098113

0

ZeroState is the first marketplace in the world, applying the emotional intelligence, which effective manage the customer loyalty through…

5

The apply of emotional intelligence ZeroState is the first marketplace in the world, applying the emotional intelligence, which effective manage the customer loyalty through the emotional involvement. The platform is based on emotional intelligence and provides an effective interaction between business and consumers, based on emotional bonds and attachments. The path to customer loyalty goes through emotional engagement. If you think about the brand you buy it. The Japanese can do advertising that affects to the subconscious mind. The emotional intelligence in the advertisement The most striking examples of such effects can be taken from Japanese advertisement. It may contain nothing at all about the advertised product, except the image of product or brand. And the whole concept is aimed to create a connection between the strong emotions and product. And it works. It is the simplest of all connections: stimulus — reaction. The consumer at the deepest level of information is fixed that the joy, comfort, pleasure or friendship is associated with a very specific product. Among the examples we know are: Tefal takes care of us, L’oreal is the beauty, quality, status and pleasure which we worthy. Even the main efforts of SSM-promotion in social networks are focused on human emotions. In order to strengthen them are using the fashionable memes and bright images. If you managed to cause the strong emotions in a person, light a spark in him or her and fascinate with your message, you will get an emotionally involved client. Such client is a grateful client. Thanks this there is a “word of mouth” that sells your product when you are resting. The emotional intelligence is a door to success The emotional intelligence (EQ) is the door to self — knowledge. The term, introduced by American researchers is studying the factors of success and uniting them in the concept of “emotional intelligence”. This term has used in business, because the development of emotional intelligence or emotional competence can significantly improve the efficiency of employees, owners and increase the profits. To be realized in all the spheres of life, important for you, you need to have the communication skills and organization. Scientists say that our success depends on emotional intelligence (EQ) for 80% and only for 20% on relative intelligence (IQ). This is because the EQ components are directly responsible for decision-making, performance, and communication. The basement of ZeroState is an application for reviews about anything in one word. This word has a sensual-emotional color or grade. How does the Emotional intelligence work in ZeroState The emotional intelligence (EQ) is what a person feels in relation to his environment, the feeling that arises in us before coming into mind and using social filters and evaluation of “good” and “bad”. These feelings person wants to share and discuss. We are collecting all these feelings through our app and giving you the opportunity to share them, build a rating, use to gain knowledge about the product or service, as well as earning on them! #EQ #zerostate #emotionalintelligence

The apply of emotional intelligence

0

the-apply-of-emotional-intelligence-1c608147e68

2018-03-13

2018-03-13 12:05:17

https://medium.com/s/story/the-apply-of-emotional-intelligence-1c608147e68

false

503

null

Emotional Intelligence

emotional-intelligence

Emotional Intelligence

5,148

ZeroState

EMOTIONAL INTELLIGENCE POWERED BY BLOCKCHAIN Effective solution for customer loyalty management through emotional involvement

93e5d0a577e0

zerostate

130

94

20,181,104

null

0

null

0

null

2018-02-09

2018-02-09 08:14:26

2018-02-09

2018-02-09 08:15:54

6

false

en

2018-02-09

2018-02-09 08:15:54

3

1c60c709443b

3.478302

0

Google is on the move in China. In December 2017, Google established the AI China Center in Beijing, led by Fei-fei Li, chief scientist of…

5

Google China Focuses on AI, Help Chinese Business to Expand Overseas Google is on the move in China. In December 2017, Google established the AI China Center in Beijing, led by Fei-fei Li, chief scientist of Google Cloud, and Jia Li, the head of research and development in AI and ML. In mid-January, Google confirmed the news it would set up an office in Shenzhen. Google now has more expectations and plans to invest in the Chinese market. SEE ALSO: Google Confirms Intent to Establish AI Team in China On February 6, Google held the 2018 Think With Google conference in Beijing, the annual summit at which Google introduces its service to its partners. For Google, the Chinese market remains special and challenging. At present, Google’s strategy in China is to push AI, scramble for opportunities and help Chinese companies expand to overseas by advertising. AI is not only for large companies, Scott Beaumont, president of Google Greater China, said at the conference. He encouraged more developers to use Google TensorFlow, an AI development framework. The Google AI platform, composed mainly of the machine learning computing framework TensorFlow and the AI invocation engine ML Engine, is Google’s competitive entry in the AI field. Google said TensorFlow is the most widely used AI development framework in the world, and it has been downloaded in China for more than a million times. The top users are Sun Yat-sen University, Tsinghua University, Xiaomi, JD.com and other institutions and enterprises. ML Engine provides pre-trained translation, chat, image recognition, natural language processing and other capabilities for users through an application program interface (API). Through the AI platform, Google can offer guidance in medical care, education and other fields, such as case diagnosis, personalized tutoring and other applications. Jia Li said the goal of Google Cloud AI. It aims to help companies clarify demands and enhance product functions by combining research strength with popular products. Huang Jiezhong, Google’s chief marketing officer of Greater China, explained some of the marketing opportunities that Chinese enterprises will face in the era of AI. He said with the help of Google AI, companies will be favored by consumers for accurate marketing, deep interaction and brand experience. Through cooperation with mobile games company, AI can analyze user preferences and ensure accurate advertising, Huang said. Compared to artificial advertising optimization’s, AI could double return on investment. As Google’s consumer-oriented business can’t be fully developed on the Chinese mainland, its role in market is as an advertising platform targeted at overseas users. Chinese enterprises can enhance their popularity and attract consumers using Google’s search engine, YouTube and Google Play. SEE ALSO: BrandZ: Internet Tech Companies Make List of Top 50 Chinese Global Brand Builders At the conference, Google and WPP released a list of BrandZ’s Top 50 Chinese Global Brand Builders 2018. Lenovo, Huawei and Alibaba ranked as the Top 3. Consumer electronics, games and e-commerce providers are the most common brand types in the list. When Chinese brands go abroad, they will need strong advertising support. Some 84 percent of Google’s revenue comes from advertising, so Chinese enterprises are undoubtedly worth striving for. In 2017, Google made frequent moves in China, and the rumor that Google planned to return to China came up from time to time. At this conference, Google mentioned that its translation can be used normally in China and will offer better experience with help of AI. The news that Google set up an office in Shenzhen caused concern. Scott Beaumont said the Shenzhen office is mainly a background service office — not a technical institution. He mentioned there are many high-tech enterprises in Shenzhen. Google and these enterprises have a lot of joint projects in hardware development, so it is natural for Google to establish an office in Shenzhen. Google recently released its earnings, creating a new record of growing more than 20 percent in revenue for 32 consecutive quarters. This article originally appeared in Jiemian and was translated by Pandaily.

Google China Focuses on AI, Help Chinese Business to Expand Overseas

0

google-china-focuses-on-ai-help-chinese-business-to-expand-overseas-1c60c709443b

2018-06-05

2018-06-05 05:34:20

https://medium.com/s/story/google-china-focuses-on-ai-help-chinese-business-to-expand-overseas-1c60c709443b

false

670

null

China

china

China

27,999

Pandaily

Everything about China’s innovation

a0de6420e386

pandaily

169

6

20,181,104

null

0

null

0

null

2018-01-14

2018-01-14 13:31:51

2018-01-14

2018-01-14 13:44:32

4

false

en

2018-01-14

2018-01-14 13:47:30

0

1c62c9a5ddcf

8.326415

1

0

With one of the world’s largest youth population, India is poised to become one of the world’s leading markets in Gaming Sector. Currently…

5

Machine Learning in Games Development With one of the world’s largest youth population, India is poised to become one of the world’s leading markets in Gaming Sector. Currently valued at USD 890 million, the Indian Gaming Industry is estimated for the annual growth rate of 14.3 per cent with mobile Gaming taking the lead at 71% share. The growth is driven by rising younger population, higher disposable incomes, introduction of new gaming genres, and the increasing number of smartphone and tablet users. With a growing consumer base and e-commerce solutions in place, the Indian game developer ecosystem will have ample opportunity to innovate and provide compelling content with social and cultural context, which is key in engaging the Indian consumer. With experience outsourcing, QA testing and developing products combined with availability of creative talent, advanced technology and low-cost, Indian markets provide the best opportunity for collaboration and development work. Machine Learning has been greeted with a certain amount of caution by games developers, and until recently, has not been used in any major games releases. Why is this — surely there must be potential demand for games that can learn — games that can adjust strategy to adapt to different opponents? There are several major reasons for the lack of enthusiasm which has, for a long time, been exhibited. Another question to be asked, is just how important is it for a game to ‘learn’? Is the average games player going to appreciate any significant advance in gameplay, or will all the effort be a waste of time and money? This definitely depends on the game. Many game developers are currently looking at the possibility of making games that can match the player’s ability by altering tactics and strategy, rather than by improving the ability of opponents. This sounds similar to the standard ‘difficulty level’ feature which is hardly a rarity, but don’t be fooled — there are few games on the market which can uncover a player’s tactics and adapt to them. Even on the toughest difficulty settings of most games (FPS’s especially), most players have a routine, which if successful, will mean that they win more often than not. However, they would surely not be so smug if the AI could work out their favourite hiding places, or uncover their winning tactics and adapt to them! This could become a very important feature of future releases, as it would prolong game-life considerably. Varieties of Learning The greatest temptation for designers, is to create a false impression of learning. It is commonplace within the gaming industry to create cheating AI systems, and I suppose there can be few moral objections to this as it does simplify things a great deal. An impression of learning can be easily implemented by controlling the frequency of errors in tactical decisions made by the AI, and reducing them with ‘experience’ as the game is played. This creates a realistic illusion of an intelligent learning process, but cannot be used unless the desired behaviour is already known — in other words, this is useless for learning to counter player strategy. Central to the process of learning, is the adaptation of behaviour in order to improve performance. Fundamentally, there are two methods of achieving this — directly (changing behaviour by testing modifications to it), and indirectly (making alterations to certain aspects of behaviour based on observations). There are positive and negative sides to each, but direct adaptation does have the advantage of not limiting behaviour, which means that ultimately, a better goal may be achievable. Here’s a look at how Youtuber SethBling used neural networks and artificial intelligence in generating a simulation for the popular Super Mario. In his own words: “MarI/O is a program made of neural networks and genetic algorithms that kicks butt at Super Mario World.” Creating a Learning Agent There are various ways in which a learning agent can be created for gameplay. For example, game developers of Quake, Age of Empires, Warcraft etc. use the example of a team-strategy based paintball game. The aim of the program is for a team of seven agents to capture the opponent’s flag and bring it back to their starting position. They must do this without being hit by the opposing team’s paintballs. So, what elements are involved in the tactics behind this type of game? Consider the factors which will influence an agents performance in the game. Terrain is an obvious start point, as this is all that stands between the two teams, so it must be used to the agent’s advantage. Secondly, there must be an element of stealth behind the behaviour of each agent, as otherwise it will be simple to undermine any tactics used during the game by simply picking off the naïve agents one-by-one. A learning agent is composed of a few fundamental parts : a learning element, a performance element, a curiosity element (or ‘problem generator’), and a performance analyser ( or ‘critic’). The learning element is the part of the agent which modifies the agent’s behaviour and creates improvements. The performance element is responsible for choosing external actions based on the percepts it has received (percepts being information that is known by the agent about its environment). To illustrate this, consider that one of our agents is in the woods playing paintball. He is aware of an opposing paintballer nearby. This would be the percept that the agent responds to, by selecting an action — moving behind a tree. This choice of action is made by the performance element. The performance analyser judges the performance of the agent against some suitable performance measure (which in this case could be how close the agent is to being hit by the enemy, or how many enemies have been hit). The performance must be judged on the same percepts as those received by the performance element — the state of affairs ‘known’ to the agent. When the analysis of performance has been made, the agent must decide whether or not a better performance could be made in the future, under the same circumstances. This decision is then passed to the learning element, which decides on the appropriate alteration to future behaviour, and modifies the performance element accordingly. It does occur at one point of time as to how do we make sure that the agent advances in its learning, and doesn’t merely confine itself to previously observed behaviour? This is dealt with by the curiosity element (so-called because it searches for a better solution) which has a knowledge of the desirable behaviour of the agent (i.e. it knows that being shot is not desirable, and that finding the opponent’s flag is!). To achieve optimal performance, this element will pose new challenges to the agent in an attempt to prevent (bad) habits developing. To understand the benefits of this, consider a paintballer who is hiding behind a tree. From his past experience, he knows that he is safe to stay where he is, and this would result in an adequate performance. However, the curiosity element kicks in, and suggests that he makes a break from his cover and heads to a nearby tree which is closer to the enemy flag. This may result in the agent ultimately being shot at, but could also achieve a more desirable goal. It is then up to the performance analyser and the learning element to consider whether there is a benefit to this change in strategy. At this point, it would be a good idea to mention the fact that this style of learning is known as reinforcement learning, which means that agent can see the result of its actions, but is not told directly what it should have done instead. This means that the agent must use, what is really trial and error, to evaluate its performance and learn from mistakes. The advantage to this is that there is no limitation on the behaviour, other than the limit to alterations suggested through the curiosity element. If after each action, the agent was told what its mistake was and how it should correct its behaviour, then the desired behaviour must already be understood, and therefore the learning is, in effect, obsolete. As the learning agent is ultimately part of a game, it must not be left simply to work out for itself how to play. The agents must be imparted with a fair degree of prior knowledge about the way to behave. In the case of paintball, this could include methods for avoiding fire by using cover, which may later be adapted during the learning process. Many games developers use learning algorithms in their games to create better computer player AI, but the resulting AI is then ‘frozen’ before shipping. Decision trees are widely believed to be a good method of ‘reasoning’ — as are belief networks and neural networks, but these are beyond the scope of this article. Problems with Learning Despite the obvious potential that learning has to offer the gaming world, it must be used carefully to avoid certain pitfalls. Here are but a few of the problems commonly encountered when constructing a Learning AI: Mimicking Stupidity — When teaching an AI by copying a human player’s strategy, you may find that the computer is taught badly. This is more than likely when the player is unfamiliar with a game. In this situation, a reset function may be required to bring the AI player back to its initial state, or else a minimum level must be imposed on the computer player to prevent its performance dropping below a predetermined standard. Overfitting — This can occur if an AI agent is taught a certain section of a game, and then expected to display intelligent behaviour based on its experience. Using a FPS as an example, an agent which has learnt from its experience over one level will encounter problems when attempting a new level, as it may not have learnt the correct ‘lessons’ from its performance. If it has found that when opening doors, it has been able to escape the line of fire by diving behind a wall to its left, it will assume that this is a generalized tactic. As you can imagine, this could lead to amusing behavioural defects if not monitored in the correct way… Local Optimality — When choosing a parameter on which the agent is to base its learning, be sure to choose one which has no dependency on earlier actions. As an example, take a snow-boarding game. The agent learns, through the use of an optimization algorithm, the best course to take down the ski slope, using its rotation as a parameter. This may mean that a non-optimal solution is reached, in which any small change cannot improve performance. Think about the data being stored — a sequence of rotations clockwise and anticlockwise. An alteration to a rotation in the first half of the run may lead to a better time over the course in the long run, but in the short-run, could cause a horrific crash further down the slope, as the rest of the rotations are now slightly off course! Set Behaviour — Once an agent has a record of its past behaviour and the resulting performance analysis, does it stick to the behaviour which has been successful in the past, or does it try new methods in an attempt to improve? This is a problem which must be addressed or else an agent may either try to evaluate every possible behaviour, or else stick to one without finding the optimal solution. Conclusion Having looked at possible applications for learning, and seen some of the problems associated with it, it seems that there is great potential for learning in games, but it must be used with caution. The majority of games have little use for any kind of learning techniques — except in the development and testing stages. Despite the price of developing this kind of software, it looks as if learning will have a large part to play in the next generation of games.

Machine Learning in Games Development

10

machine-learning-in-games-development-1c62c9a5ddcf

2018-05-09

2018-05-09 12:59:39

https://medium.com/s/story/machine-learning-in-games-development-1c62c9a5ddcf

false

2,021

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Nabeel Abdul Latheef

Pray. Wander. Eat. Love. Sing. Game. Gym. Repeat. Views are my own! #DiabolicalSoul #Revenant #Rebel #IndignationHegemony

bf11535b549d

nabil_lathif

23

16

20,181,104

null

0

null

0

null

2017-10-30

2017-10-30 15:41:32

2017-10-30

2017-10-30 15:49:28

2

false

fr

2017-10-30

2017-10-30 15:49:28

1

1c62e71b7536

7.405975

3

0

Octobre 2017. On ne compte plus le nombre d’articles de presse publiés chaque semaine expliquant comment l’Intelligence Artificielle va…

5

Quels champs pour une expérimentation rapide et concrète de l’I.A. ? l’exemple de la santé Octobre 2017. On ne compte plus le nombre d’articles de presse publiés chaque semaine expliquant comment l’Intelligence Artificielle va révolutionner, pour le bien ou pour le mal, le quotidien de l’humanité. Que l’on soit d’accord ou non avec cette affirmation, il y a des choses certaines : l’Intelligence Artificielle a passé un cap décisif en 2017, elle a commencé à être appliquée et industriellement déployée dans de grandes entreprises. Intelligence artificielle et santé, une promesse forte Dans ce contexte, il est légitime pour tout dirigeant du secteur de la santé de s’interroger sur la façon d’expérimenter, d’appliquer ou de déployer des applications d’IA dans leur domaine activité. Mais par où commencer ? il existe de nombreux champs d’applications, en particulier dans le domaine médical, diagnostic et thérapeutique : L’aide au diagnostic : probablement le champ le plus connu et médiatisé de l’I.A. en santé. Se basant sur l’exploitation d’un ensemble de données patient (résultats d’examens cliniques et paracliniques, profil génétique, etc…), l’aide au diagnostic permet de faciliter les tâches quotidiennes des médecins. Loin de les remplacer, elle « augmente » les compétences des praticiens en les déchargeant des tâches les moins complexes et chronophages. Théoriquement applicable à tous types de cas, l’aide au diagnostic se révèle particulièrement utile dans des cas lourds et complexes. A titre d’exemple, face à l’échec des chimiothérapies préconisées, la plateforme Watson d’IBM a réussi à détecter une anomalie de la moelle osseuse sur une patiente japonaise, là où les praticiens n’avaient diagnostiquer qu’un simple cancer. Le monitoring de signes physiologiques : il existe aujourd’hui sur le marché des dispositifs médicaux connectés permettant le suivi permanent du patient : semelles connectées pour le diabète, vêtements connectés pour la surveillance de personnes épileptiques, pour la mesure de signes cliniques, pour l’envoi d’alertes en cas d’anomalie, patch permettant la mesure de température et l’envoi d’alertes… L’analyse automatisée de l’imagerie : L’intelligence artificielle permet d’interpréter plus vite et avec une meilleure précision que l’œil humain les imageries médicales. La société Cardiologs commercialise une solution permettant de mieux détecter la présence de maladies cardiovasculaires grâce à des bases de données d’électrocardiogramme déjà interprétés. La prévision et détection précoce de maladies chroniques : La multiplication des données patients (sur la maladie ou le mode de vie) combinée à la capacité d’apprentissage des algorithmes d’intelligence artificielle permet aujourd’hui aux ordinateurs d’être parfois plus performant que les humains dans la prévision des maladies, à l’image de ce qui se fait pour les maladies cardio-vasculaires [1]. La médecine personnalisée : basée sur l’ensemble des informations personnelles d’un patient, la médecine personnalisée permet de traiter chaque patient de façon individualisée en fonction de ses spécificités génétiques et environnementales. Dans le cas d’un Cancer, elle permet de cibler la thérapie la mieux adaptée à un patient parmi plusieurs traitements possibles et dont le choix est guidé par l’identification d’un biomarqueur prédictif de réponse[2]. Au contraire de la Chimiothérapie, la thérapie ciblée n’interagit que sur certaines molécules ayant un rôle crucial dans la progression du cancer. L’analyse prédictive appliquée au cycle du médicaments : L’Intelligence Artificielle permet de réduire sensiblement le cycle de mise sur le marché d’un médicament. Grâce à l’analyse prédictive[3], via simulations informatiques répliquant des expériences d’ordinaires réalisées en laboratoire, il est possible de prédire des réactions moléculaires « futures » en se basant sur un historique de plusieurs millions d’expériences passées. En conséquence, les résultats seraient jusqu’à 150 fois plus rapides que ceux des outils actuels. Cette liste est loin d’être exhaustive, et il n’y a que peu de doutes quant au potentiel de croissance de ce marché dans le temps. Au-delà des attendus thérapeutiques et médicales, l’espoir placé dans l’I.A. est fort, notamment en matière d’amélioration du fonctionnement et de l’efficience des systèmes de santé : Face à la pénurie annoncée de médecin et la pénibilité croissante du métier, le développement d’applications d’assistance aux diagnostics peut être perçue comme une solution, permettant aux praticiens d’améliorer leur productivité, de se focaliser sur les tâches à plus forte valeur ajoutée, et d’améliorer le confort au travail Face au développement croissant des maladies chroniques dans les pays occidentaux et à leurs implications sur les systèmes de santé, l’Intelligence Artificielle, par sa capacité à faciliter le suivi régulier et la prévention des complications et incapacités, pourrait permettre de simplifier significativement la prise en charge des patients. Face aux coûts élevés de fonctionnement des systèmes de santé, le développement de la médecine préventive / prédictive basée sur des algorithmes d’I.A. pourrait limiter fortement le recours à des interventions lourdes et coûteuses liées à la prise en charge tardive de certaines pathologies. Des interrogations lourdes quant à la mise en place des applications à fort contenu médicale Bien que les promesses des applications d’IA en thérapeutique et médical soient particulièrement séduisantes, la mise en œuvre de ces dernières suscite des interrogations lourdes pour tout décideur souhaitant se lancer dans un projet. Tout d’abord, ces applications pourraient impacter plus ou moins fortement l’organisation des opérateurs de soin. Prenons l’exemple d’un diagnostic préliminaire et automatisé des patients au moment de l’entrée en hospitalisation : La mise en œuvre du diagnostic pourrait induire une redéfinition des rôles et des responsabilités des acteurs dans l’organisation, le diagnostic pouvant théoriquement se faire sans la présence du médecin. Dans ce cas, qui est responsable ? A l’instar de la télémédecine, la tâche pourrait faire l’objet d’une délégation auprès des infirmiers, ce qui nécessiterait une réorganisation des pratiques de soins, ainsi qu’une éventuelle autorisation spécifique délivrée par les autorités compétentes. Cette mise œuvre pourrait également impacter les protocoles et processus internes, notamment en matière de prise en charge des patients. La mise en œuvre du diagnostic nécessitera la réalisation d’une refonte, plus ou moins lourde, des systèmes d’informations pour permettre la communicationentre les différentes applications et l’accès aux données nécessaires au diagnostic. D’autre part, le recours à des applications d’IA nécessite d’avoir bien défini et cadré les risques en matière réglementaire et juridique : Certains projets complexes d’Intelligence Artificielle nécessitent de recourir à une variété et un volume important de données personnelles pour fonctionner. Tout décideur doit donc évaluer l’adéquation entre les besoins du projet (objets de la collecte des données, stockage, exploitation) et la nécessité de se conformer à la réglementation nationale et communautaire en vigueur en matière de données personnelles (notamment le RGPD). Au-delà de l’évaluation de conformité réglementaire se pose la question de la responsabilité juridique. En cas d’erreur du diagnostic conduit par la machine, quelle est la responsabilité du praticien en charge ? quelle est la responsabilité de l’établissement de santé ? Enfin, le recours à des applications médicales, diagnostiques et thérapeutiques d’Intelligence Artificielle amène les décideurs à devoir se positionner sur des questions cruciales en matière d’éthique : Si une application est en mesure d’évaluer la probabilité de survie d’un patient, quel arbitrage du décideur en matière sélection du traitement adéquat ? En d’autres termes, si l’application évalue que la probabilité de survie du patient est inférieure à 1%, va-t-on dépenser des centaines de milliers d’euros pour un traitement lourd ou pour une molécule innovante ? Si l’Intelligence Artificielle permet d’améliorer la détection précoce de maladies rares ou de malformation, quels risques en matière d’Eugénisme ? Si l’intelligence Artificielle limite le libre arbitre des praticiens, quels en seraient les impacts en matière de relations humaines ? Des domaines d’applications à privilégier pour une première expérimentation ? Bien qu’elles s’appliquent à différents degrés, fonctions de la maturité et de la nature des projets, les questions soulevées par les applications thérapeutiques et médicales peuvent rallonger fortement l’horizon de mise en œuvre effective pour qui voudrait expérimenter l’Intelligence Artificielle. Il existe toutefois des domaines d’application, nécessitant une mise en œuvre moins contraignante et pouvant bénéficier non moins fortement à l’amélioration globale du système de santé. Des solutions pour améliorer la relation entre les patients et les établissements de santé Développées pour des entreprises évoluant en dehors du secteur de la santé, certaines applications peuvent se révéler non moins utiles et efficaces pour les établissements de santé. C’est le cas des « chatbots », largement répandus dans le commerce en ligne, qui permettent grâce à des scénarios préétablis et à l’analyse prédictive de données de tenir des conversations « humaines » avec les clients et prospects d’une entreprise. Ces applications sont intéressantes car elles ne remplacent pas forcément le travail « humain » mais permettent soit de le compléter (disponibilité d’un interlocuteur en dehors des heures d’ouverture), soit de le recentrer sur des tâches complexes à plus forte valeur ajoutée (demandes clients complexes). Il est facilement imaginable de déployer ce type de solutions pour améliorer l’accueil et l’information des patients à l’entrée d’un établissement. Des solutions pour réduire les coûts de fonctionnement des établissements de santé Il existe par ailleurs des applications d’Intelligence Artificielle spécifiques au secteur de la santé, hors du domaine thérapeutique et médical, à l’image de Lifen. Crée par un des fondateurs de Criteo[4], Lifen propose une solution permettant d’améliorer la communication des informations médicales entre l’hôpital et la ville, via une technologie prenant en charge automatiquement et quel que soit le format (Postal ou Numérique) les envois et réception de comptes rendus médicaux. L’application de cette technologie permet non seulement de gagner en qualité d’utilisation (comptes rendus stockés sous une seule forme et dans un même endroit), mais également en terme économiques, en limitant drastiquement le recours au courrier papier. Des solutions pour simplifier le transport sanitaire Par ailleurs, il est largement possible d’imaginer d’autres applications possibles à partir de solutions existantes en dehors du secteur de la santé à l’exemple des gestionnaires de flottes. Le développement de dérivés de « Uber » ou « Blablacar » appliqués au transport sanitaire ne pourrait que conforter les plateformes actuelles mises en place par les pouvoirs publics pour réduire les dépenses de transport. Une stratégie de transformation « progressive » S’il est probable que l’Intelligence Artificielle change fondamentalement les métiers et l’organisation de la santé dans la (les) décennie(s) à venir, il est en revanche moins probable, au vu des freins énoncés, que cette transition ne puisse se faire de manière aussi brutale et rapide, comme cela a était le cas dans certains secteurs ayant subi la « digitalisation » de plein fouet (transports, commerces ou des médias). Ainsi face au risque de s’enliser dans un projet de transformation long et très ambitieux, il peut être recommandé pour un décideur souhaitant faire le « pari » de l’Intelligence Artificielle, de démarrer par des expérimentations à faible enjeu médical, ces dernières ayant des implications moins fortes en terme d’organisation, de freins juridiques et éthiques, et permettant potentiellement de susciter dans un premier temps une meilleure adhésion des parties prenantes. David Rudnianski [1] Journal Science — 14/04/2017 [2] Dr Frederic Eberlé — Le Figaro — 24/11/2014 [3] http://www.chematria.com/ [4] Criteo est une entreprise française de reciblage publicitaire. Fondée en 2005, elle est cotée au Nasdaq en octobre 2014, et réalise environ 2 milliards d’euros de chiffre d’affaires. Criteo est souvent cité comme exemple français de start up ayant passé le stade de « licorne »

Quels champs pour une expérimentation rapide et concrète de l’I.A. ? l’exemple de la santé

5

quels-champs-pour-une-expérimentation-rapide-et-concrète-de-li-a-l-exemple-de-la-santé-1c62e71b7536

2018-06-07

2018-06-07 17:08:10

https://medium.com/s/story/quels-champs-pour-une-expérimentation-rapide-et-concrète-de-li-a-l-exemple-de-la-santé-1c62e71b7536

false

1,861

null

Intelligence Artificielle

intelligence-artificielle

Intelligence Artificielle

478

David Rudnianski

Strategy Consultant // Former Entrepreneur & Economist

df1f3f330b3f

david.rudnianski

49

56

20,181,104

null

0

(bash) >>> brew install cmake (bash) >>> brew install libomp (bash) >>> git clone --recursive https://github.com/Microsoft/LightGBM ; cd LightGBM (bash) >>> mkdir build ; cd build (bash) >>> cmake .. (bash) >>> make -j4 OSError: dlopen(/Users/xxx/anaconda3/lib/python3.6/site-packages/lightgbm/lib_lightgbm.so, 6): Library not loaded: /usr/local/opt/gcc/lib/gcc/7/libgomp.1.dylib Referenced from: /Users/xxx/anaconda3/lib/python3.6/site-packages/lightgbm/lib_lightgbm.so Reason: image not found (bash) >>> brew update (bash) >>> brew upgrade (bash) >>> pip uninstall lightgbm (bash) >>> git clone --recursive https://github.com/Microsoft/LightGBM ; cd LightGBM (bash) >>> export CXX=g++-8 CC=gcc-8 (bash) >>> mkdir build ; cd build (bash) >>> cmake .. (bash) >>> make -j4 (bash) >>> pip install --no-binary :all: lightgbm (bash) >>> python (python) >>> import fbprophet

4

6c3da2525339

2018-09-20

2018-09-20 10:23:56

2018-09-21

2018-09-21 00:00:18

0

false

en

2018-09-21

2018-09-21 00:00:18

1

1c640861e914

0.758491

0

If you fail on installing LightGBM, you might wanna take a look at this post.

2

Install LightGBM on macOS High Sierra If you fail on installing LightGBM, you might wanna take a look at this post. I followed the official guide and tried to install LightGBM on macOS High Sierra, but I failed. According to the official guide, you will have to do the following steps: I got the errors like this: The solution I found on the internet is here. Instead of doing the steps mentioned above, try the steps below. If you didn’t get any errors during installing, then you have succeeded. Now, let’s test!

Install LightGBM on macOS High Sierra

0

install-lightgbm-on-macos-high-sierra-1c640861e914

2018-09-21

2018-09-21 00:00:19

https://medium.com/s/story/install-lightgbm-on-macos-high-sierra-1c640861e914

false

201

My learning stamps

null

Programming Notes

null

tinghaochen

null

Machine Learning

machine-learning

Machine Learning

51,320

Ting-Hao Chen

Machine Learning Enthusiast & Python Lover

7958641d1e91

tinghaochen

39

35

20,181,104

null

0

null

0

null

2018-08-26

2018-08-26 17:51:26

2018-08-26

2018-08-26 18:41:36

1

false

pt

2018-08-26

2018-08-26 18:41:36

16

1c66065c0072

1.792453

9

0

Entre os dias 20 e 22 de Agosto aconteceu a primeira semana de apresentações da edição 2018 do Azure Tech Nights, um evento online…

5

Azure Tech Nights 2018 - 1a semana: saiba como foi cada uma das apresentações [Vídeos Gratuitos] Entre os dias 20 e 22 de Agosto aconteceu a primeira semana de apresentações da edição 2018 do Azure Tech Nights, um evento online, gratuito e noturno do Canal .NET focado no uso de tecnologias e serviços que integram o Microsoft Azure. Seguem alguns dados sobre o Azure Tech Nights: 9 apresentações realizadas até o momento; Picos com mais de 100 usuários online durante as apresentações; Um total de mais de 3400 visualizações, considerando as gravações dos diferentes dias de evento (até 26/08); Áreas: Desenvolvimento Web, Segurança, NoSQL, Inteligência Artificial, DevOps e Docker; Organizadores: Renato Groffe (Microsoft MVP, MTAC), Joel Rodrigues (Microsoft MVP) e André Secco (Microsoft MVP, MTAC). Pensando naqueles que não puderam acompanhar ou, até mesmo, gostariam de rever alguma apresentação, foram agrupados neste post os links (indicados abaixo) para as gravações de todas as palestras realizadas ao longo da primeira semana (dias 20 a 22/08). É importante lembrar ainda que o evento ainda está em andamento, com novas palestras previstas para os dias 27 e 28/08 (segunda e terça desta semana). Para maiores detalhes e efetuar a sua inscrição acesse os seguintes links no Meetup: Azure Tech Nights - Dia 4: App Service, Machine Learning, CDN Azure Tech Nights - Dia 5 - Disaster Recovery em APIs, IoT, Azure Maps E faça também sua inscrição na página oficial no Facebook, a fim de receber notificações sobre tudo o que está acontecendo. Dia 1 - 20/08/2018 Palestra 1 - Conheça o AKS, o serviço de Kubernetes do Azure — Giovanni Bassi (Microsoft MVP) Palestra 2 - Conhecendo as APIs do Azure Cosmos DB — Dani Monteiro (Microsoft MVP) Palestra 3 - Implementando APIs seguras na nuvem — Renato Groffe (Microsoft MVP, MTAC) Dia 2 - 21/08/2018 Palestra 1 - Cognitive Search nos seus Aplicativos - Thiago Custódio (Microsoft MVP) Palestra 2 - Automatizando a entrega das suas Aplicações ASP.NET Core com Docker, VSTS e Azure - Milton Câmara Gomes Palestra 3 - Azure SignalR + Functions + Logic Apps: um exemplo prático - Ericson da Fonseca (Microsoft MVP) e Robson Araújo (Campinas .NET) Dia 3 - 22/08/2018 Palestra 1 - Monitorando uma Aplicação ASP.NET com Application Insights e Power BI - Rafael Cruz (Microsoft MVP) Palestra 2 - Boas Práticas com o Microsoft Azure - Jaqueline Ramos (Microsoft MVP) Palestra 3 - Inteligência artificial para desenvolvedores .NET - Angelo Belchior (Microsoft MVP) E para finalizar, ainda não segue o Canal .NET nas redes sociais? Faça sua inscrição então, para ficar por dentro de novidades sobre eventos, tecnologias Microsoft e outros conteúdos gratuitos: Facebook: https://www.facebook.com/canaldotnet/ YouTube: https://www.youtube.com/canaldotnet

Azure Tech Nights 2018 - 1a semana: saiba como foi cada uma das apresentações [Vídeos Gratuitos]

35

azure-tech-nights-2018-1a-semana-saiba-como-foi-cada-uma-das-apresentações-vídeos-gratuitos-1c66065c0072

2018-08-26

2018-08-26 18:41:37

https://medium.com/s/story/azure-tech-nights-2018-1a-semana-saiba-como-foi-cada-uma-das-apresentações-vídeos-gratuitos-1c66065c0072

false

422

null

Azure

azure

Azure

3,345

Renato Groffe

Microsoft Most Valuable Professional (MVP), Multi-Plataform Technical Audience Contributor (MTAC), Software Engineer, Technical Writer and Speaker

6698199cd3bb

renato.groffe

2,320

27

20,181,104

null

0

null

0

null

2017-10-27

2017-10-27 02:19:38

2017-10-27

2017-10-27 03:18:37

4

false

en

2017-10-27

2017-10-27 03:18:37

0

1c6ba4632f23

0.918868

3

0

Recipe1M: 1M cooking recipes and 800k food images

1

Review: Learning Cross-modal Embeddings for Cooking Recipes and Food Images Recipe1M: 1M cooking recipes and 800k food images historgram of dataset Joint embedding of recipes and images ingredients -> word2vec -> bi-directional LSTM cooking instructions: -> sentence level -> all food images: VGG-16 resnet-50 Network structure of joint embedding minimize the cosine similarity of positive recipe-image pairs and maximize the similarity of non-matching pairs. Loss function Semantic regularization

Review: Learning Cross-modal Embeddings for Cooking Recipes and Food Images

3

review-learning-cross-modal-embeddings-for-cooking-recipes-and-food-images-1c6ba4632f23

2018-01-26

2018-01-26 06:19:43

https://medium.com/s/story/review-learning-cross-modal-embeddings-for-cooking-recipes-and-food-images-1c6ba4632f23

false

58

null

Machine Learning

machine-learning

Machine Learning

51,320

Jackie Loong

National University of Singapore

5662ae3e2c83

Dragen_36706

1

11

20,181,104

null

0

null

0

null

2017-09-23

2017-09-23 17:31:17

2017-09-23

2017-09-23 17:52:38

0

false

en

2017-09-23

2017-09-23 17:52:38

0

1c6da0c455d7

1.671698

0

So you must be wondering why I put my class as the title of this? Well I’m taking this class and I want to track my progress on my future…

2

EL-GY 9123: Intro to Machine Learning So you must be wondering why I put my class as the title of this? Well I’m taking this class and I want to track my progress on my future project. The goal of the project is to demonstrate my understanding of the topics discussed in class. At this point I have submitted my proposal and I’m waiting for it to be accepted before I start working on it (or telling people what it is). Let me tell you what you can expect: 1. The prep I wanted to do a project I find fun and could connect with. The thing is I’m an API newbie sooo before I can do much I need to learn how to use a few apis. The first one is going to be twitter because that’s what I said I was going to use in my proposal. Of course I’m going to build an Android app but everything back-end wise will be python (language used for this course). 2. Acceptance Well I can really control this step but I can’t move forward without it. I haven’t thought about what I’d do if this doesn’t get accepted either. Either way with acceptance comes project planning. I will put on my project manager hat to setup some deadlines given the project due date. 3. Research Every project needs research so I’ll create a way to collect that research using technology of course! Also because I’m me I’ll be doing some UI designs for this in the long run. Let’s be real I want my project to run on my phone because I’m a mobile dev duh. lol 4. Publishing This will be when I drop the mobile app for the real world, turn in my project for class, and tell you all how much I’ve cried about getting this done. Who knows I might have even learned how to do a decent press package by this point, if so I’ll publish it for experience. 5. Celebrate I’m not sure how to write about this one but I guess I’ll just link my instagram (@keheirathadev) when the time comes. I also have a friend that wants to help me potentially pitch my idea if it’s really that cool so let’s just keep this one as an optional step. Well hope guys enjoy this journey as much as me. I’m going to try to do weekly-ish on updates. Of course if you have any questions/comments you can always hit me up online, email me, submit a github pull request, or comment. Until next time I think I’ll go finish my lab. lol

EL-GY 9123: Intro to Machine Learning

0

el-gy-9123-intro-to-machine-learning-1c6da0c455d7

2017-09-23

2017-09-23 17:52:39

https://medium.com/s/story/el-gy-9123-intro-to-machine-learning-1c6da0c455d7

false

443

null

Machine Learning

machine-learning

Machine Learning

51,320

Keheira tha Dev

🤓 . 🎶. 👨🏽‍💻. Hardware & security enthusiasts. Just out here to learn and talk about as much random stuff as possible. she/her/they/them

2c5e3e9fc6c

keheira

169

131

20,181,104

null

0

null

0

d0c08339b41f

2018-01-19

2018-01-19 18:44:46

2018-01-19

2018-01-19 18:52:09

1

false

en

2018-01-19

2018-01-19 18:52:09

21

1c6e4c3191a6

1.720755

0

TWiML Talk 95

5

Embedded Deep Learning at Deep Vision with Siddha Ganju TWiML Talk 95 In this episode we hear from Siddha Ganju, data scientist at computer vision startup Deep Vision. Siddha joined me at the AI Conference a while back to chat about the challenges of developing deep learning applications “at the edge,” i.e. those targeting compute- and power-constrained environments. Subscribe: iTunes / SoundCloud / Google Play / Stitcher / RSS In our conversation, Siddha provides an overview of Deep Vision’s embedded processor, which is optimized for ultra-low power requirements, and we dig into the data processing pipeline and network architecture process she uses to support sophisticated models in embedded devices. We dig into the specific the hardware and software capabilities and restrictions typical of edge devices and how she utilizes techniques like model pruning and compression to create embedded models that deliver needed performance levels in resource constrained environments, and discuss use cases such as facial recognition, scene description and activity recognition. Siddha’s research interests also include natural language processing and visual question answering, and we spend some time discussing the latter as well. Giveaway Update! Thanks to everyone who took the time to enter our #TWiML1MIL listener giveaway! We sent out an email to entrants a few days ago, so please be on the lookout for that. If you haven’t heard from us yet, please reach out to us at team@twimlai.com so that we can get you your swag! TWiML Online Meetup The details for our January Meetup are set! Tuesday, January 16, we will be joined by former TWiML guest and Microsoft Researcher Timnit Gebru. Timnit joined us a few weeks ago to discuss her recently released, and much acclaimed paper, “Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States”, and I’m excited that she’s be joining us to discuss the paper, and the pipeline she used to identify 22 million cards in 50 million Google Street View images, in more detail. I’m anticipating a lively discussion segment, in which we’ll be exploring your AI resolutions & predictions for 2018. For links to the paper, or to register for the meetup, or to check out previous meetups, visit twimlai.com/meetup. About Siddha Siddha’s Personal Page Siddha on Linkedin Siddha on Twitter Mentioned in the Interview Deep Vision METEOR BLEU CIDEr SPICE TWiML Presents: Series Page TWiML Events Page TWiML Meetup TWiML Newsletter

Embedded Deep Learning at Deep Vision with Siddha Ganju

0

embedded-deep-learning-at-deep-vision-with-siddha-ganju-1c6e4c3191a6

2018-01-19

2018-01-19 18:52:10

https://medium.com/s/story/embedded-deep-learning-at-deep-vision-with-siddha-ganju-1c6e4c3191a6

false

403

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

Machine Learning

machine-learning

Machine Learning

51,320

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

2017-10-10

2017-10-10 23:49:38

2017-10-15

2017-10-15 00:43:33

4

false

en

2017-10-18

2017-10-18 22:10:27

9

1c6e8a743994

3.816981

0

How can we build a better world for everyone using innovative technologies?

5

Teaching Teens The Future of Work How can we build a better world for everyone using innovative technologies? I had some interesting conversations with Salt Lake City teens about STEAM (Science, Technology, Engineering, Art, Math) stuff. Over the course of the class, we looked at the implications of automation and artificial intelligence, as well as tools for modern careers and life-skills. The team from West High School were challenged with thinking on macro and microeconomic scales in an increasingly globalized world. What kind of jobs will be here for them in 5–15 years and what skills will they need to get them? Or better yet, to create them. Experiential learning is an important component in getting people to understand what they are capable of. Without tangible examples, it’s harder to use fragmented dots to solve new problems. Jobs of the future will not follow post-industrial models. Many jobs with repetitive patterns will become automated. (yes, even the professions of law and general medicine) Work that requires deep empathy, niche creativity, or unpredictability will be in demand. Having broad skill-sets and being able to pull from multiple knowledge pools is the new normal. It is important to be a creator, not just a passive consumer. Here’s 4 of their experiences: 1 We partnered with Virtualities, a cool virtual reality cinema & arcade experience at The Gateway Mall. They use state-of-the-art gear like the HTC VIVE and Oculus Rift to fully immerse you in a land far, far away or killing zombies, of course, they preferred the zombies. They’re also working on a collaborative community experience where a user can create a virtual model and have it 3D printed in the space, which lowers the prohibitive cost barrier to 3D printing. This was the student’s first time trying VR and they were astounded by how real it felt. Some of them are now exploring careers in VR. 2 They visited Avenues Courtyard, an assisted living community for seniors. Their aim was to help seniors use technology to better connect with their loved ones and staying safe online. From using mobile phones to setting up Facebook accounts, they learned that elders are just like them and the value of giving back to the community. And no matter what age you are, everyone loves posting selfies. 3 A visit to Comcast HQ gave them a glimpse into what working in the real-world was like. They are now part of the exclusive club allowed to see “The Internet” aka the server hub for Salt Lake’s metro area. Only a select few Comcast team members even have the authorization to see it in real life. 4 Next stop, the award winning Lassonde Studios, recently ranked #1 College for Aspiring Entrepreneurs by LendEDU, at the University of Utah. Full disclosure, I had the privilege of joining the elite Launched at Lassonde program, so I may be a bit biased, but it really was a great experience. Launched at Lassonde supports student-run, whether in concept or growth, companies with great resources: mentoring, dedicated office space, unlimited coffee, monthly member meetings where founders help each other find solutions to their problems, and networking opportunities, to name a few. The studios operate on a collaborative startup incubation model with 400 live-in students. The best part about Lassonde is that the facilities are available to anyone on campus. You don’t need to be a business major or resident to utilize their facilities, whether it’s the 3D printers, Wood shop, Focus Workshops, or cafe, they make it easy for you to take your vision, gather the materials free of charge (Made possible by Pierre Lassonde & Zions Bank, thanks!), build it, test it, iterate, test again, pivot, and boom you’ve built the next unicorn. CLOSING THOUGHTS No one can predict the future and uncertainty can be frightening, but having rose colored glasses on doesn’t solve any problems. Many of the students have never been exposed to the coming tide of pervasive automation and its impacts on their futures (economical, sociopolitical, societal). Suffice it to say, they had mini existential crises with eyes wide open. This was only an after-school program. We need to be exposing all kids to what is already occurring at an exponential rate. Our current system of K-12>College>Career = Stability, is unrealistic. It’s not all doom and gloom, new jobs will be created. People will figure out ways to innovate. These students now have a head-start with new tools to navigate their lives. Special thanks to our wonderful community partners, we couldn’t have done it without your support! To Centro de la Familia de Utah, for allowing me the privilege of curating the curriculum and teaching the next generation. To Comcast for providing delicious snacks, great resources, and laptops for the teens. To Virtualities for allowing us to have a VR experience in your space. To Lassonde Studios and the University of Utah for the wonderful tour. To West High for providing a computer lab for the students. To Avenues Courtyard for allowing us to share our knowledge with the elders.

Teaching Teens The Future of Work

0

teaching-teens-the-future-of-work-1c6e8a743994

2018-02-14

2018-02-14 21:30:32

https://medium.com/s/story/teaching-teens-the-future-of-work-1c6e8a743994

false

826

null

Future Of Work

future-of-work

Future Of Work

8,540

Jerry Hansana

Founder @Tellomee & @PhotonicHouse ⏣ Social Entrepreneur. Building the future of purposeful work and meaningful life experiences.

eb541578da82

jerryhansana

17

24

20,181,104

null

0

null

0

null

2017-11-11

2017-11-11 02:31:15

2017-11-11

2017-11-11 05:29:01

5

false

en

2017-11-16

2017-11-16 19:03:15

1

1c6f2edc99f7

4.814465

5

0

Couple of weeks ago, linear modelling was a distant memory in my brain. It is one of those things you learn in school and you brush it of…

3

How I applied linear modelling to solve a ‘real life’ problem Couple of weeks ago, linear modelling was a distant memory in my brain. It is one of those things you learn in school and you brush it of as something you’ll never apply in ‘real life’. But then I had an ‘aha’ moment when I was trying to find a solution to a problem that a client had asked. Through relentless research and late nights, I found a technique to solving the problem — Linear models. Goal of the blog post is to distill the discovery and learning journey into a practical example and implementation practices in a most intuitive way. Although I can’t share the details of the problem because of confidentiality, I can go through an example that proxies the problem closely. Before we jump into it, it’s important to define a structure around forming the problem statement and key aspects of linear programming: Forming a problem statement — this is an important phase as you want to think about the problem you are solving. Is it maximization (i.e maximizing profit) or minimization (i.e. minimizing cost or time) problem? What are some of the constraints you are dealing with (i.e. capacity) Decision Variables — as the name states, it is variables that equate to decisions that are to be made. For example, what is the maximum number of books you can read in a year given limited time you have. Here the decision variable is number of books. Objective — Element of this is coming from the problem statement. Is it a maximization or minimization problem? Constraints — these are essential as they define boundaries for feasible and optimal solution. Testing — this is my own addition to the formal structure. Testing is important in any application and no exceptions for linear models. Here our goal is to make sure that constants x decision variables add up to constraints. Intuition — another one of my own additions to the structure. Not all models will spit out results that make sense. As you do more and more, you will gain an intuition of defining constraints, making reasonable assumptions to obtain results that make more sense. You will see this in practice below. Implementation — a data science problem such as this is solved best using programming. Python is a popular choice as it has ample mathematical libraries and frameworks. I will be using NumPy and PuLP modelling framework to find optimal solutions. Alright, enough of stories and words — let’s jump into defining our problem Every week, my spouse and I go to a very busy grocery store to do shopping. I am OCD about getting the shopping done in as little time as possible. We buy approximately 50–60 items that are probably in 10–15 locations. By locations I mean different aisles and different section of aisles. Usually we take around 45–60 minutes to finish shopping (checkout time is extra). I take about 32 seconds to find/pick item and about 50 seconds on average to walk to a location. My spouse takes 50 seconds to find/pick item but takes around 32 seconds on average to walk to a location. What is the optimal amount of items and locations in a grocery store that my husband and I should be allocated so that shopping time is minimized? Now that a very ‘real life’ problem has been defined, let’s pick this apart and extract our decision variables, objective and constraints. Decision Variable Decision here is how many items and locations we both should have given our efficiency to find/pick and walk to a location. Objective Minimize shopping time given the decision variable above is our objective. Let’s keep things `apples to apples` and convert the time in seconds to minutes as it makes sense to read the results in minutes denomination. Ok here is the kicker — you must ensure your objective function is linear in nature. That means that sum of constants x decision variables must sum up. Above minimization function only has two individuals participating in shopping madness. However, this can work for n number of individuals. Following represents generalization of the above equation: Constraints I will take the constraints as picking 60 items from 15 locations as this seems to be majority of the scenario. Constraint 1 — constraints the number of picks to 60 Constraint 2— constraints the number of location to 15 Intuition & Assumptions It doesn’t serve any purpose (because I walk slower) for me to pick items at only one location. But it’s very possible the optimal result for a machine is me picking at one location 40 items. For results to make more sense, I will assume minimum locations for my spouse and I are 5 as after that fatigue kicks in from walking in a crowded space. Also, minimum each person picks is 20 because intuitively that seems like a fair ratio to 5 locations. There is quite a bit to chew on here in terms of problem definition and linear model parts. In the next blog post, I will go through how to solve for this problem using Python mathematical libraries such as NumPy matrices, some historic data (that calculates efficiency) and modelling open source library called PuLP. Optimal Results If you are curious about what is the optimal items and locations for my husband and I, here are the results: So looks like we’ve been doing this all wrong — if I took approximately 40 items in 5 locations and my spouse took 20 items in 10 location we could get our shopping done in a total of 25 minutes (we are shopping in parallel so the maximum of the slowest shopper is the total time — which is me!). I will go in detail on how to get here in the next series. Food for thought As an exercise, formulate an objective function with perhaps more individuals doing the shopping. Can you think of adding any other scenarios to this problem? For demonstration purposes, I’ve kept it relatively straight forward but there are other opportunities to add interesting use cases to this situation. Go to part 2 series to learn more about implementation details! (As as aside, if you have an efficient or better way of solving this, please feel free to comment/suggest!)

How I applied linear modelling to solve a ‘real life’ problem

27

how-i-applied-linear-modelling-to-solve-a-real-life-problem-1c6f2edc99f7

2018-03-09

2018-03-09 06:21:02

https://medium.com/s/story/how-i-applied-linear-modelling-to-solve-a-real-life-problem-1c6f2edc99f7

false

1,055

null

Machine Learning

machine-learning

Machine Learning

51,320

Megha Bambra

null

bdded3311deb

mbambra

206

144

20,181,104

null

0

null

0

null

2018-02-09

2018-02-09 03:06:48

2018-02-09

2018-02-09 03:10:48

5

false

en

2018-02-09

2018-02-09 03:10:48

0

1c6fc11da7d8

2.848428

8

0

Sometimes it can be challenging to quantify how sure a prediction is.

2

Residual Machine Learning: Continuous as Categorical Sometimes it can be challenging to quantify how sure a prediction is. When asked to predict a continuous variable (e.g. how much energy will be consumed tomorrow?) many machine learning algorithms will predict a single result without offering any guidance on how certain that result is. This can be a problem if you want to understand the likelihood of passing key thresholds (e.g. will peak demand exceed capacity?), evaluate risk (e.g. what is the best case / worst case of a forecast storm), or highly uncertain events (e.g. individual energy consumption). Individual energy consumption doesn’t follow a nice, smooth curve for individual customers; it fluctuates wildly for a variety of reasons: <Figure 1, individual energy consumption Fitting a distribution to the errors of a model can be a solution, unless there’s a pattern, (i.e. heteroskedasticity). Individuals are far more predictable at 4am than they are at 7pm: <Figure 2, residential energy consumption heat map> While it would be possible to build a separate model for each hour, the sheer number of heteroskedastic factors involved in individual consumption (hour of the day, day of the week, weather, ownership of PV system, an air conditioner, swimming pool etc) makes it impractical to split the data into enough groups for separate models. One underutilised solution is to either split target variables, or the residuals from another model, into groups (sometimes referred to as bins) and modelling the bins as categorical variables. This approach has two benefits: exploring the uncertainty around the most likely prediction by examining the probabilities of the other categories, and automatically handling latent variables. Latent variables occur when there is something you don’t or can’t measure that is influencing the result. Consider the electricity consumption of a hypothetical factory: <Figure 3, factory consumption> The pattern here is quite strong, except when the machinery doesn’t turn on. There could be any number of reasons why this happens: maybe there were no orders that day, maybe something broke, maybe everyone is at a party. We’ll never know what happened, or if or when it will happen again, but our model will implicitly factor these events into the predictions: <Figure 4, factory model> An important consideration is how big the residual bins should be, too small and the predictions may not be reliable, too big and they may not be useful. One option is to use quantiles (perhaps deciles or percentiles) to give a nice, even spread to the number of records in each bin. Another option is to convert the continuous value into a Z score (divide by the mean and subtract the standard deviation to form a bell curve) and then apply a formula like this: round(10^(round(log10(abs(residual)), 0.01)), 0.1) which gradually increases the size of the steps as you move away from the mean, giving you more precision where it’s likely to be important: <Figure 5, step chart> The big learning is that with a model of the residuals it is possible answer in terms of confidence, precisely identifying the risk of key thresholds being passed. This allows better understanding where the resources are likely to be needed; improving the prioritisation of maintenance and upgrades, and minimising unnecessary spending.

Residual Machine Learning: Continuous as Categorical

10

residual-machine-learning-continuous-as-categorical-1c6fc11da7d8

2018-02-12

2018-02-12 02:13:32

https://medium.com/s/story/residual-machine-learning-continuous-as-categorical-1c6fc11da7d8

false

534

null

Machine Learning

machine-learning

Machine Learning

51,320

Pete Condon

null

b962640d199b

pete.b.condon

5

4

20,181,104

null

0

null

0

null

2018-04-16

2018-04-16 18:02:52

2018-04-23

2018-04-23 13:55:10

17

false

en

2018-04-23

2018-04-23 13:55:10

0

1c7048b9618d

7.324528

75

1

Introduction

4

Why MobileNet and Its Variants (e.g. ShuffleNet) Are Fast Introduction In this article, I give an overview of building blocks used in efficient CNN models like MobileNet and its variants, and explain why they are so efficient. In particular, I provide intuitive illustrations about how convolution in both spatial and channel domain is done. Building Blocks Used in Efficient Models Before explaining specific efficient CNN models, let’s check computational cost of building blocks used in efficient CNN models, and see how convolution are performed in spatial and channel domain. Letting HxW denotes the spatial size of input feature map, N denotes the number of input channels, KxK denotes the size of convolutional kernel, and M denotes the number of output channels, the computational cost of standard convolution becomes HWNK²M. An important point here is that the computational cost of the standard convolution is proportional to (1) the spatial size of feature map HxW, (2) the size of convolution kernel K², (3) the numbers of input and output channels NxM. The above computational cost is required when convolution is performed on both spatial and channel domain. CNNs can be speeded up by factorizing this convolution as shown below. Convolution First of all, I provide an intuitive illustration about how convolution in both spatial and channel domain is done for standard convolution, whose computational cost is HWNK²M. I connect lines between input and output to visualize dependency between input and output. The number of lines roughly indicate the computational cost of convolution in spatial and channel domain respectively. For instance, conv3x3, the most commonly used convolution, can be visualized as shown above. We can see that the input and output are locally connected in spatial domain while in channel domain, they are fully connected. Next, conv1x1 [1], or pointwise convolution, which is used to change the size of channels, is visualized above. The computational cost of this convolution is HWNM because the size of kernel is 1x1, resulting in 1/9 reduction in computational cost compared with conv3x3. This convolution is used in order to “blend” information among channels. Grouped Convolution Grouped convolution is a variant of convolution where the channels of the input feature map are grouped and convolution is performed independently for each grouped channels. Letting G denote the number of groups, the computational cost of grouped convolution is HWNK²M/G, resulting in 1/G reduction in computational cost compared with standard convolution. The case of grouped conv3x3 with G=2. We can see that the number of connections in channel domain becomes smaller than standard convolution, which indicates smaller computational cost. The case of grouped conv3x3 with G=3. The connections become more sparse. The case of grouped conv1x1 with G=2. Thus, conv1x1 can also be grouped. This type of convolution is used in ShuffleNet. The case of grouped conv1x1 with G=3. Depthwise Convolution In depthwise convolution [2,3,4], convolution is performed independently for each of input channels. It can also be defined as a special case of grouped convolution where the numbers of input and output channels are same and G equals the number of channels. As shown above, depthwise convolution significantly reduces the computational cost by omitting convolution in channel domain. Channel Shuffle Channel shuffle is an operation (layer) which changes the order of the channels used in ShuffleNet [5]. This operation is implemented by tensor reshape and transpose. More precisely, letting GN’ (=N) denote the number of input channels, the input channel dimension is first reshaped into (G, N’), then transpose (G, N’) into (N’, G), and finally flatten into the same shape as input. Here, G represents the number of groups for grouped convolution, which is used together with channel shuffle layer in ShuffleNet. While the computational cost of channel shuffle can not be defined in terms of the number of multiply-add operations (MACs), there should be some overhead. The case of channel shuffle with G=2. Convolution is not performed, and simply the order of the channels is changed. The case of channel shuffle with G=3. Efficient Models In the following, for efficient CNN models, I provide intuitive illustrations about why they are efficient and how convolution in both spatial and channel domain is done. ResNet (Bottleneck Version) Residual unit with bottleneck architecture used in ResNet [6] is a good start point for further comparison with the other models. As shown above, a residual unit with bottleneck architecture is composed of conv1x1, conv3x3, and conv1x1. The first conv1x1 reduces the dimension of the input channel, reducing the computational cost of subsequent relatively expensive conv3x3. The final conv1x1 recover the dimension of the output channel. ResNeXt ResNeXt [7] is an efficient CNN model, which can be seen as a special case of ResNet whose conv3x3 is replaced by grouped conv3x3. By using efficient grouped conv, the channel reduction rate in conv1x1 becomes moderate compared with ResNet, resulting in better accuracy with the same computational cost. MobileNet (Separable Conv) MobileNet [8] is a stack of the separable convolution modules which are composed of depthwise conv and conv1x1 (pointwise conv). The separable conv independently performs convolution in spatial and channel domains. This factorization of convolution significantly reduces the computational cost from HWNK²M to HWNK² (depthwise) + HWNM (conv1x1), HWN(K² + M) in total. In general, M>>K² (e.g. K=3 and M ≥ 32), the reduction rate is roughly 1/8–1/9. The important point here is that the bottleneck of the computational cost is now conv1x1! ShuffleNet The motivation of ShuffleNet is the fact that conv1x1 is the bottleneck of separable conv as mentioned above. While conv1x1 is already efficient and there seems to be no room for improvement, grouped conv1x1 can be used for this purpose! The above figure illustrates the module for ShuffleNet. The important building block here is the channel shuffle layer which “shuffles” the order of the channels among groups in grouped convolution. Without channel shuffle, the outputs of grouped convolutions are never exploited among groups, resulting in the degradation of accuracy. MobileNet-v2 MobileNet-v2 [9] utilizes a module architecture similar to the residual unit with bottleneck architecture of ResNet; the modified version of the residual unit where conv3x3 is replaced by depthwise convolution. As you can see from the above, contrary to the standard bottleneck architecture, the first conv1x1 increases the channel dimension, then depthwise conv is performed, and finally the last conv1x1 decreases the channel dimension. By reordering the building blocks as above and comparing it with MobileNet-v1 (separable conv), we can see how this architecture works (this reordering does not change the overall model architecture because the MobileNet-v2 is the stack of this module). That is to say, the above module be regarded as a modified version of separable conv where the single conv1x1 in separable conv is factorized into two conv1x1s. Letting T denote an expansion factor of channel dimension, the computational cost of two conv1x1s is 2HWN²/T while that of conv1x1 in separable conv is HWN². In [5], T = 6 is used, reducing the computational cost for conv1x1 by a factor of 3 (T/2 in general). FD-MobileNet Finally, I introduce Fast-Downsampling MobileNet (FD-MobileNet)[10]. In this model, downsamplings are performed in earlier layers compared with MobileNet. This simple trick can reduce total computational cost. The reason lies in the traditional downsampling strategy and the computational cost of separable conv. Starting with VGGNet, many models adopt the same downsampling strategy: perform downsampling and then double the number of channels of subsequent layers. For standard convolution, the computational cost does not change after downsampling because it is defined by HWNK²M. However, for separable conv, its computational cost becomes smaller after downsampling; it is reduced from HWN(K² + M) to H/2 W/2 2N(K² + 2M) = HWN(K²/2 + M). This is relatively dominant when M is not so large (i.e. earlier layers). I end up this article with the following cheat sheet, thank you :P References [1] M. Lin, Q. Chen, and S. Yan, “Network in Network,” in Proc. of ICLR, 2014. [2] L. Sifre, “Rigid-motion Scattering for Image Classification, Ph.D. thesis, 2014. [3] L. Sifre and S. Mallat, “Rotation, Scaling and Deformation Invariant Scattering for Texture Discrimination,” in Proc. of CVPR, 2013. [4] F. Chollet, “Xception: Deep Learning with Depthwise Separable Convolutions,” in Proc. of CVPR, 2017. [5] X. Zhang, X. Zhou, M. Lin, and J. Sun, “ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,” in arXiv:1707.01083, 2017. [6] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in Proc. of CVPR, 2016. [7] S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He, “Aggregated Residual Transformations for Deep Neural Networks,” in Proc. of CVPR, 2017. [8] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient Convolutional Neural Networks for Mobile Vision Applications,” in arXiv:1704.04861, 2017. [9] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. Chen, “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” in arXiv:1801.04381v3, 2018. [10] Z. Qin, Z. Zhang, X. Chen, and Y. Peng, “FD-MobileNet: Improved MobileNet with a Fast Downsampling Strategy,” in arXiv:1802.03750, 2018.

Why MobileNet and Its Variants (e.g. ShuffleNet) Are Fast

304

why-mobilenet-and-its-variants-e-g-shufflenet-are-fast-1c7048b9618d

2018-06-19

2018-06-19 05:38:54

https://medium.com/s/story/why-mobilenet-and-its-variants-e-g-shufflenet-are-fast-1c7048b9618d

false

1,517

null

Machine Learning

machine-learning

Machine Learning

51,320

Yusuke Uchida

null

aa28d0c8f0c5

yu4u

98

43

20,181,104

null

0

null

0

9c321fdafce5

2017-09-18

2017-09-18 08:06:18

2017-09-18

2017-09-18 08:12:05

4

false

en

2017-09-20

2017-09-20 13:10:36

10

1c729ca00d39

3.330189

6

0

The march of the chatbot continues with users as diverse as Barcelona Football Club and Citi Bank in Singapore launching interactive…

5

Chatbots rising globally as proof of cost and time savings mount up The march of the chatbot continues with users as diverse as Barcelona Football Club and Citi Bank in Singapore launching interactive avatars for people to converse with. As more companies reveal the financial and time saving benefits of bots, more will follow. For the people who ignore any technology until they can no longer escape, the noose is tightening when it comes to chatbots. They are appearing in our health services, sporting obsessions and finance institutes. Soon there won’t be anywhere to hide from them, as growing popularity and demonstrations of fiscal value continue to pop up. One of the leading benefits is that services are linking chatbots to their communications chain, making it a part of the process rather than an outlying novelty. So, while UK medical patients could start off talking to a chatbot, those with reason to take a query further can soon be talking to a medical professional who can better assess any concerns and provide next steps. In Spain, millions of FC Barcelona fans will soon be able to vote for their man of the match and enjoy other interactions, helping bring the world of the chatbot to sports fans in a format that is easy to enjoy and replicate. Perhaps the key role they will take in our daily lives is in finance, with banks rolling them out at pace. In Asia, Citi is going big with a chatbot currently in testing. They plan to add features including card activation, locking and unlocking of credit cards plus having bots show transaction alerts for cards. In the UK, the first car insurance Chatbot is on the way with Co-Op helping to provide quotes without having to talk to a customer service agent. Also key to the success of chatbots will be high profile placement on services like Facebook, and the the arrival of Business Chat on Apple’s iOS 11 offering businesses chat via Apple iMessage for a more direct line to customer support. Any company looking to deploy a chatbot should ensure it is available across the widest user bases. Products like Snatchbot deploy to a range of apps, sites and social media services. It’s not all business though, as Schwartz the spice seller has got a chatbot up on Messenger that will take a few ingredients and come up with the right recipe and spice combination for hungry indecisive types. With users comes a world of statistics Businesses love some hard data when it comes to investing in new technology. The slow rise in smart cities is largely due to the huge sums involved, but also because productivity or beneficial data is being protected or obscured by the cities or vendors. That leaves others waiting to see if smart cities are really are all they are cracked up to be before investing. For chatbots, the results are easier to come by and are starting to make impressive reading. With the rise in use comes the needed statistics to drive other businesses to get into the chatbot way. In India, HDFC Bank announced that its chatbot had helped address 2.7 million user queries in six months. It has had 1.2 million conversations, talking to around one-in-five of the bank’s customers, a massive saving in human time. In 2017, 1–800-Flowers launched its own GWYN bot, and earlier this year the company revealed it added around $10 million of revenue. While looking forward, National Australia Bank reckons its new chatbots will save the bank $16 million, dealing with simple items and freeing up account handlers for more serious issues. These non-trivial sums will pique the interest of others and help launch further generations of chatbot, all doing their bit for the digital economy. Our survey says: A recent survey shows that users are increasingly keen to use chatbots. Also, a raft of aggregated statistics show that chatbots can save four minutes per call, could be used for around $55 in purchases before customers want to start getting a better look at goods, and could be responsible for 85% of customer interactions by the end of the decade.

Chatbots rising globally as proof of cost and time savings mount up

70

chatbots-rising-globally-as-proof-of-cost-and-time-savings-mount-up-1c729ca00d39

2018-04-25

2018-04-25 20:41:34

https://medium.com/s/story/chatbots-rising-globally-as-proof-of-cost-and-time-savings-mount-up-1c729ca00d39

false

697

Follow us to explore a new way of storytelling.

null

iampopin

null

I AM POP

hello@iampop.in

iampop

FACEBOOK MESSENGER,MARKETING,CHATBOTS,MUSIC BUSINESS,TECH

iampopin

Chatbots

chatbots

Chatbots

15,820

Chris Knight

Tech writer interested in mobile, digital business, IT, smart homes and gadgets - anything with a GHz pulse.

56e4127fc713

ChrisKnightcms

68

5

20,181,104

null

0

def content_loss(base, combination): return K.sum(K.square(combination - base)) def gram_matrix(x): features = K.batch_flatten(K.permute_dimensions(x, (2, 0, 1))) gram = K.dot(features, K.transpose(features)) return gram def style_loss(style, combination): S = gram_matrix(style) C = gram_matrix(combination) channels = 3 size = img_height * img_width return K.sum(K.square(S - C)) / (4. * (channels ** 2) * (size ** 2)) def total_variation_loss(x): a = K.square( x[:, :img_height - 1, :img_width - 1, :] - x[:, 1:, :img_width - 1, :]) b = K.square( x[:, :img_height - 1, :img_width - 1, :] - x[:, :img_height - 1, 1:, :]) return K.sum(K.pow(a + b, 1.25)) from keras import backend as K target_image = K.constant(preprocess_image(target_image_path)) style_reference_image = K.constant(preprocess_image(style_reference_image_path)) # This placeholder will contain our generated image combination_image = K.placeholder((1, img_height, img_width, 3)) # We combine the 3 images into a single batch input_tensor = K.concatenate([target_image, style_reference_image, combination_image], axis=0) # We build the VGG19 network with our batch of 3 images as input. # The model will be loaded with pre-trained ImageNet weights. model = vgg19.VGG19(input_tensor=input_tensor, weights='imagenet', include_top=False) outputs_dict = dict([(layer.name, layer.output) for layer in model.layers]) content_layer = 'block5_conv2' style_layers = ['block1_conv1', 'block2_conv1', 'block3_conv1', 'block4_conv1', 'block5_conv1'] total_variation_weight = 1e-4 style_weight = 1.0 content_weight = 0.025 # Define the loss by adding all components to a `loss` variable loss = K.variable(0.) layer_features = outputs_dict[content_layer] target_image_features = layer_features[0, :, :, :] combination_features = layer_features[2, :, :, :] loss += content_weight * content_loss(target_image_features, combination_features) for layer_name in style_layers: layer_features = outputs_dict[layer_name] style_reference_features = layer_features[1, :, :, :] combination_features = layer_features[2, :, :, :] sl = style_loss(style_reference_features, combination_features) loss += (style_weight / len(style_layers)) * sl loss += total_variation_weight * total_variation_loss(combination_image) # Get the gradients of the generated image grads = K.gradients(loss, combination_image)[0] # fetch the values of the current loss and the current gradients fetch_loss_and_grads = K.function([combination_image], [loss, grads]) class Evaluator(object): def __init__(self): self.loss_value = None self.grads_values = None def loss(self, x): assert self.loss_value is None x = x.reshape((1, img_height, img_width, 3)) outs = fetch_loss_and_grads([x]) loss_value = outs[0] grad_values = outs[1].flatten().astype('float64') self.loss_value = loss_value self.grad_values = grad_values return self.loss_value def grads(self, x): assert self.loss_value is not None grad_values = np.copy(self.grad_values) self.loss_value = None self.grad_values = None return grad_values evaluator = Evaluator() from scipy.optimize import fmin_l_bfgs_b from scipy.misc import imsave import time result_prefix = 'vgg19_try1' iterations = 10 x = preprocess_image(target_image_path) x = x.flatten() for i in range(iterations): print('Start of iteration', i) start_time = time.time() x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x, fprime=evaluator.grads, maxfun=10) print('Current loss value:', min_val) # Save current generated image img = x.copy().reshape((img_height, img_width, 3)) img = deprocess_image(img) fname = result_prefix + '_at_iteration_%d.png' % i imsave(fname, img) end_time = time.time() print('Image saved as', fname) print('Iteration %d completed in %ds' % (i, end_time - start_time)) model = vgg16.VGG16(input_tensor=input_tensor, weights='imagenet', include_top=False) def preprocess_image(image_path): img = load_img(image_path, target_size=(img_height, img_width)) img = img_to_array(img) img = np.expand_dims(img, axis=0) img = vgg16.preprocess_input(img) return img model = resnet50.ResNet50(input_tensor=input_tensor, weights='imagenet', include_top=False, pooling='max') content_layer = 'res5b_branch2a' style_layers = ['res3a_branch2a','res4a_branch2a','res5a_branch2a'] total_variation_weight = 0 style_weight = 400000 content_weight = 0.0001

26

e36da56c4596

2018-05-05

2018-05-05 03:36:12

2018-05-10

2018-05-10 03:37:30

22

false

en

2018-05-10

2018-05-10 04:07:24

1

1c750c0cb458

8.208491

4

0

Experimenting with neural style transfer with VGG16, VGG19, and ResNet50 pre-trained models. Code for this project can be found here.

5

Image Style Transfer using Pre-trained ConvNets Experimenting with neural style transfer with VGG16, VGG19, and ResNet50 pre-trained models. Code for this project can be found here. Neural style transfer is an image-modification technique (first developed and introduced by Leon Gatys et al) that applies the style of one image to the content of another image. In this blog post, we’ll look at how we can apply this using a pre-trained ConvNet model from Keras. An example of neural style transfer. Above is a concrete example of neural style transfer. In this scenario, style is defined as the textures, colors, and general patterns (such as brush strokes) found in an image. Van Gogh’s ‘The Starry Night’ is a popular choice as a style image because of its obvious style. The key to successful style transfer lies behind the loss function. We want to minimize the content loss between the content image and the generated image, while also minimizing the style image between the style image and the generated image. Let’s first take a look at content loss: Content Loss: Earlier layers in our networks are associated with more local information, while the higher layers will contain activations with global information. Because content is defined by the macrostructure of the image (for example the building structure in the example above), the topmost layers in our network will capture image content. Thus, we’ll simply compute the content loss as the L2 distance (sum of square differences) between the activations of a top layer computed over the content image and the generated image. Style Loss: To capture style loss, we’ll be using the Gram matrix method outlined in the original 2015 style transfer paper. The Gram matrix is the inner product of the feature map with a given layer. This results in a map of correlations between features of a layer. These feature correlations are what capture the texture patterns at a certain spatial scale, which correspond to the physical appearance of textures/colors/patterns at this scale. Thus, the style loss function aims to minimize the difference between feature correlations in each activation of a layer from the style image to the generated image. We’ll also be using total variation loss, which is used to encourage spacial continuity and discourage over-pixilation: Now we’re ready to start our style transfer. These are the general steps we’ll be taking for each of the next 3 parts in this post: Set up the network to compute activations for style, content, and generated images at the same time. Use these activations to compute loss. The total loss will be a weighted average of style_loss, content_loss, and total_variation_loss . Use gradient descent to minimize the loss function to obtain the final generated image. Part 1 using VGG19 We’ll start by setting up our model. We batch together our style, content, and generated image (placeholder for now). This allows us to compute the activations for all three at once. We’ll pass in our combined input_tensor as an argument to our VGG19 model in order to change the shape of the layers accordingly. Next, we’ll define which layers to use for defining content and style. 'block5_conv2' is a top level Conv2D layer in our VGG19 model, so we’ll use it to evaluate content. As for style_layers , we’ll use a variety of Conv2D layers ranging from the top to bottom of the network in order to calculate style error. We’ll also assign weights to style, content, and total-variation which will be used to calculate the weighted total loss. We’ll also be using a Python class Evaluator . This class allows us to compute the loss function and gradients at the same time. Essentially it removes the redundant computations when computing these two values and and speeds up our program by a factor of 2. Now we’ll run our gradient descent process using SciPy’s L-BFGS algorithm. We’ll run 10 iterations, where each iteration consists of 10 steps of gradient descent. We also save the generated image at each iteration so we can track our progress. Here are the results: Content, Style, and Generated image for VGG19 Part 1 (left to right). Generated images at iterations 1, 4, and 9 (left to right). The style transfer worked decently well. However, the style image we’ve chosen above isn’t great. It lacks clearly defined patterns and textures. Let’s try again with a different model and a new style image. Part 2 using VGG16 This time we’ll be using the VGG16 model from Keras shown below: We’ll also have to update our preprocess_image function to reflect this change: We’ve also increased the style_weight and decreased content_weight in an attempt to receive better results (also to make up for downgrading to VGG16 from VGG19). Here are the results: Content, Style, and Generated images for VGG16 Part 2 (left to right). Generated images at iterations 1, 4, and 9 (left to right). We are able to achieve great results. The content of the image is definitely preserved, as well as the colors and textures found in the style image. Part 3 with ResNet50 This time, we’ll be attempting style transfer using ResNet50 from Keras: On our first attempt, we’ll be using the following values: These are the results after 10 iterations: Content, Style, and Generated images for ResNet50 Part 3 (left to right). Generated images at iterations 1, 4, and 9 (left to right). Unfortunately, we’ve failed the capture the full style of the style image. We see the beginnings of texture inspired by the style image. However, after examining the generated image, I realized many of the colors found in the style image weren’t found in content image. We can see a majority color of the content image is some shade of blue. Because the style image doesn’t contain much blue, there isn’t much style that can be applied from the style image to the content image. Thus, we end up with a generated image that looks half-finished. Tip: Make sure to use a style image with a similar color palette to your content image (otherwise style won’t be applied well). Going back to the image we used in Part 2, we get the following results: Content, Style, and Generated images for ResNet50 Part 3 Trial 2 (left to right). We’ve done slightly better. However, it is clear that ResNet50 is much worse at style transfer than the VGG networks. Here are some reasons why that may be: The VGG networks are very big (both are 500MB+ compared to ResNet50’s 99MB) — thus could incidentally capture and store more information than other models. VGG’s are relatively shallow and modular — there are no residual connections or shortcuts to skip raw data through layers. This results in a clear, hierarchical series of abstractions. ResNet’s on the other hand are too spread-out through layers; the individual features exist but can be mixed up through the many layers of abstraction. VGG doesn’t downsample as aggressively compared to other models (max pooling layers only after multiple convolutions). Future Work: Here are some proposed ways to test the reasons given above: Train much bigger ResNet’s/much smaller VGG’s to see if the gap in performance shrinks. If a small VGG can’t perform style transfer better than a ResNet of the same size, this suggests VGG’s advantage lies in the size of the model. Experiment with how features are calculated in the ResNet. Because the features may be spread out thin across many layers, create a method to sum feature maps depth-wise before calculating the Gram Matrix. Brute-force a lot of combinations for content and style layers for ResNet50. I’ve already tried a good of combinations (this may be image specific as in different style layers work better for different types of images).

Image Style Transfer using Pre-trained ConvNets

25

image-style-transfer-using-pre-trained-convnets-1c750c0cb458

2018-06-07

2018-06-07 05:50:38

https://medium.com/s/story/image-style-transfer-using-pre-trained-convnets-1c750c0cb458

false

1,685

It can be challenging to understand and keep up with the ever-advancing field of Artificial Intelligence. As a student, I attempt to build a series of projects to experiment with and better understand AI concepts through concrete examples.

null

The College Coder

khnoh@brown.edu

training-convnet-on-small-binary-classified

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

null

Machine Learning

machine-learning

Machine Learning

51,320

Ken Noh

Rising Junior at Brown University (CS), Software Engineering Intern at AIBrain.

b2bd22b3443e

kyung_hyun_noh

10

20

20,181,104

null

0

null

0

54ae3167fa36

2017-09-11

2017-09-11 15:03:02

2017-09-13

2017-09-13 21:45:41

6

false

en

2017-09-16

2017-09-16 23:29:10

3

1c7596065e48

7.146226

2

0

It’s been a while since we are building our algorithms here at Decissio and finally it was about time to take them out for a spin with a…

5

We tried our algorithms in the real world and they worked: Our First Pilot It’s been a while since we are building our algorithms here at Decissio and finally it was about time to take them out for a spin with a real use case. For that purpose, we got in touch with the best Central European accelerator: StartupYard. In this post we’ll go over some of the key findings from our experiments and how they can help investors and founders think about the early stages of the startup process. You can see StartupYard’s evaluation of our process here StartupYard had over a hundred applications to sort through in order to get to its six final entrants. 43 of the most promising applicants were chosen for a Skype interview after which eventually six went on to participate in StartupYard’s program. Utilizing our research of previous datasets from other accelerators we have and the dataset provided by StartupYard. We created our own ranking system for the recent applicants to StartupYard. We wanted to see how do our system stands up to StartupYard’s extensive selection process with a professional jury. [The selection accelerators have at hand involves data and several interviews. On the other hand we only had the information given by the applications. That information was augmented by our own data collection, categorization and data management server.] Our initial task was to correctly classify applicants into domains that StartupYard was interested in AI, blockchain, IoT, robotics, and virtual reality. We had the names of the startups, their teams, duration of their projects, revenue, and funds invested to date (when available), and answers to questions built into the application that described their business concepts and the backgrounds of the team involved. Our process produced about 200 variables for each startup that told the story of that company, their founders and their progress. Which we used to automate the initial screening part of the evaluation process. Instead of human evaluators at StartupYard going and screening using interviews, we automated the process to save StartupYard’s staff a tremendous amount of time otherwise devoted to sorting through subpar applications. The applicants themselves, stand to benefit. A star applicant will hopefully not have to worry about their application being missed simply because a human evaluator is exhausted after reading hundreds of applications that day. Cleaning out the mess We started by measuring the completeness of each application. StartupYard has a detailed application form that expects applicants to have well developed understandings about everything from their business plans, their end users, and how they differentiate themselves from their competitors. One obvious way to cut down on the number of applications a human evaluator would potentially have to go through is to sort the applications based off on how complete the application is. Our first step was recording the percentage of questions answered and sorting our applications accordingly. We also calculated into this score the presence of working social media accounts, websites, video advertising, profiles, and other pieces of their digital trace. Distribution of raw percentage of application completion within our dataset. The median represents the middle of our dataset where half the data is on one side or the other. Q25 is the 25th quartile. To the left of that point we have the bottom 25% in terms of completing their applications. One of the first thing we noticed was that while for the most part applicants filled out the information requested of them, a portion didn’t, 7.8% of applicants filled out less than 60% of the available fields. At 60% these applicants had left many crucial pieces of information missing and possibly weren’t taking the application very seriously. We sorted these to the bottom of the evaluator’s pile. But this methodology alone doesn’t take into account the content of each response, just whether or not there was a response. Ideally we would like analyze the content of each response in order to gauge its relevance to what StartupYard is looking for. Luckily we didn’t have to conceptualize an AI literary critic to solve this problem. We had taken into account incomplete applications but now we want to sort out the poorly written ones. During our initial conversations with StartupYard, we learned that grammar and spelling mistakes were common in many applications and a source of frustration for human evaluators. We analyzed applicants responses for spelling and grammar errors. The median applicant had 11 spelling mistakes throughout their application and we can see in the visualization below that this is pretty standard for most. To err is human and even though we took into account proper nouns that could have set off false positives in our spell check program, for the average applicant a few misspelled words are the norm. Kernel Density plot showing the distribution of applicant spelling errors. Q25 marks the bottom 25% of the applicants in terms of spelling errors. In other words the top 25% of our applicants in terms of not making spelling mistakes. The y axis is the proportion of the total applicant pool with the number of spelling mistakes listed on the x axis. Looking at the distribution above we can see that there is a skew to the left in our distribution. In the bottom 25% of applicants in terms of spelling mistakes the number of mistakes descend rapidly. It would be reasonable to suppose that in more complete applications there would be more spelling errors. You can’t make spelling mistakes in questions you don’t answer. But our analysis shows that this isn’t the case with our datasets, where the two scores were slightly negatively correlated. One hypothesis could be that applicants who took more chances at making a spelling mistake by answering more questions, were more likely to have the sense to use spell check. The conscientious are consistent in that regard. We also checked the applications to see if they fit into technology domains StartupYard was interested in. Applicants were asked to self describe their domains but we went a step further. With our algorithms we analyzed the vocabulary used by applicants. We have managed to get a robust algorithm that can pinpoint the category that a startup falls in based on the vocabulary they use to describe themselves. Scaling these words appropriately, we created a robust score of how relevant an applicant was to their claimed field of expertise. We brought the most relevant to the top and punished the rest. Top legend shows domain when categorized by keywords associations. The length of each bar indicates the number of applicants classified by keywords according to their own self designations. For example there are 8 applicants who designated themselves as VR/AR companies but our methodology would classify them as “Other” which is our designation for applicants who do not fit into a domain based on keyword usage. From our own analysis a large chunk of startups don’t articulate their self described domain through their applications. While talking to StartupYard we discussed that, while anyone can claim to be an AI or VR firm what was important to StartupYard is that they actually demonstrate competence in their applications. Rather than simply trusting applicants to accurately self-describe their domain, Decissio does its own Natural Language Processing evaluation to ferret out a startup’s real tech expertise or lack thereof. After these previous factors were analyzed we took into account more of the hard numbers behind each applicant. StartupYard asked applicants how long they and their teams had been working on their startups. We transformed the number of man hours completed over the lifetime of the startup into a normalized score, which we built into what we call the effort ratio. Each individual line represent a single applicant and their effort score on the bottom of the figure. The 90th percentile is labeled as Q90 and marked by a red line. Everything to the right of the Q90 mark is in the top 10 percent in terms of effort scores. What we see here is a large rightward skew in the dataset. As most applicants are early stage startups their combined hours, transformed into an effort score largely hover around the low end of the scores. But there is a significant subset of applicants who worked on their project far more than other applicants. For some investors this may be a useful metric to follow while for others not so much. Decissio is here to help bring clarity to your investment decisions, and with this metric investors can get a fuller picture of their business prospects. Some applicants already had established revenue streams and previous investments. We automatically tabulated the ratio of revenue to investments used them for our analysis of the applicants, favoring startups with a better record of turning investments into revenue. We did this while still taking into account startups that had received investor money but were pre-revenue at this stage. Like with effort scoring, this is a metric that investors can filter when looking at their prospects. Below is a chart of applicant reported revenue per investments: This chart should be interpreted in a similar manner as the above. One line equals on applicant, the vast majority of applicants are overlapping lines below the 90th percentile. The top 10% of our distribution, everything to the right of the the Q90 mark is largely a return on investment (a score of 1 to 1 on revenue to investment dollars). There are some applicants claiming to be making much higher returns. Through our own current capabilities investors can not only filter prospects based off of previous ROI, but also quickly and easily analyze a prospect financials, in the context of other factors. As a final variable we utilized our own Natural Language Processing techniques to tabulate a media mentions score for each applicant, both the team members and the startup as an organization. We pulled information from hundreds of news startup and business sources to create a score based on the sentiment, quantity, and quality of the media mentions. In this dataset applicants either had either positive media exposure or none, so we don’t see any negative media score. For investors using Decissio, our media mentions score can inform research into firms that have been shortlisted for consideration. A positive media score could indicate to an investor that an uninspiring prospect on paper, may be worth looking into if they had managed to attract positive media attention. Also a negative media score could signal that a promising startup could be hiding something that the media had already picked up on. Taking a step back within each of our metrics we can see that within StartupYard’s applicant pool a normalized set of data hovering around the median but have significant groups of outliers that stand way apart from other applicants. While the average applicant had an average revenue per investment ratio of .44, the top 5% had a ratio of 2. In all of our scores we have tracked the notable few that outperform weather that’s in terms of effort, ROI, or publicity. On the flipside we have tracked those on the opposite end when it comes to sloppiness in their communications (spelling mistakes, domain classification, completeness scores) or bad publicity. In the end for each investor these metrics will mean different things, but with Decissio these metrics will be at the fingertips of investors.

We tried our algorithms in the real world and they worked: Our First Pilot

2

findings-from-recent-collaboration-with-startupyard-1c7596065e48

2018-05-29

2018-05-29 19:28:07

https://medium.com/s/story/findings-from-recent-collaboration-with-startupyard-1c7596065e48

false

1,642

We help early stage investors make better investment decisions

null

Decissio

dite@decissio.com

decissio

null

Startup

startup

Startup

331,914

Corey Kiyoshi Clippinger

null

ff3446a56eb9

coreykiyoshiclippinger

41

50

20,181,104

null

0

null

0

bdf435deaf29

2018-05-01

2018-05-01 21:59:17

2018-05-02

2018-05-02 13:26:51

6

false

en

2018-11-02

2018-11-02 16:31:29

6

1c75cfdb108b

6.123585

21

1

0

Today we launch a 30 day pre-sale campaign for our new robot, the Misty II…

5

Misty II — A Powerful and Refined Robot It feels like every day is an exciting day at Misty, however, today is one of those few exceptionally exciting days! Today we launch a 30 day crowdfunding campaign for our new robot, the Misty II. The Misty II incorporates all of the learnings we’ve been collecting from the Misty I Developer Edition that started shipping earlier this year and puts them into a highly refined, fully injection-molded and manufactured body — on top of more great features. So, what have we learned from Misty I and how? Most of what we’ve learned has come from people like you who’ve joined our community, attended our Robothons, and spent time developing for the Misty I. These people range from professional software engineers, to artists, to mobile developers, to high school students, and everyone in between. Here are a few of the top things we learned and have iterated on: Robustness — The robot needs to be solid and well built. Misty II is entirely injection molded with most parts cast in polycarbonate. Voice Input — We learned that people really want high quality voice interaction with their robot. With a three MEMS microphone far-field array that’s capable of beam forming, we’re putting a lot more energy into making sure that voice interaction works great and is easy to use. High Fidelity Audio System + Bass Port — Great sound is really important. We hired an audio engineer and went through many iterations to get to a design that is loud and clear. Automatic Charging — It’s a pain to have to always plug and unplug your robot. Misty II can now run skills at any time and charge herself when her battery gets low. For instance, if you want to write a skill that can check on your pet while you’re away from home, Misty will be ready to go. Misty II automatically charges herself Edge Detection — People want to know Misty is safe. If left unattended, Misty II won’t fall down your stairs. Carrying Handle — Moving your robot can be difficult. The new carrying handle on the back of Misty’s head makes this easy. Touch — More interaction with Misty is better. We added six capacitive touch panels to Misty’s head to make this possible. Track Drive — The robot shouldn’t get stuck in the hostile environment of your home. This meant using a track-based drive system Expandability — We learned that people wanted more options for making additional hardware for Misty. Misty II has a magnetic panel on her head where you can add things like a projector. Remove one screw, and her arm pops off if you want to create your own gripper arm. A trailer hitch allows you to tow a custom delivery trailer. And a backpack with serial and USB ports give you power and data to your add-ons. Arduino Backpack — Make existing projects robot enabled. With the Arduino backpack you can simply plug your existing Arduino shields into Misty and load in the same code you wrote before. Manipulation — More than just pushing objects, Misty can use her arms to point, or you can attach a pen to allow her to draw. Expressiveness — In addition to her eyes, Misty II has a three degree of freedom neck, so she can look up, down, side to side, and tilt her head if she’s curious. Personality Engine — Misty II has a lot of personality built in, and we make it easy for you to carry this personality into your skills. If you want, you can even load in your own eyes and sounds to create your own custom personality for Misty. And these are all on top of the features already built into the hardware of Misty I, like: Depth Sensors — For detailed room mapping and navigation. Floor plan of part of our office generated by Misty 3D Version4K Camera — For things like face detection and recognition or creating a telepresense video chat skill. 4.3" LCD Display — To display Misty’s face or whatever data you want to show from your skill. Bump and Time of Flight Sensors — So Misty doesn’t get stuck or run into objects. Two Powerful Snapdragon Processors — One running Windows 10 IoT Core, as the main processor, and the second running Android 7, which handles the navigation and perception systems (but you don’t need to know how any of that works, we make it super easy). But why Misty II? To set the context, our vision is to rapidly progress robotics and eventually put a robot in every home and office — taking all of the mundane tasks out of our lives so we can spend our time doing the things we really want to be doing. We spent a long time doing research on how we could accomplish this and came to the conclusion that we’d have to satisfy four requirements in our products: Has to be friendly, simple to use, and easy to program. Can be iterated on in future versions to allow for even more capabilities to do even more useful things. Has the capability to begin to do real useful stuff for us. Must be able to do many different things. Not single purpose. To expand on items 3 and 4 above — to do useful things you need a mobile platform (otherwise it’s just a really expensive Amazon Echo). You also need navigation so the robot can find its way around from room to room. You need to be able to interact with the robot using voice. Computer vision is important for identifying things and following you around. And finally you need to be able to easily take advantage of all of these features to program skills for the robots (and share them with others to use and build upon). We can’t keep spending weeks just getting the thing to move like we currently do in robotics today. When we looked at existing solutions we found quite a few platforms but none even got close to satisfying our four criteria. Here’s what we found: Many are STEM/educational robots that are geared towards learning programming with sometimes some flexibility on the mechanical side. On the mid to higher end this could be Lego Mindstorms, MakeBlock, Meccano, or the Alpha 1S as examples. They teach coding but you generally can’t do useful things with them. And if you did find something useful it’s very difficult and time consuming for others to duplicate. Creations are also fragile and single purpose. If you want more advanced capabilities you can buy a Roomba that does a task well but you can’t program a Roomba to do anything else. You can also buy other robots that can start to do useful tasks but the price points are well out of reach of consumers in the tens of thousands of dollars for the PR2, Baxter, Pepper, etc. They can also take weeks to learn how to use — even if you’re an expert. Then you have DIY robots where you’re on your own. You have to buy a mobile base, select your processor, sensors, etc and then integrate it all. The closest to a complete solution is the Turtlebot 3 at $1400-$1800 depending on the version but it’s very difficult to use if you aren’t an expert in Linux and ROS. Integrating components is INCREDIBLY difficult and just getting something that can drive around and avoid all of the crazy obstacles you’ll find in a home without getting stuck would take many months. And again, it’s a one off creation that can’t be shared with others. Not to mention, you haven’t even gotten to useful tasks because you’re still trying to figure out how to get it to not get stuck on your socks. Misty is a huge leap forward! We hope you find her friendly and incredibly easy to program. We will iterate on Misty in the future, but she has the capabilities to start to do useful tasks for us today. Today is an exciting day! We have 50% off for the next 30 days to reward our early customers: Click here for the crowdfunding website and an additional $100 off And even if you aren’t ready to buy a robot, please join our community. We would love to hear from you! https://community.mistyrobotics.com ~Ian

Misty II — A Powerful and Refined Robot

540

misty-ii-a-powerful-and-refined-robot-1c75cfdb108b

2018-11-02

2018-11-02 16:31:29

https://medium.com/s/story/misty-ii-a-powerful-and-refined-robot-1c75cfdb108b

false

1,371

The blog of the Misty Robotics team

null

MistyRobotics

null

MistyRobotics

info@mistyrobotics.com

mistyrobotics

ROBOTS,ROBOTICS,TECHNOLOGY,ARTIFICIAL INTELLIGENCE,PROGRAMMING

MistyRobotics

Robotics

robotics

Robotics

9,103

Ian Bernstein

Founder/Head of Product at Misty Robotics. Former Co-Founder/CTO of Sphero. Been building robots since I was 12.

3a730b3a1603

arobodude

187

66

20,181,104

null

0

null

0

a434d803b44f

2018-08-26

2018-08-26 21:28:39

2018-09-06

2018-09-06 20:29:45

9

true

en

2018-10-12

2018-10-12 03:39:25

11

1c75d6b4af92

5.166038

5

0

Motivation

5

Fast Annotation Net: A framework for active learning in 2018 Motivation Machine learning models, specifically deep learning based approaches for computer vision, require training data. Here are two examples: Left: bounding box example. Right: polygon or “semantic segmentation” http://cocodataset.org/#home. Training data, or annotations, is what the deep learning model uses as ground truth in order to learn. On the left we have a bounding box, a common type of annotation. On the right we have two polygons. Polygons take longer to draw as they represent the exact shape of the object. That’s not too hard right? Well what if you want 10,000 images, or 14.2 million? Imagenet https://medium.com/@mozesr/script-to-get-images-from-image-net-org-7fe8592e6650 There are services that offer to do it for you, such as Scale API. Their posted rate for a semantic segmentation image, is $6.40 USD per image. So that 10k image dataset will cost you $64,000. Scale API represents the market rate to deliver a quality annotation. The cost will go up over time as the level of annotations required gets more complex and the level of knowledge and experience required for the human annotation increases. Data is the #1 roadblock to building machine learning models As machine learning models become easier to train and computation power improve, training data becomes the #1 roadblock to building applied machine learning models. What if there was a way we could use the a machine learning model to help reduce that cost? Here’s an example: Left: train image, Center: predicted result shown in dashed lines, Right: Predicted result with error corrected by human On the left, an image that was annotated by a human. In the center is an image predicted by FAN that may be marked as “correct”, no further annotation needed. We just went from having to draw 4 boxes, to simply reviewing the image and marking it complete. On the right, an example where the green marker was missed, and the human adds the green marker label in. Here we had to only add 1 annotation. Or 1/4 the work. We see here is the heart of it — FAN is learning alongside you. As you annotate, you train FAN, and FAN gets better over time, helping you annotate. As you annotate, you train FAN, and FAN gets better over time, helping you annotate. Before we take a deeper dive into how this works, let’s look at some of the prior art. A brief history Active learning, the concept of the human in the loop, has been around for a while. Even as specific as in the context of machine learning, published in 1996 Cohn et. al. Active Learning with Statistical Models According to them: “ The goal of machine learning is to create systems that can improve their performance at some task as they acquire experience or data. … This passive” approach ignores the fact that, in many situations, the learner’s most powerful tool is its ability to act, to gather data, and to in influence the world it is trying to understand. Active learning is the study of how to use this ability effectively.” (emphasis added) Other approaches More recently the approaches generally fall into two buckets: Specific and general. Coarse-to-Fine Annotation Enrichment for Semantic Segmentation Learning by Luo et al. And Efficient Interactive Annotation of Segmentation Datasets with Polygon-RNN++ by Acuna, Ling, Kar, et. el. are two examples. And Efficient Interactive Annotation of Segmentation Datasets with Polygon-RNN++ . by Acuna, Ling, Kar, et. el. Let’s take a closer look at Polygon-RNN++. Their approach is interesting in that they appear to predict the polygon points directly instead of predicting per pixel. This reminds me of the way an objector detector predicts a value for min and maximum points to form a box. It can be “off” by a significant number of pixels and still get a good result. Where as if you had to have every pixel correct, it would be a lot more difficult. And there have been general approaches. For example UC Berkeley: BDD100K: A Diverse Driving Video Database with Scalable Annotation Tooling. 60% reduction in annotation time Here’s an example of results from UC Berkeley’s active learning approach: By Yu et. el. https://arxiv.org/pdf/1805.04687.pdf “As shown in Fig. 3(a), the object detector is able to label 40% of bounding boxes at a minimal cost. On average, the time of drawing and adjusting each bounding box is reduced by 60%.” by Yu et. el. On average, the time of drawing and adjusting each bounding box is reduced by 60%.” — Yu et. el. To put the statistic into context here’s an example of a FAN network trained on a small portion of the Cityscapes data. The image on the left has no pre labels, you have to annotate every box yourself. The image on the right was pre-labeled with FAN (shown in dashed lines). How it works A user annotates images or video A network is trained. Inference is performed and results are fed back into annotation system The user reviews, corrects, and or adds new content Example use cases For annotations: Continually refine an existing large scale model at a significant reduction in annotations needed. Use an already high performing model to handle the majority of existing classes while annotating only new classes. Use the network built through FAN for your own processes. Fast annotation net is an important piece of the puzzle for reducing the cost of annotations. Limitations and failure cases This is still a very new concept and requires a certain general level of machine learning knowledge to get good results There has been less research on the true effectiveness of human + computer ground truth data. It’s a very ill posed questioned so we may never have a definitive answer. (A comparison would be, does a software developer write better quality code with tab based auto complete features? Well since software quality is an ambiguous concept this is hard to define beyond a general answer like “probably”.) The time to correct an annotation can sometimes be just as long to do it in the first place — so in the worst case it’s about the same time. It takes some time and compute resources to train a FAN network and run inference. I demonstrated some of my work on May 24, 2018 at LDV vision summit. Here’s the video: I’m working on making FAN available to everyone through Diffgram. If you are interested in participating in the beta signup here. And if you are interested in working with me on this please reach out to me on LinkedIn. Thanks for reading!

Fast Annotation Net: A framework for active learning in 2018

69

fast-annotation-net-a-framework-for-active-learning-in-2018-1c75d6b4af92

2018-10-12

2018-10-12 03:39:25

https://medium.com/s/story/fast-annotation-net-a-framework-for-active-learning-in-2018-1c75d6b4af92

false

1,051

Diffgram product updates

null

diffgram

null

diffgram

null

diffgram

MACHINE LEARNING,COMPUTER VISION,SELF DRIVING CARS,ANNOTATIONS,TENSORFLOW

diffgram

Programming

programming

Programming

80,554

Anthony Sarkis

https://diffgram.com

149e0bf17d7b

anthony_sarkis

231

74

20,181,104

null

0

null

0

null

2018-03-17

2018-03-17 07:51:30

2018-03-18

2018-03-18 07:51:42

8

false

en

2018-04-05

2018-04-05 09:33:33

0

1c76223b1992

4.978616

2

0

A recommendation engine predict what a user may or may not like among a list of given item. A recommendation system is one of the most…

5

All you know about Recommendation System (Explained) A recommendation engine predict what a user may or may not like among a list of given item. A recommendation system is one of the most common and most successful practical examples for applying a machine learning algorithm in real life. Also they are playing a significant role in people’s daily lives. There are also interests-based social networking site which provides users with recommendations of books, music CDs, movies, and articles, and recommendations of people who might share the similar tastes based on users’ ratings for the mentioned items. This makes recommendation engines a great part of web sites and services such as Facebook, YouTube, Amazon, and more. Recommendation engines work ideally in one of two ways. It can rely on the properties of the items (products, movies, events, articles) that a user (customers, visitors, app users, readers) likes, which are analysed to determine what else the user may like; or, it can rely on the likes and dislikes of other users, which the recommendation engine then uses to compute a similarity index between users and recommend items to them accordingly. It is also possible to combine both these methods to build a much more robust recommendation engine. A variety of techniques have been proposed for performing recommendation, including content-based, collaborative, knowledge-based and other techniques here we will discuss all of them. Types of Recommendation System 1. Knowledge based recommendation systems Knowledge-based recommendation systems are a specific type of recommender systems that are based on knowledge about the item assortment, user preferences, and recommendation criteria (i.e., which item should be recommended in which context). These systems are applied in scenarios where alternative approaches such as collaborative filtering and content-based filtering cannot be applied. The system gathers the user’s requirements on a particular product by questioning the user and consulting its knowledge base to find the items that best meet the user’s requirement. This system uses Natural language processing (NLP) tools and understand the requirement of user and recommend similar products. This type of system have knowledge of the domain where it is used. So it is also called as Domain or Conversational based recommendation system. It is also used in Chatbot were user is ask about best restaurant and hotels and it recommends best restaurant or hotel which meets user may like. Example : When you go to shop to buy dress you tell your requirement and the shopkeeper recommends dresses which meets your requirement. 2. Demographic: This type of system uses information about the user i.e age, sex and location of the user . And recommends the item which is liked by similar user to your profile. The assumption is that different recommendations should be generated for different demographic niches. Many Web sites adopt simple and effective personalization solutions based on demographics. Fashion websites show female products when user is female and shows sports item is the user is male. 3. Most Popular Items — The Simplest Strategy The simplest strategy is to simply offer the customer whatever is most popular, be that a movie, a book, or an article of clothing. Without doing anything more than looking in your sales records you could accomplish this. No data science required. Such approach also works in places like news portals where people likes to read the popular news article of there interest. 4. Association or Market Basket Analysis Association Analysis and Market Basket Analysis looks almost exclusively at content. This type of statistical analysis relies on only the simplest of calculations to find items that are frequently consumed together. Association and Market Basket analysis are mathematically the same. When customers typically acquire the items or services one at a time (like banking services) we call this Association. When customers potentially buy several things at once we call this Market Basket. So Association Analysis is conducted at the customer level (what’s in their account) while Market Basket Analysis is conducted at the transaction level (what’s in their basket). Example: If customer buys Banana, she may also buy milk. 5. Content-based: The system learns to recommend items that are similar to the ones that the user liked in the past. The similarity of items is calculated based on the features associated with the compared items. Such systems are recommending items similar to those a given user has liked in the past, regardless of the preferences of other users. Basically, there are two different types of feedback. Explicit feedback is intentionally provided by users in form of clicking the “like”/”dislike” buttons, rating an item by number of stars, etc. In many cases, it is hard to obtain explicit feedback data, simply because the users are not willing to provide it. Instead of clicking “dislike” for an item which the user does not consider interesting, he/she will rather leave the web page or switch to another TV channel. Implicit feedback data, such as “user viewed an item”, “user finished reading the article” or “user ordered a product”, however, are often much easier to collect and can also help us to compute good recommendations. Various types of implicit feedback may include: Interactions (implicit feedback): - user viewed an item - user viewed item's details - user added an item to cart - user purchased an item - user have read an article up to the end Again, you can expect better performance of recommender system, when the feedback is rich. 6. Collaborative filtering Last group of recommendation algorithms is based on past interactions of the whole user-base. These algorithms are far more accurate than the algorithms described in previous sections, when a “neighborhood” is well defined and the interactions data are clean. User based : Find similar user to me and recommend what they liked. Item based : Find similar items to those that i have previously liked. 7. Hybrid recommendation Systems(Mix of recommender system) All the above systems have there strengths and weaknesses. So we come up with an system which uses strengths of above system and recommends items. Soon I”II update more on this topic. Happy Reading. Any suggestion and questions are welcome.

All you know about Recommendation System (Explained)

12

all-you-know-about-recommendation-system-explained-1c76223b1992

2018-04-05

2018-04-05 09:33:34

https://medium.com/s/story/all-you-know-about-recommendation-system-explained-1c76223b1992

false

1,019

null

Data Science

data-science

Data Science

33,617

Shubham Gupta

Data Science & Machine Learning Enthusiast. Building Neural network and sharing my knowledge to all.

ab6fdc580b9a

shubham02gupta

3

20,181,104

null

0

null

0

2b74c7736d2a

2018-05-06

2018-05-06 15:04:20

2018-05-06

2018-05-06 15:06:18

1

false

en

2018-05-06

2018-05-06 15:17:24

1

1c7640fb4d4e

1.215094

109

2

0

AiFintek has decided to give away 5 million$ worth AFTK tokens for our world wide Crypto support community.

5

AIRDROP Announced — 5 mil $, 50,000 members AiFintek has decided to give away 5 million$ worth AFTK tokens for our world wide Crypto support community. the give away can be claimed here: https://goo.gl/forms/eydgTgCCBazcUSwg1 Why are we doing this? As the development team, we have decided that we want to focus on global community development and have our community be involved at every stage. This is Web 3.0 We have three focus areas — AI Logic development that is for the people and by the people, automated “Heartbeat” Smart Contracts and its platform development to facilitate open P2P transactions, and Game Development with a Social Learning focus. We love China, we love India, we love Japan, we love Europe, we love Africa and we definitely love Latin America and the North — we want to have a global community of democratic participants in our project that help us to keep focused on product development that generates social benefit for all. Our community involvement and AFTK Token giveaway campaign starts on Twitter. We are also extending this on WhatsApp and WeChat platforms. Members — Please forward to Friends and Family on your preferred platforms so that the project can reach scale for development — Lets get connected on this. This is our Guarantee — if we do not raise 5 million $ in our ICO — which is our soft cap, ALL funds received will be returned to the ICO contributors’ public wallet addresses. Also, the initial AIRDROP announced today is totally free. So, Come one, Come all — Claim your share in this emerging new platform. Welcome Aboard !!

AIRDROP Announced — 5 mil $, 50,000 members

954

airdrop-announced-5-mil-50-000-members-1c7640fb4d4e

2018-06-18

2018-06-18 00:14:13

https://medium.com/s/story/airdrop-announced-5-mil-50-000-members-1c7640fb4d4e

false

269

AI Logic Trading, “Heartbeat” Smart Contracts and Game development

null

AiFinTek

tek@AiFinTek.com

aifintek

BLOCKCHAIN,ARTIFICIAL INTELLIGENCE,GAME DEVELOPMENT,FINTECH,ICO

AiFinTek

Blockchain

blockchain

Blockchain

265,164

Oscar Wilde

null

28a3e51b4655

oscarwilde2100

92

1

20,181,104

null

0

null

0

null

2018-08-07

2018-08-07 05:45:57

2018-08-07

2018-08-07 05:46:38

0

false

en

2018-08-07

2018-08-07 05:46:38

1

1c766ebd1327

0.335849

0

Big data is a revolution led to the explosion in data science. Machine learning is helping with prophesying behavior and spotting patterns…

2

Top 6 advance revolutionary things about Machine learning 6 revolutionary things to know about Machine Learning We are stepping into an avant-garde period, powered by advances in robotics, the adoption of smart home appliances…bigdata-madesimple.com Big data is a revolution led to the explosion in data science. Machine learning is helping with prophesying behavior and spotting patterns that humans fail to predict. List of 6 revolutionary things like machine learning- Generalization is the core, data alone cannot do the job, feature engineering is the key to success etc.

Top 6 advance revolutionary things about Machine learning

0

top-6-advance-revolutionary-things-about-machine-learning-1c766ebd1327

2018-08-07

2018-08-07 05:46:38

https://medium.com/s/story/top-6-advance-revolutionary-things-about-machine-learning-1c766ebd1327

false

89

null

Machine Learning

machine-learning

Machine Learning

51,320

NexSoftSys

Technology Consulting Firm for Customized #Offshore #Software & Mobile #Apps #Development for Healthcare, Telecommunication and Banking System.

e2bc0f6834bf

nexsoftsys

30

229

20,181,104

null

0

null

0

null

2017-10-22

2017-10-22 12:38:44

2017-10-22

2017-10-22 18:12:23

7

false

en

2017-10-22

2017-10-22 18:27:13

3

1c798e70ee08

4.740566

41

3

0

Hi there, this is a continuation from my previous blog post about SWISH activation function recently published by a team at Google. If you…

5

Swish in depth: A comparison of Swish & ReLU on CIFAR-10 Hi there, this is a continuation from my previous blog post about SWISH activation function recently published by a team at Google. If you are not familiar with the Swish activation (mathematically, f(x)=x*sigmoid(x)), please be sure to check out the paper for an in-depth understanding or my blog post for a TLDR. According to the paper, Swish often performs better than ReLUs. But many people have pointed out that it is also more computationally intensive than ReLUs. In my previous post I showed how Swish performs relative to ReLU and sigmoids on a 2 hidden layer neural network trained on MNIST. Although a simple 2 layer network is a good starting point, one cannot really generalize the results to most problems one might encounter in practical use cases. In this post, I will compare the performance of Swish on Convolutional Networks (CNN) and show you exactly how slow is Swish relative to ReLUs. Lets get started. BYOC (Build Your Own CNN) Since I was going to make my code open source, I decided to write it such that someone with more computational resources can just change one line of code and train as deep a model as they like. This is the idea behind BYOC. To make this possible, I used a ResNet like architecture with its famous computational bottleneck units consisting of 3 convolutional layers of 1x1, 3x3 and 1x1 convolutions, as shown below: Structure of the computational bottleneck (source: ResNet paper) With this idea, I designed my model such that adding convolutional layers is as simple as setting a variable n_layers to some number. Hence the acronym BYOC. Feel free to check out my GitHub repo and train a better model. For the purposes of this blog, I trained 2 models, a 6 layer and a 12 layer one. These are end to end convolutional networks and there is no fully connected layer except at the output. Results: To recap, I am comparing three activations: ReLU, standard swish and swish_beta (f(x)=x*sigmoid(beta*x), where beta is learned during training). Since the objective was to compare activation functions and not to build a great model, I did not tune the hyperparameters and trained for only 10 epochs. The results I got do not paint a rosy picture for Swish activation. Here are the results for training accuracy from the six layer network: Performance Its clear to see that ReLU performs quite well here, much better than Swish. The variables for all networks were initialized starting from the same random seed. So, initialization is not a factor here. I really wanted Swish to do better especially considering the fact that it is more computationally involved. A comparison of training time for these activations is shown below: Training time To be clear, I have compared the time of forward and backward passes through the whole network, as opposed to just the time of applying the activations, since one won’t use the activations in isolation. On my AWS g2.2xlarge AMI, with a batch size of 128 on an average ReLU took 200 milliseconds to make one full pass, Swish took 11.2% more time, and swish_beta took 12% more time than ReLU. Inference time For practical applications, it is more important to know the inference time of any network since when a network is deployed on a product, we only want to do inference. Here again, as expected, Swish is slower. With a batch size of 100 samples, on an average, ReLU took 44 milliseconds, whereas Swish took ~21% more time and swish_beta took ~28% more time. 12 layer Network: The results from 12 layer network are similar. ReLU has higher accuracy, and is much faster during training and inference. The results are shown below: Accuracy Train time: Inference time: Interestingly, I the gap between ReLU and Swish for training and inference has increased as the network has gotten deeper. This is not a great sign. Since even a ‘light weight’ ResNet is also about 50 layers deep, it would be just terrible if the gap between training and inference time keeps on increasing with the number of layers. But one cannot really extrapolate from just two data points (6 & 12 layers). I wanted to train an 18 layer network as well, but the results from these two cases made me less motivated to do so. I really wanted Swish to work better than ReLUs but, at least in these experiments, it didn’t. I would be happy to have been proven wrong as more people apply it to real world problems, because I, like everyone else, want to be able to train better models. Takeaways: Here are a few takeaways from these results: Swish does not really perform as well as in these experiments as I expected. ReLUs consistently beat Swish on accuracy. There is a large difference between training times required for these activations. Even if Swish performs better than ReLUs on a problem, the time required to train a good model will be about 15–20% more than ReLUs. More importantly, the run time performance of Swish seems to be much slower than ReLUs, by 20–30% or more. This is a non trivial slowdown for cases where real time performance is needed. Between standard Swish and Swish_beta, the beta version certainly performs better. If you find that Swish works for your problem and if you care ddeply about the accuracy, it is a good idea to try learning the parameter beta during training as well. Of course, the higher accuracy would come at the cost of at higher training time and inference time, as we saw from the graphs above. Code All the code for reproducing these results and training more models with BYOC is uploaded on my GitHub. If you find any bugs or have difficulty in understanding the code, feel free to contact me.

Swish in depth: A comparison of Swish & ReLU on CIFAR-10

185

swish-in-depth-a-comparison-of-swish-relu-on-cifar-10-1c798e70ee08

2018-06-10

2018-06-10 06:36:22

https://medium.com/s/story/swish-in-depth-a-comparison-of-swish-relu-on-cifar-10-1c798e70ee08

false

978

null

Machine Learning

machine-learning

Machine Learning

51,320

Jaiyam Sharma

Blogs about replicating research papers in machine learning

5917295ec799

jaiyamsharma

171

82

20,181,104

null

0

null

0

null

2018-08-20

2018-08-20 10:29:07

2018-08-20

2018-08-20 10:31:19

3

false

en

2018-08-20

2018-08-20 10:31:19

5

1c79c2d6182a

3.217925

1

0

Technological progress does not stand still, the revolution in industry has already occurred, This has contributed to the introduction of…

4

Improve Your Business with BLACKBOX Technological progress does not stand still, the revolution in industry has already occurred, This has contributed to the introduction of high technology in industrial production. At the enterprises, the tendency to automation of all working process, by means of introduction of an artificial intellect is pursued. Naturally, for business, such innovations are economically viable. The less manpower, the more their revenue figures. This working process in the future will allow many times to improve the economy of developed countries. The overwhelming majority of people are not particularly satisfied with the situation when they are a simple mechanism for obtaining benefits. They get the idea that their work is undervalued and requires more reward. Blackbox OS is a unique block-based platform that allows you to create a decentralized work environment through a global business operating system to manage the automation of the entire workflow. This helps improve the workflow. A platform is an ideal option for large organizations that seek to become advanced and keep pace with the times. The usual methods of documenting on paper or electronic media are transformed into a system of smart contracts. The introduction of blockade technology and artificial intelligence will give huge opportunities for companies to conduct their business. Platform The decentralized platform facilitates the formation of an independent ecosystem for collective work with common management. Blackbox will use the BBOS token to manage and coordinate all platform processes in a decentralized environment. Important characteristics of Blackbox can be divided into 5 components: Control Any enterprise has its own management system. Thanks to Blackbox OS, the existing problems in the work of the whole team can be eliminated. Blackbox will conduct an analysis of the management system of a particular organization and provide an appropriate model for improving the process of achieving the objectives. Identity The platform will provide each employee of the enterprise with the creation of a profile in which his personal data will be stored, as well as information on the reputation obtained while working. Security Of course, protection from hacker attacks will be one of the most important values for the enterprise. Blocker will provide full protection of all company data and a variety of business tools to reduce the risk of third-party interference in the workflow. Compensation It’s no secret that a worthy reward gives the employee the incentive to perform their work more qualitatively. The personal contribution to the working process of each worker should be adequately evaluated. For these purposes, the platform provides a protocol Proo of Value. To increase the speed of transactions, reduce fees, employee compensation, Blackbox will use its own BBOS token. Management Naturally, any company has its own specific workflow. The platform does not set itself the goal of forced optimization of the entire work process. Blackbox provides the enterprise with new solutions to improve the efficiency of the workflow. Blackbox Platform Operating Modules To solve specific tasks related to the optimization of the workflow on the platform, modules are created that support artificial intelligence. The main task of these modules is to integrate into the existing enterprise work system and improve all its parameters. Module DVP. This module contributes to the improvement of the solution of the current tasks of the enterprise and the management of its projects. It combines and improves the interaction between the best tasks and management functions. Market Module This module will work as a centralized solution. Modules can integrate directly into the operating system of the company. It is planned that the new modules will be added by the developers. HR module, AI module and detection module — all of them will be available. Decision Definitely in the future, the entire workflow will be automated and the involvement of a person will be minimized. Already now it is possible to trace the tendency when many professions are simply not needed. This every year occurs more often, due to the gradual introduction of high technology. The developers set themselves the goal of automating low-skilled jobs, as well as improving the efficiency of the workflow through decent remuneration and incentives for employees. WebSite: https://token.blackboxfoundation.org/ White Paper: https://docsend.com/view/zn2axya Telegram: https://t.me/blackboxtoken ANN thread: https://bitcointalk.org/index.php?topic=4517962.0 My Bitcointalk Profile: https://bitcointalk.org/index.php?action=profile;u=1712011 My Eth: 0xce74878c0c0E431b9615d5Bc0A79B931F68293F8

Improve Your Business with BLACKBOX

41

improve-your-business-with-blackbox-1c79c2d6182a

2018-08-20

2018-08-20 10:31:19

https://medium.com/s/story/improve-your-business-with-blackbox-1c79c2d6182a

false

707

null

Blockchain

blockchain

Blockchain

265,164

icologist

null

ad67d8b36cd9

icologist

293

453

20,181,104

null

0

null

0

null

2017-11-26

2017-11-26 10:27:00

2017-11-26

2017-11-26 10:29:28

0

false

en

2017-11-27

2017-11-27 05:21:48

9

1c7a34bcafe

3.415094

1

0

By Malisa Nusrat Huda

5

Why do we need to collect data? By Malisa Nusrat Huda You don’t need to be Amazon to know what your customers will demand and what to recommend. You just have to start small and simple and keep it going. Not only will leveraging customer data support sales every day of the year, but your customer database will be a valuable asset when you sell your business. Similarly, all data that you collect will in turn become an asset that gives you a competitive edge when used effectively. Big data and data collection and data analytics Big data describes voluminous amounts of structured, semi-structured and unstructured data collected by organizations. But because it takes a lot of time and money to load big data into a traditional relational database for analysis, new approaches for collecting and analyzing data have emerged. To gather and then mine big data for information, raw data with extended metadata is aggregated in a data lake. From there, machine learning and artificial intelligence programs use complex algorithms to look for repeatable patterns. Data analytics is the science of collecting and examining raw data with the purpose of drawing conclusions about the information. These analyses are used by organizations to make better business decisions, especially where change is the outcome. What’s the purpose of collecting so much data? 1. To support the need for a change in operations or process. People often fear change, especially without supporting data to validate the need for it. Many times we see a need for a change in a process within our organizations but approaching policy-makers with your opinion or gut instinct will not hold much ground without concrete data to support it. In today’s metrics-driven world, data is required in order to institute change or justify a business case for it. 2. To gain visibility to the unknown. Industry professionals are smart! Our intuition often guides us when making key decisions, however data is useful in proving our instincts right. Actionable insights add clarity to the unknown which will support or challenge a theory, both of which are invaluable in decision-making. 3. To reduce risk. Often, people tasked with planning meetings and events in an organization are not meeting professionals, but rather meeting planning is a part of their job responsibilities. As such, they are not experts in understanding and negotiating the terms of a meeting contract and sign agreements that may not support the best interests of their company. This practice exposes both the individual signing the agreement as well as the organization to risk should an issue arise with the contract, or in the event the meeting cancels and penalties are assessed. There are also terms that, if not properly addressed or omitted, could jeopardize the safety or well-being of attendees of that meeting. Having historical bid data or analytics that identify suppliers with whom it’s safe to do business, or to support professional procurement practices, can reduce or eliminate this risk all together. 4. To demonstrate compliance with regulations. No one wants to be called out for non-compliance. Whether it’s following company policies, regulatory guidelines, accounting principles, or laws, being compliant is key to job security. It’s not enough to have regimented processes for transparency, government entities want to see that the data is reported in such a way that it meets the letter of the law(s). 5. For Executive Management support of a new idea. Executive management, especially in publicly held companies, are being tasked with ensuring that all the decisions they make support quarterly earnings. It’s fun to be the initiator of a new idea with the potential to dramatically streamline a process, improve the bottom line and increase those earnings. Suggesting new concepts may be admirable, but they won’t go very far without analytics to support them. Often predictive analytics are necessary to project how that idea will benefit the organization going forward. Leverage data to build a business case for specific elements that will yield the highest ROI or provide other tangible benefits to the organizations. 6. To build stronger partnerships with key suppliers. Knowing what you spend and with whom is the first step in establishing preferred or strategic supplier partnerships. These relationships work best when they provide a win-win for both parties. Suppliers are looking for increased market share in exchange for discounts and other financial incentives. A better approach is to align with partners with whom your organization naturally gravitates towards, whether because they are geographically well-positioned, and/or they offer financial benefits already that may be improved upon as market share increases. Regardless of the approach in selecting key suppliers with whom to partner, having the data that supports who to approach in the first place is a critical success factor. “The collection of data is not an end in itself, but is essential for informed decision-making.” In today’s economy, the value data might bring into business is limitless. It’s well-known that it is a lot more effective to sell to the customers you already have than to find new ones. But if you don’t know who they are, how to reach them or what they have already bought and like, you will just have to wait. — and hope. And that’s no way to grow a business. That’s when your database based on all information you have collected over the years comes into action.

Why do we need to collect data?

1

why-do-we-need-to-collect-data-1c7a34bcafe

2018-04-23

2018-04-23 11:04:58

https://medium.com/s/story/why-do-we-need-to-collect-data-1c7a34bcafe

false

905

null

Big Data

big-data

Big Data

24,602

Cramstack

Cramstack provides a platform that makes data access simple by searching through millions of rows of company data by asking questions in plain English.

1ce64b5d41

cramstackdata

15

4

20,181,104

null

0

null

0

null

2018-02-16

2018-02-16 00:54:44

2018-02-20

2018-02-20 08:45:17

4

false

en

2018-02-20

2018-02-20 09:09:26

2

1c7c42cbb160

8.933962

1

0

Machine learning is one of the strategies many companies are using as they try to improve and differentiate their product by augmenting the…

5

UX Design for AI: AI specific issues while user testing Machine learning is one of the strategies many companies are using as they try to improve and differentiate their product by augmenting the user’s experience. Building artificial intelligence into apps can improve what can be done for customers and users, but can also affect how customers and users interact with the machine. In this article I am covering how, as the designer, specific issues for you to look out for when testing user scenarios, preparing for doing user testing, and doing user testing. I’ve previously covered how choosing an algorithm affects the user’s experience where I cover the research that is needed before design begins. The information from the previous article will help when doing user testing too. Photo courtesy of www.amenclinics.com PROTOTYPE TESTING & VERIFICATION Even before doing user testing you want to make sure all of the algorithms are working. If you can hand off UX requirements to the QA testing group it’s helpful. But, I’ve found a lot of the time the QA group is also adjusting to the new reality of working with AI; and a lot of the time, as long as they can verify that the app gives a reasonable response, it passes. However, AI can give many responses that make sense but do not help the user to achieve their goals. So is the AI helping? I’ve found the easiest way to find out is an A/B test of AI vs no AI. The problem is that once software has been written it is hard to disable the AI and have the software still work. A good way around this is to have the developers set up an alternative algorithm to just take all the data and find the mean average. If the AI can not preform better to achieve the stated goals than averaged data then back to the drawing board before user testing. USER TESTING As soon as there is anything usable for the AI it is best to test with the users to verify what it is giving what is important to them. Even after choosing a specific algorithm developers can modify the algorithm using human imposed variables called Hyper-parameters. These settings can tweak how the machine learning algorithm preforms. It can be things like how many layers of neurons, or the number of neurons in a layer. The important thing to remember is that the more accuracy achieved is usually associated with higher processor requirements. So what problems can you discover? Let’s start with an example that during user testing you all the users getting too similar of recommendations from the AI or the AI is always making guesses that are too similar to each other. This could be caused by Underfitting. The best way to explain underfitting is with an X-Y graph. In this example we are only using two lists of measurements/data called Features. Of course what you are working on will have more features, but the idea is the same. Figure 1 shows the hypothetical AI giving a perfect recommendation line ignoring all noise. Figure 1: Best Case Scenario Underfitting is ignoring or generalizing too much data during training. It is treating everything as noise. Figure 2: Underfitting The solution is to train on more data (or data closer to what the users are using), add more features (measure more things), create a more more complex model (add more neurons or layers to the model). Developers will have to know which hyperparameters make the most sense to change in your situation; and there is a good chance they will change all of them to see which ones create the best results. Your part is knowing this can be a problem, so you can tell them where 1) the data needs to be sampled, 2) and where the answers needs to be right. The opposite problem is Overfitting. This is trying to fit every point on the graph. The algorithm will work great when all the points are part of the training data but the AI can’t generalize when it sees something new. This can happen when no Cross-validation is done (Splitting out some data to test the AI on instead of using it all for training.) Sometimes cross-validation can’t be done, but sometimes it is forgotten on the list of things to do. Figure 3: Overfitting Overfitting shows up during user testing as the edge cases being wildly off, so there is actually a better chance of catching it when validating the prototype. Of course there are the basic remedies of removing layers or neurons but another solution is something called regularization. Regularization is a fix for the problem machine learning has if some of the features are small numbers and some are large. The algorithm will tend to forget the small numbers. For example, in a home buying AI with the features as number of rooms and price, since the price is much higher, the number of rooms won’t matter. The fix is to make all the data sets look similar. For example make the price of the homes in 100k (eg. $2.1 instead of $210,000). For natural language processing regularization will strip off all of the verb endings so working, worked, and work all count as the same word. Usually regularization works without your intervention unless the thing the user is interested in relates to something that was regularized away. While this can pop up in user testing it is easier to catch by talking with the developers to see what data they have regularized and make sure it does not conflict with the users goals. DETECTING BIAS Bias is a big enough problem it needs its own section. As smart as machine learning gets, when a bias is detected it can remind you just how dumb AI still is. First, I will cover Statistical bias. This is when the “bias” is part of the error term. Basically the model is not being the true model. So if the model is off this is something that should be designed out, hopefully with your help the developers never see this problem. Using the user stories and goals take into account the variance and bias from reality. Variance is how far from the average numbers can get. Think of a shotgun pattern on a wall. The further back you stand when shooting the gun the more the dots will be spread out, or the higher the variance. The tolerance for a high or low variance depends on what the user goals are. Also know that variance interacts with bias. If variance is how spread out the answers are, bias is how far off the center of the target is. So based on what the user goals are (like consistency is more important than accuracy), you might want a high bias and a low variance. This would make all the answers close together with a predictable distance from reality. The second type of bias is biased training data. Biased training data is when there is a problem with how the data is gathered. Discovering the AI being effected by training data was first detected in 1964 ( with what might be apocryphal) when they were trying to use an AI to detect tanks in images. Testing off the training data went well but it could not work with new photos. It turns out all of the tank photos were from sunny days and non-tank photos on cloudy days. So the AI was good at detecting the weather in the photo instead of finding tanks. Another example is using an AI to decide who to keep in the hospital with extra care and who to send home. The AI kept recommending people with asthma to be sent home which went against obvious medical knowledge. It turns out people with asthma were getting extra attention from the doctors because of the extra risk. this extra attention was not factored into the training data and therefore leading the AI to the bias not matching reality. Detecting biased training data is something you will need to look for on both ends. During the design phase sit down with the developers to make sure the training data parallels the user goals, actions, and stories. Also make sure the training and testing data are split randomly. For example you don’t want stock market data to be trained with all the data from last year and tested in the data from the last two weeks. Once the algorithm is ready make sure you are covering all of your persona types when doing your user testing. Also, this is a good place to find the subject matter experts (SMEs) and go through heuristic evaluations with them. Like the doctors in the previous example they can say when recommendations do not match up with reality. A subcategory of biased training data is data normalization. As you know, users will put in fake data if there is forced collection to reach their goal. A good example is known as the Schenectady problem. There is a zip code 12345. It is for the GE factory in Schenectady, New York. If you try to use unnormalized data the amount of people showing up as living inside of a factory will be unusually high. Other anomalies include the number of people sharing the birthday January 1, 1900 and phone numbers that start with 555. It is easier to design in catch questions for user surveys. But, if you are working with already collected data the algorithms used to normalize data differs on what you are working on so the main things to verify is 1) the data is normalized before AI training and 2) that if your user groups and personas do have “out of the norm” peculiarities that they are not normalized away out of the data. The last type of bias to cover is social bias. This is when you know the data goes against the company values. All data collected is from the past and from people who are acting on what they learned in the past. So things like racism and bigotry will show up in real data sets and need to be adjusted for. The biggest problem is remembering there is a problem since most development is done within one bubble or other. As the designer it is something to check for: 1) during design, 2) verifying in the training and testing data, 3) making sure the personas cover race, culture, personal and group identity, and gender and 4) to test against those persona groups. Social bias gets split up into two groups (and I’m quoting directly from Kate Crawford, NIPS 2018) Harms of allocation covers discrimination in the product or service, for example approving a mortgage, granting a parole, or deciding insurance rates. The second is Harms of representation. This covers the social inequalities and stereotypes we don’t want to perpetuate. There are plenty of examples of companies getting embarrassed by this. With Google, being a leader in machine learning, run into this problem a lot. There was the time their image recognition app was recognizing black people and tagging them as gorillas. The problem was there was not enough racial diversity in the developers working on the project so when they were testing with their own pictures they never saw the problem (causing biased training data). Google also had a problem with their search recommendations. Since they build up their recommendation lists based on what other people typed in; when searching, racist recommendations used to pop-up when minorities were words used. They fixed this by being aware of the problem, detecting the racist searches, moving them to a different list than the training data so when racist data shows up as part of the test data it can still be acknowledged but will not count against the accuracy of the algorithm and will not show up as a suggestion. Other examples include the profiling algorithm being less likely to recommend high paying jobs to women (story by Prachi Patel) and searching for historically black names are more likely to show ads for prison background checks (research by L. Sweeney). Or, when doing language translation, translating from non gendered sentences for doctor and nurse will add male to doctor and female to nurse. Even the tools used by developers like word2vec (a tool to categorize words to other words they are most likely to be used with)is more likely to associate male associated words with brilliant, genius. commit, and firepower, while female associated words are near babe, sassy, sewing, and homemaker. CONCLUSION This is just scratching the surface of problems you can run into. There are so many different areas of AI I can only cover some of them. Not to mention everyone is working on a customized version of an algorithm to get it to work for their own specific needs. If you come across a design problem or solution with AI, you are free to contact me and let me know so I can help or get to word out about good solutions.

UX Design for AI: AI specific issues while user testing

11

design-for-ai-ai-specific-issues-while-user-testing-1c7c42cbb160

2018-05-28

2018-05-28 13:34:35

https://medium.com/s/story/design-for-ai-ai-specific-issues-while-user-testing-1c7c42cbb160

false

2,182

null

Machine Learning

machine-learning

Machine Learning

51,320

Mark Bailey

UX Designer specializing in AI and machine learning apps. Portfolio at http://baileydesign.com

bd366f904033

baileydesign

14

74

20,181,104

null

0

null

0

null

2018-10-31

2018-10-31 15:04:32

2018-07-24

2018-07-24 20:45:42

1

false

en

2018-10-31

2018-10-31 15:09:33

7

1c7cc02dadf4

5.932075

0

The difference between cloud software design and traditional architectures is that cloud software design takes advantage of principles…

5

5 Steps to Take Before Migrating Your Apps to Azure The difference between cloud software design and traditional architectures is that cloud software design takes advantage of principles such as decentralization, self-healing, asynchronous, elasticity, and polyglot persistence. As a result, before migrating existing applications, some housekeeping is in order. Below are 5 steps to take before migrating your apps to Azure. 1. Security You should already be performing penetration tests on any apps exposed to the internet. You should also be doing source code vulnerability scans with your favorite tool and taking steps to remediate any potential problems. Other typical practices include endpoint security, data encryption, connection string vaulting, logging, and securing any access secrets. These operations should be a part of your current development process and are major steps toward ensuring cloud application security. If you feel there are gaps in your security posture, then get an independent assessment. Once you’ve done all you can on the ground, it’s now time to turn your attention to cloud security. Azure’s Assumed Breach Strategy is where you’ll start. Understanding that it’s not if, but when a breach will occur, helps to prepare you organization for the inevitable attack. Furthermore, on the application side, Azure’s Cloud App Security Framework will help ensure you’re protected (see below). Cloud Discovery — Discover all cloud use in your organization and continually assess risk Data Protection — Monitor and control your data, enforce policies, ensure privacy, and be informed Threat Protection — Use artificial intelligence and machine learning to detect anomalies and suspicious behavior patterns in your traffic 2. Get your code off the ground The goal is to eliminate the clutter, so you can focus on the business of software, including DevOps infrastructure. Stop using TFS and build servers and move all source code repos to Visual Studio Team Services (VSTS) and Git. Take the opportunity to define or refine your pipelines and delivery processes. Automate everything including tests, builds, deployments, notifications, defect entry, and task promotion. Remove any human interaction with these processes and you’ll decrease errors, bugs, and defects dramatically. TFS server hardware maintenance…gone. TFS server patching…gone. TFS version upgrades…gone. Punching network holes for remote developers…gone. Global access…enabled. Seamless and secure Azure integration…enabled. Let’s stop treating our servers like household pets and more like cattle. Put those TFS servers to pasture and let someone else handle walking the dog. 3. App Dependency Mapping You could not have picked a better time to move your apps to Azure. There are few reasons left for not moving to the cloud. The advantages far outweigh the disadvantages, and your CEO is probably wondering why you’re not already there. But let’s tap the brakes a bit and continue cloud preparations through documentation and diagrams. In-house apps, developed over time, will have forgotten complexities and integration points that if not considered before cloud migration, could stifle success. However, in most cases simple consideration and remediation will typically thwart any issues. This starts with knowing and depicting app dependencies. Knowledge that your apps use Windows GAC, are wired into antiquated vendor services, are dependent on data movement processes, require special hardware integrations, or use 3rdparty libraries that are no longer supported, is key to a successful and timely migration. These items are easily handled by various Azure components, but planning for them and expecting them is more effective than being reactive. There are multiple tools to help with this endeavor; however, you’ll still want to examine and debug the code, interview developers, and meet with app users as this discovery process will help complete the picture. The final output should include app documentation and workflow diagrams that depict application dependencies. 4. Understand Azure Services Organizations are looking for ways to innovate, reduce IT spend, and improve performance and they are banking on cloud technologies to help get them there. Understanding Azure cloud offerings will aid in all the aforementioned categories. Now you’ll have the ability to match right-sized and right-fit cloud components to your in-place applications. We are not interested in jamming the shiny new cloud toy into production, but pairing cloud services to app components that will indeed give you the benefits of leveraging Microsoft’s cloud. Start with Platform-as-a-Service (PaaS) As developers we generally adhere to the philosophy that simple is better and less is more, so let Azure handle the infrastructure while you focus on delivering business value. We do not want to patch servers, upgrade hardware, or worry about scale and whether my app will be resilient in a storm. We want to focus solely on developing great products for our business and customers. This is what Azure PaaS services do for us. Below is a listing of traditional hosted services mapped to the Azure PaaS respective equivalent. This should serve as a starting point. It will take discovery and analysis to ultimately select the appropriate Azure PaaS solution. Hosted Azure PAAS Website Web App, Blob Storage wcf sERVICE Web App, API App, Relays API, Web API API App, Logic App, Function, API Management Files File Storage images Blob Storage, CDN video Media Services messaging, Bus, msmq Queue Storage, Service Bus SQL database Azure SQL, Elastic Pools nosql Table Storage, Cosmos DB Document DB Cosmos DB warehouse Data Warehouse, Data Lake data movement Data Factories MySQL MySQL DB postgresql PostgreSQL DB integrations Logic Apps, Connectors, Relays. NOTE: Azure PaaS services can handle most of the common programming languages and their platforms including: ASP.NET, Node.js, Python, C#, JavaScript, Java, PHP, TypeScript, Ruby, and Scripts Move to Containers Why should I implement containers? What business use cases do containers solve? Below is a high-level list of reasons to use containers. Modernize existing applications Provide software consistency through like environments Abstract network topology issues Abstract the underlying OS and infrastructure They are self-healing They run anywhere and/or in isolation Application decentralization Cons of containerization can include a substantial team learning curve, difficulty of tracking and debugging, maintenance of container images, and not all scenarios fitting the container model. Your CI/CD pipelines and general DevOps posture will also need attention. Again, careful consideration and building a suitable use case will ensure you are using containers wisely. One example is Microservices. You cannot implement microservices properly without containers and container orchestration, but this architecture typically assumes massive scale as a business requirement. Below are the current Azure container offerings. Azure Offering Service Notes Container services Container orchestration Kubernetes, Docker Swam, DC/OS Container Registries Repository of container images Container housing/storage Service fabric clusters Standalone node orchestrator Run anywhere on Windows Server Container instances Simple and scalable PaaS service Deploy single containers quickly and easily Kubernetes services PaaS service Managed Kubernetes orchestration OpenShift Red Hat container platform Container orchestration from Red Hat Docker EE for Azure Enterprise-grade cluster management Turnkey Docker container system in Azure NOTE: As we speak, Azure is going to start offering Windows containers as an App Service. Currently ,App Services (PaaS) containers are only supported using the Linux OS. Finally, If you must… Infrastructure-as-a-Service (IaaS) You can simply “lift & shift” your applications to cloud infrastructure and mirror what on have on premises. This in part is why we can say, “if you can run it on premises, then it will run in the cloud”. Azure has done a marvelous job of making this scenario a reality. The only apps that will not run in any cloud are ones that are poorly architected, highly proprietary, or perform unpredictably. Avoid these apps. We will leave IaaS options, design, and architecture for another post as they are extensive, but take note that most organizations have a mix of IaaS, PaaS, and SaaS offerings to serve up their line of business applications. Choose the best tool for the job and use it wisely. Then optimize, optimize, and optimize again and remember: If it’s already written, don’t re-write it (SaaS) If you can abstract hardware, do so (PaaS) 5. Staff Readiness As developers and technologists, it is our duty to keep pace. It is what we signed up for. The good news for developers is that between Visual Studio, Visual Studio Code, and Azure Cloud Shell we have everything we need for Azure cloud development. Microsoft, in transforming their business, has done a fantastic job of inclusion, open source, and BYOT (Bring your own tools), so many of your daily activities will be familiar. You simply need to focus on Azure and utilizing the educational resources provided by Microsoft. Below is a list of resources to help get you started. Conclusion At Quisitive our mantra is ‘Start Right, Finish Right’. Preparing your applications as well as yourself for cloud migration will ensure success. Don’t go at this alone. Partner with trusted experts like us who can get you there quickly and then teach you how to maintain and optimize your estate. Migrating applications to Azure doesn’t have to be difficult. Iterative planning and careful preparation will make sure you’re on the right path. Learn more about us at https://quisitive.com Originally published at quisitive.com on July 24, 2018.

5 Steps to Take Before Migrating Your Apps to Azure

0

5-steps-to-take-before-migrating-your-apps-to-azure-1c7cc02dadf4

2018-10-31

2018-10-31 15:09:33

https://medium.com/s/story/5-steps-to-take-before-migrating-your-apps-to-azure-1c7cc02dadf4

false

1,519

null

Azure

azure

Azure

3,345

Quisitive

Quisitive is a premier Microsoft solutions provider that helps customers navigate the ever-changing technology climate that their business relies upon.

f8c5d53de45d

BeQuisitive

0

1

20,181,104

null

0

ffmpeg \ -i results/SonicTheHedgehog-Genesis-GreenHillZone.Act1-0001.mp4 \ -i results/SonicTheHedgehog-Genesis-GreenHillZone.Act1-0729.mp4 \ -filter_complex '[0:v]pad=iw*2:ih[int];[int][1:v]overlay=W/2:0[vid]' \ -map [vid] \ -c:v libx264 \ -crf 23 \ -preset veryfast \ output.mp4

1

null

2018-04-22

2018-04-22 05:07:49

2018-04-24

2018-04-24 03:57:40

1

false

en

2018-06-10

2018-06-10 00:12:14

15

1c7d171e5bc8

2.85283

2

1

0

Days 16–18 of the OpenAI Retro Contest

5

Running the PPO baseline and giving up on local evaluation Days 16–18 of the OpenAI Retro Contest Now that I am back in the land of decent internet, I could build my docker image for local evaluation. It looked like it was going to take some time, but the results definitely seemed worth it. Just take a look at the agent’s first run vs the one 700+ iteration into the future: The only issue was that it was taking hours to run on my MacBook’s cpu. I would never be able to do much experimentation if I had to wait for hours or running overnight for each iteration and I was kind of worried my computer would catch fire from all the hard work. So I set out to optimize. My previous attempt to get TensorFlow compiled optimally for my computer didn’t go well, but that was also because of unrelated confusion from not having a GPU. After searching the web a bit, I tried experimented with this dockerfile from Google: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/docker/Dockerfile.devel At first I tried just adding the parts that I needed, since it didn’t seem like I needed jupyter, but I quickly reverted to seeing if I could get it working as-is. I ran into this bug, but was able to continue by adding the suggested flags. I wasn’t clear about what was already included with the retro contest base image, so my image ended up looking like this: And that seemed to work, despite being very non-optimal and redundant, or at least I didn’t run into any errors that exited and a a few warnings during compilation is normal right? Well a few hours after that I realized that I was getting the same warnings over and over again. Somehow it seems like I was able to get stuck in an infinite loop during compilation. I have my output log here if anyone wants to tell me what is wrong. I wasn’t able to find much help online about optimizing my dockerized ubuntu compiled TensorFlow to take advantage of a couple extra instructions my CPU had. Either this is something that is not often done, or I had no idea what the right search terms were. Ultimately I decided the time spent on this fest like trying to crawl most efficiently when I should be trying to learn to walk. Even with if I get my laptop to work 3x better, that is still hours of hot high cpu time that I didn’t want. I decided that my future investment should be in getting a remote host that has gpu up and running with my agents. Submission time! Team Bobcats hadn’t submitted an agent in a while. So I built the GPU version of the PPO2 baseline and sent it off for evaluation. It scored a paltry 3280 which put us in 56th place. Looking at the video showed some room for improvement as Sonic mostly ran into a wall. Bonus! How to create split screen videos with ffmpeg. If you liked my video at the top, here is how you can make your own. First I used the playback tooling from Day 6 to convert the .bk2 files generated from my agent into .mp4 videos. Then I used ffmpeg to and the nice people on who answer questions in various StackExchanges to combine the two videos into one that was side-by-side: Thanks for reading! You might be interested in the rest of this series: Day 1: Getting the Basics Set Up Day 3: Running the Jerk Agent Days 4 & 5: Getting TensorFlow & Docker to work on my MacBook Day 6: Playback Tooling for .bk2 files Days 9 &10: Failing with the Rainbow DQN baseline code. Days 11–14: Reading the PPO2 code Days 16–18: Running the PPO2 baseline code, and failing at TensorFlow & Docker optimization. Days 22–25: A Deep Dive into the Jerk Agent Days 26–29: Visualizing batches of sonic runs Days 38–53: Discovering Q-Learning My final submission: the improved JERK agent

Running the PPO baseline and giving up on local evaluation

3

running-the-ppo-baseline-and-giving-up-on-local-evaluation-1c7d171e5bc8

2018-06-10

2018-06-10 00:12:15

https://medium.com/s/story/running-the-ppo-baseline-and-giving-up-on-local-evaluation-1c7d171e5bc8

false

703

null

Docker

docker

Docker

13,343

Tristan Sokol

Developer Evangelist for Square. When I’m not helping build a commerce platform, I’m growing succulents in my back yard. https://tristansokol.com/

fb504b57e2ab

tristansokol

245

0

20,181,104

null

0

null

0

null

2017-10-05

2017-10-05 18:29:50

2017-10-05

2017-10-05 18:48:07

1

false

en

2018-02-14

2018-02-14 02:52:28

22

1c7d3d42c9b9

6.437736

23

1

0

In Memoriam: Professor Hubert Dreyfus (1929–2017)

5

The World Is its Own Model or Why Hubert Dreyfus Is Still Right About AI In Memoriam: Professor Hubert Dreyfus (1929–2017) Abstract In this article, I argue that mainstream artificial intelligence is about to enter a new AI winter because, in spite of claims to the contrary, they are still using a representational approach to intelligence, aka symbolic AI or GOFAI. This is a criticism that Hubert Dreyfus has been making for half a century to no avail. I further argue that the best way to get rid of the representationalist baggage is to abandon the observer-centric approach to understanding intelligence and adopt a brain-centric approach. On this basis, I conclude that timing is the key to unlocking the secrets of intelligence. The World Is its Own Model Hubert Dreyfus is a Professor of philosophy at the University of California, Berkeley. Dreyfus has been the foremost critic of artificial intelligence research (What Computers Still Can’t Do) since its early days. The AI community hates him for it. Here we are, many decades later, and Dreyfus is still right. Drawing from the work of famed German philosopher, Martin Heidegger and the French existentialist philosopher, Maurice Merleau-Ponty, Dreyfus’s argument has not changed after all those years. Using Heidegger as a starting point, he argues that the brain does not create internal representations of objects in the world. The brain simply learns how to see the world directly, something that Heidegger referred to as presence-at-hand and readiness-to-hand. Dreyfus gave a great example of this in his paper Why Heideggerian AI Failed and how fixing it would require making it more Heideggerian (pdf). He explained how roboticist Rodney Brooks solved the frame problem by moving away from the traditional but slow model-based approach to a non-representational one: The year of my talk, Rodney Brooks, who had moved from Stanford to MIT, published a paper criticizing the GOFAI robots that used representations of the world and problem solving techniques to plan their movements. He reported that, based on the idea that “the best model of the world is the world itself,” he had “developed a different approach in which a mobile robot uses the world itself as its own representation — continually referring to its sensors rather than to an internal world model.” Looking back at the frame problem, he writes: And why could my simulated robot handle it? Because it was using the world as its own model. It never referred to an internal description of the world that would quickly get out of date if anything in the real world moved. Deep Learning’s GOFAI Problem By and large, the mainstream AI community continues to ignore Dreyfus and his favorite philosophers. Indeed, they ignore everyone else including psychologists and neurobiologists who are more than qualified to know a thing or two about intelligence and the brain. AI’s biggest success, deep learning, is just GOFAI redux. A deep neural network is actually a rule-based expert system. AI programmers just found a way (gradient descent, fast computers and lots of labeled or pre-categorized data) to create the rules automatically. The rules are in the form, if A then B, where A is a pattern and B a label or symbol representing a category. The problem with expert systems is that they are brittle. Presented with a situation for which there is no rule, they fail catastrophically. This is what happened back in May to one of Tesla Motors’s cars while on autopilot. The neural network failed to recognize a situation and caused a fatal accident. This is not to say that deep neural nets are bad per se. They are excellent in controlled environments, such as the factory floor, where all possible conditions are known in advance and humans are kept at a safe distance. But letting them loose in the real world is asking for trouble. As I explain below, the AI community will never solve these problems until they abandon their GOFAI roots and their love affair with representations. The Powerful Illusion of Representations The hardest thing for AI experts to grasp is that the brain does not model the world. They have all sorts of arguments to justify their claim that the brain creates representations of objects in the world. They point out that MRI scans can pinpoint areas in the brain that light up when a subject is thinking about a word or a specific object. They argue that imagination and dreams are proof that the brain creates representations. These are powerful arguments and, in hindsight, one cannot fault the AI community too much for believing in the illusion of representations. But then again, it is not as if knowledgeable thinkers, such as Hubert Dreyfus, have not pointed out the fallacy of their approach. Unfortunately, mainstream AI is allergic to criticism. Why the Brain Does Not Model the World There are many reasons. I’ll just list a few as follows. The brain has to continually sense the world in real time in order to interact with it. The perceptions only last a short time and are mostly forgotten afterwards. If the brain had a stored (long-term) model of the world, it would only need to update the model occasionally. There are not enough neurons in the brain to store a model of the world. Besides, the brain’s neurons are too slow to engage in any complex computations that an internal model would require. It takes the brain a long time (years) to build a universal sensory framework that can instantly perceive an arbitrary pattern. However, when presented with a new pattern (which is almost all the time since we rarely see the same exact thing more than once), the cortex instantly accommodates existing memory structures to see the new pattern. No new structures are learned. A neural network, by contrast, must be trained with many samples of the new pattern. It follows that the brain does not learn to create models of objects in the world. Rather it learns how to sense the world by figuring out how the world works. The brain should be understood as a complex sensory organ. Saying that the brain models the world is like saying that a sensor models what it senses. The brain builds a huge collection of specialized sensors that sense all sorts of phenomena in the world. The sensors are organized hierarchically. They are just sensors (detectors) that respond directly to specific sensory phenomena in the world. For example, we may have a high level sensor that fires when grandma comes into view but it is not a model of grandma. Our brain cannot model anything outside of it because our eyes do not see grandma. They just sense changes in illumination. To model something, one must have access to both a subject and an object. An artist can model something by looking at both the subject and the painting. The brain must sense things directly. It only has the signals from its senses to work with. To Understand the Brain, Be the Brain The most crippling mistake that most AI researchers make is that they try to understand intelligence from the point of view of an outside observer. Rather, they should try to understand it from the point of view of the intelligence itself. They need to adopt a brain-centric approach to AI as opposed to an observer-centric approach. They should ask themselves, what does the brain have to work with? How can the brain create a model of something that it cannot see until it learns how to see it? Once we put ourselves in the brain’s shoes, so to speak, representations no longer exist because they make no sense. They simply disappear. Timing is the Key to Unsupervised Learning The reason that people like Yann LeCun, Quoc Le and others in the machine learning community are having such a hard time with unsupervised learning (the kind of learning that people do) is that they do not try to “see” what the brain sees. The cortex only has discrete sensory spikes to work with. It does not know or care where they come from. It just has to make sense of the spikes by figuring out how they are ordered. Here is the clincher. The only order that can be found in multiple sensory streams of discrete signals is temporal order: they are either concurrent or sequential. Timing is thus the key to unsupervised learning and everything else in intelligence. One only has to take a look at the center-surround design of the human retina to realize that the brain is primarily a complex timing mechanism. It may come as a surprise to some that we cannot see anything unless there is motion in the visual field. This is the reason that the human eye is continually moving in tiny movements called microsaccades. Movements in the visual field generate precisely timed spikes that depend on the direction and speed of the movements. The way the brain sees is completely different from the way computer vision systems work. They are not even close. New AI Winter in the Making Discrete signal timing should be the main focus of AI research, in my opinion. It is very precise in the brain, on the order of milliseconds. This is something that neurobiologists and psychologists have known about for decades. But the AI community thinks they know better. They don’t. They are lost in a lost world of their own making. Is it any wonder that their field goes from one AI winter to the next? Artificial intelligence research is entering a new winter as I write but most AI researchers are not aware of it. See Also: AI Pioneer Now Says We Need to Start Over. Some of Us Have Been Saying This for Years Unsupervised Machine Learning: What Will Replace BackPropagation?

The World Is its Own Model or Why Hubert Dreyfus Is Still Right About AI

141

the-world-is-its-own-model-or-why-hubert-dreyfus-is-still-right-about-ai-1c7d3d42c9b9

2018-06-19

2018-06-19 09:09:46

https://medium.com/s/story/the-world-is-its-own-model-or-why-hubert-dreyfus-is-still-right-about-ai-1c7d3d42c9b9

false

1,653

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Louis Savain

I rebel, therefore I am. Working on the AGI revolution.

ebbe8e8bbd24

RebelScience

225

28

20,181,104

null

0

null

0

923e17d2b5

2017-10-08

2017-10-08 18:14:48

2017-10-08

2017-10-08 20:12:15

8

false

en

2017-10-08

2017-10-08 20:12:15

11

1c7f540105

5.374843

19

0

Questions your company should be asking before implementing machine learning.

5

Is Machine Learning Right for Your Business? Questions your company should be asking before implementing machine learning. Machine learning (ML) is all the craze right now. You hear about Elon Musk and Mark Zuckerberg debate the future of artificial intelligence and machine learning, but you wonder, how is machine learning going to actually help my business? In this article, we briefly explain what ML is and then dive into the ML-related questions your company should be asking. What is Machine Learning? Machine learning is revolutionary because it gives computers the ability to solve problems without being explicitly programmed. In a conventional computer algorithm, a programmer will specify the rules that explicitly determine what their software will do. ML algorithms work differently. At a high level, they make decisions/predictions by ingesting large quantities of historical data and using that knowledge to guide their results. Some examples of ML currently being used in businesses include: Email filters marking messages as either spam or not spam Netflix recommending what movies/shows you are likely to enjoy Google maps predicting how difficult parking will be at your destination Facebook’s facial recognition identifying people in photos Anomaly detection algorithms that can identify fraudulent purchases Here is the typical setup for doing machine learning (at a very high level): Pick a ML model/algorithm Train your model by feeding it data Use the trained model to make decisions/predictions Let’s use email spam filters as an example. The ML model an email provider might use to detect spam is the naive bayes classifier (but other applicable models exist as well). They train this model by feeding in millions of emails that are marked as spam and emails that are marked as legitimate. With the model sufficiently trained, they can use it to classify incoming emails as spam or not spam with high accuracy. For instance, if you receive an email containing the phrase “Nigerian Prince”, the ML model would remember that that phrase occurs frequently in previous spam emails and mark the incoming message as spam as well. Going from a Business Problem to an ML Algorithm The mathematical nature of ML can be very daunting. So the question I hope to address is whether or not your business can benefit from machine learning at all. The answer to that question is very situation dependent. It depends on the problem you’re trying to solve and data you are able to collect. To begin with, here are some preliminary questions your company should ask before you get started: 1. Have you already tried traditional data analytics / statistics? You might not need a solution as sophisticated as machine learning. Just knowing basic statistics about the problem you are trying to solve might be enough. An engineer at a data center could use machine learning to reduce their energy usage — perhaps, by finding complex relationships between IT load, water pumps, room temperature and other factors — or they could just look at how much energy each component is using and cut back on servers using too much energy. A retail store could use a ML model like k-means clustering to find patterns in consumer purchases (e.g. “what time do people age 20–30 go shopping?”) or they could just open a spreadsheet of the store’s transactions and manually deduce what they want to know. Basic statistics, in lieu of machine learning, might give you sufficient insight while saving you time. At the very least, it’s a good starting point. 2. Do you have data that is relevant to solving the problem? Suppose, for example, that your company is trying to perform predictive maintenance on factory equipment. In other words, you want to estimate how long a particular machine will last before it breaks. In this hypothetical scenario, you would need sensors attached to machines collecting information such as: How frequently it’s being used Vibrations it’s experiencing How old it is Room temperature, and more. Generally speaking, a machine learning algorithm without relevant data is like a detective without useful clues. The old adage holds true: garbage in, garbage out. IoT For All Newsletter Sign up for our weekly newsletter and exclusive content! 3. Do you have lots of relevant data? You have to train a ML model with a large amount of data before you can use it. For them to work with sufficient accuracy, they need to have at least thousands of data points (and preferably more). It is possible to get pre-trained models, but it’s unclear if a pre-trained model will exist for the specific type of problem you’re trying to solve. Next Steps If you still think ML is applicable, it’s worth consulting with someone knowledgeable about the different ML models. Surprisingly, the difficult part is not building these machine learning models. TensorFlow, MATLAB and R are examples of open-sourced programs that provide pre-built ML models. The difficult part is retrieving and reformatting your data from your SQL database (or whatever storage option you use) to your ML program. To illustrate the difficulty of this process, take this quote from the Google Cloud Next 2017 presentation on machine learning: “We’re getting a lot of free attention in this room and other rooms around machine learning because it’s new science, it’s unicorns and glitter, it’s all magic at this point. No data, no quality data, no machine data, no coalesced data out of 19 different databases into a single data store … no machine learning. I have no solution for anyone in this room if you say ‘but a lot of my transactional data is in my Oracle financial system, but my online system is in my e-commerce system which is hosted somewhere else, but don’t worry, all my logging data which I want to combine into learnings as well sits on my Apache servers which is at my hoster … let’s do some machine learning’. And I’ll say, ‘come back to me when you have big data’.” Again, the solution to this problem is to consult with someone familiar with both machine learning and database technology. Conclusion In summary, when thinking of implementing machine learning in your business always start simple with traditional statistics. From there, you can start to consider if it is worth consulting with someone familiar with the variety of ML models out there. They can help you put together a complete ML solution — from data retrieval, to data storage, to actually training the ML model — and deliver powerful functionality to your product or company. Alternatively, you could look into AutoML programs that programmatically do this process for you. Want all the latest advances and tech news sent directly to your inbox? Originally published at www.iotforall.com on September 24, 2017.

Is Machine Learning Right for Your Business?

334

is-machine-learning-right-for-your-business-1c7f540105

2018-04-18

2018-04-18 12:59:18

https://medium.com/s/story/is-machine-learning-right-for-your-business-1c7f540105

false

1,124

Expert analysis, simple explanations, and the latest advances in IoT, AR/VR/MR, AI & ML and beyond! To publish with us please email: contribute@iotforall.com

null

iot4all

null

IoT For All

hello@iotforall.com

iotforall

IOT,INTERNET OF THINGS,TECH,TECHNOLOGY,ARTIFICIAL INTELLIGENCE

iotforall

Machine Learning

machine-learning

Machine Learning

51,320

Narin Luangrath

Product Engineer at Leverege LLC

42204066ff96

narin_luangrath

131

9

20,181,104

null

0

null

0

e68e8e37db94

2018-03-06

2018-03-06 07:59:11

2018-03-06

2018-03-06 08:05:58

0

false

en

2018-03-06

2018-03-06 08:05:58

6

1c801fc2845b

3.015094

0

228-Artificial Intelligence-16–43158

1

Post 228 — Harvard psychologist Steven Pinker: The idea that A.I. will lead to the end of humanity is like the Y2K bug 228-Artificial Intelligence-16–43158 Words:790words Climate change. Killer robots. Russian bad actors spreading fake news in social networks. School shootings. It’s easy to feel dour about the future of mankind. But constant, widespread doomsday prophecies are not going to help — it’s only going to make matters worse. That’s according to famed Harvard cognitive scientist Steven Pinker. Hepenned an op-ed in Canadian paper, The Globe and Mail, published Saturday, making the case. His op-ed preceded the release of his new book, “Enlightenment Now: The Case for Reason, Science, Humanism, and Progress,” which was released Tuesday. “Doomsday is hot. For decades, we have been terrified by dreadful visions of civilization-ending overpopulation, resource shortages, pollution and nuclear war. But recently, the list of existential menaces has ballooned,” says Pinker. “We now have been told to worry about nanobots that will engulf us, robots that will enslave us, artificial intelligence that will turn us into raw materials and teenagers who will brew a genocidal virus or take down the internet from their bedrooms.” Billionaire tech titan Elon Musk is one of the loudest voices publicizing the potential threat of AI. He has said robots will be able to do everything better than humans; competition for AI at the national level will cause World War 3; and AI is a greater risk than North Korea. Renowned physicist Stephen Hawking has said AI could be the best or the worst event in the history of civilization. His warnings of the latter were frightening. “It brings dangers, like powerful autonomous weapons, or new ways for the few to oppress the many. It could bring great disruption to our economy,” said Hawking. While preparing, or even over-preparing, for threats may seem harmless, there are risks, says Pinker. “Apocalyptic thinking has serious downsides. One is that false alarms to catastrophic risks can themselves be catastrophic,” says Pinker. “Sowing fear about hypothetical disasters, far from safeguarding the future of humanity, can endanger it.” The Harvard professor pointed to the nuclear arms race of the 1960s and the 2003 invasion of Iraq as examples of catastrophic thinking doing harm. Pinker says constant fear-mongering can make it harder for the human brain to correctly distinguish between a legitimate and a false threat. “Some of the threats facing us, such as climate change and nuclear war, are unmistakable, and will require immense effort and ingenuity to mitigate,” he writes. “Folding them into a list of exotic scenarios with minuscule or unknown probabilities can only dilute the sense of urgency.” If every doomsday scenario feels possible, then people are actually disincentivized to take action, says the cognitive scientist. “If humanity is screwed, why sacrifice anything to reduce potential risks? Why forgo the convenience of fossil fuels or exhort governments to rethink their nuclear weapons policies? Eat, drink and be merry, for tomorrow we die!” says Pinker, explaining the thinking process that results. An overblown notion that technology will be our end is not new. “Some threats strike me as the 21st-century version of the Y2K bug,” he says, referring to the mistaken panic that because of a flaw, dates with the year 2000 and beyond would cause computers around the world to go haywire. “This includes the possibility that we will be annihilated by artificial intelligence, whether as direct targets of their will to power or as collateral damage of their single-mindedly pursuing some goal we give them,” writes Pinker in The Globe and Mail. Intelligence does not necessarily translate to evil, says Pinker. Also, if humans are able to create unbelievably smart machine intelligence, they will also be smart enough to test said technology before giving it control of the world, he says. Further, the idea that AI is both smart enough to take over and dumb enough to do so by accident is not logical, Pinker says. Other overhyped doomsday threats include mass starvation and resource scarcity, says Pinker. Even when it comes to realistic threats, Pinker is relatively optimistic that things like nuclear war and climate change can, with careful and diligent work, be mitigated. “Unsolved does not mean unsolvable,” says Pinker. “Pathways to decarbonizing the economy have been mapped out, including carbon pricing, zero-carbon energy sources and programs for carbon capture and storage. So have pathways to denuclearization, including strengthening international institutions, de-alerting nuclear forces, stabilizing systems of deterrence and verifiably reducing (and eventually eliminating) nuclear arsenals,” he says. “The prospect of meeting these challenges is by no means utopian,” says Pinker. “But we know that there is one measure that will not make the world safer: moaning that we’re doomed.” #A.I. #Utopian #Y2K #Harvard #StevenPinker

Post 228 — Harvard psychologist Steven Pinker: The idea that A.I.

0

post-228-harvard-psychologist-steven-pinker-the-idea-that-a-i-1c801fc2845b

2018-03-06

2018-03-06 08:05:59

https://medium.com/s/story/post-228-harvard-psychologist-steven-pinker-the-idea-that-a-i-1c801fc2845b

false

799

uhemi

null

UhemiEducation

khchanaq@gmail.com

uhemieducation

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Uhemi

null

9ea0ab50d247

cfo_20738

1

20,181,104

null

0

null

0

null

2016-05-17

2016-05-17 19:15:08

2016-05-17

2016-05-17 19:17:32

1

false

en

2017-01-07

2017-01-07 23:43:06

0

1c80abccfb29

3.10566

2

1

0

On progress, simplicity and the reasons we fight complexity even though we have lost this fight a long time ago.

5

Photo by Greg Rakozy The End of Simple On progress, simplicity and the reasons we fight complexity even though we have lost this fight a long time ago. We live in a time of unprecedented change. For the past 70.000 years we have witnessed the progress of our cognitive abilities, we have grown to shape the world we live in and we have learned to rely on collective intelligence which allowed us to make progress on a global scale. 70.000 years ago we went through the first major revolution as human beings: we became homo sapiens with the power to change the world. And from that moment on we have changed everything. Thousands of years have enabled us to explore our own potential. The evolution of our ideas has turned us into innovators who go above and beyond what seems possible. We have built cities, we have learned to fly and we connected everyone and everything. 12.000 years ago the agricultural revolution allowed us to form economies and societies and we developed elaborate political systems. Just 500 years ago the scientific revolution kick started exponential growth in the technologies we use today. By connecting the minds of millions we have created a collective intelligence that spans around the globe and provides us with what we call progress. Today we are looking at the most complex network of minds in our history. Technologies are crossing boundaries in artificial intelligence and robotics, corporations around the globe rethink the meaning of work, diverse world views, cultures and religions collide and politicians everywhere desperately try to protect local power and authority while cutting diplomatic ties. We live in a world of chaos. And our minds are not prepared for this. For 70.000 years we have trained our minds to look for patterns and understand the world in terms of cause and effect. In a chaotic world this no longer works for us. The speed of technological change, the globalization of markets, the increasing cultural diversification and the lack of a global political system create a level of uncertainty and ambiguity that our minds fight in the most dysfunctional way possible: we oversimplify. The need to understand, to predict and to control leads us in the wrong direction. We are looking for simple answers, simple solutions and simple truth. Yet, this is not the time for simple solutions. Simple solutions no longer work. This is not the time to build walls, this is not the time to stop exploring the unknown, this is not the time to fight complexity with single-cause explanations. This is the time to adopt a new belief system, the time to question our own assumptions and the time to see the opportunity to create the future we want. This is the end of simple answers to complex questions. This is the end of slow and steady progress. And this is the end of linear growth. We are entering the age of collaboration with strangers, an age in which we accept the unknown and the fear surrounding us without letting it shape our decision making towards simple solutions. Instead of isolating ourselves in the false belief that this will protect the status quo we need to learn to leave our comfort zones despite the fear of losing control. We need to go into this future with confidence, patience and forgiveness. Rather than constantly looking for simple solutions we need to relearn to ask better questions. Robots and artificial intelligence will be around us, companies will be virtual and global, the workforce will be international and diverse, data will become the currency of this future and we will further explore outer space. But no matter what, we are the ones deciding how this future is going to impact the wellbeing of humans worldwide. We are the ones who provide the ideas that make the world a better place or not. We are the ones demanding purpose beyond greed and fear. This is not the time to demand oversimplified answers. This is the time to ask the right questions. We need to rethink our vision of the next 70.000 years. Simple answers won’t define this future. Instead our progress will be guided be the questions we ask today. I’m Christoph Burkhardt and I have helped organizations around the globe for years to innovate in this chaotic environment by asking better questions. I believe that things get better and that we are making progress in many areas of human life. Yet, if you want to change things for the better it’s not a simple answer you should be looking for, we all should be looking for better questions.

The End of Simple

2

the-end-of-simple-christoph-burkhardt-1c80abccfb29

2017-11-02

2017-11-02 17:20:27

https://medium.com/s/story/the-end-of-simple-christoph-burkhardt-1c80abccfb29

false

770

null

Innovation

innovation

Innovation

59,190

Christoph Burkhardt

Innovation Strategist | Award-winning Speaker | #complexity #antifragility #evolution #biases

5d9ed1099dfb

chrisburkhardt

275

476

20,181,104

null

0

null

0

f5af2b715248

2018-01-13

2018-01-13 17:33:54

2018-01-13

2018-01-13 17:36:23

3

false

en

2018-02-18

2018-02-18 18:29:38

6

1c8259f5d87c

2.425472

22

1

0

TL;DR: It’s the future

5

WTF is Decentralized Artificial Intelligence? TL;DR: It’s the future The artificial intelligence revolution has started and organizations in manufacturing, transportation, retailing, finance, entertainment, education, and nearly every other industry are transforming their core processes and business models to take advantage. Not only is AI transforming industries and companies, it plays a huge role in our daily life. How we get around, what we decide to eat, how much we pay for a beer are often dependent on AI solutions. Working with AI (specifically machine learning) is a blast and leveraging tools like PyTorch and Tensorflow make it super easy to build really interesting and valuable tools. There are enormous advantages to be gained from AI and we are just now starting to see individuals, organizations, and governments reap the rewards. I’ve always enjoyed working with AI but my blood really started pumping when I began trying to implement AI solutions on blockchains. For anyone who is unfamiliar with the blockchain, it’s a digital and decentralized ledger technology that records all transactions chrnologically and publicly. It is the foundation that cryptocurrencies are built on because it’s transparent, speeds settlements, reduces transaction costs, and it’s controlled by the users. Even though this technology hasn’t fully matured, it’s already proving to be a key piece in the advancement of AI. This is because “old school” or traditional AI follows a centralized distribution pattern (one agent controls the world). We get access through an API that is part of a cloud based service and the software packages are on remote servers of different AI providers. Thankfully, we are moving towards the “new school”. Imagine AI being a collaborative solution by a distributed group of intelligent agents. AI can run, train, and even make decisions on local devices in decentralized networks like the blockchain. That is decentralized AI! As we move forward, I see three tremendous advantages of decentralized AI over traditional AI: 1. Minimal latency (no dependency on network connection) 2. Training is more efficient (done in a decentralized way) 3. Less Power Consumption (again, no dependency on network connection) This concept is gaining ground fast. Recently, computers (Google’s TPU) and phones have been optimized with AI in mind. Also, Google really stepped up to the plate with their Federated Learning concept which boasts a decentralized, collaborative approach to training data. By keeping all training data on the device, there is no dependency of storing data in the cloud. We are building the foundation for “new school” AI. Make sure you take note. Personally, I’m most intrigued with leveraging AI in decentralized autonomous organizations (DAOs). In a nutshell, DAOs have some or all of the decision making responsibilities done by intelligent agents on the blockchain. But that will be a blog post for another day :) The future is bright because it’s closer than you think. Onward, Ben Stewart This story is published in The Startup, Medium’s largest entrepreneurship publication followed by 293,189+ people. Subscribe to receive our top stories here.

WTF is Decentralized Artificial Intelligence?

260

wtf-is-decentralized-artificial-intelligence-1c8259f5d87c

2018-05-30

2018-05-30 20:36:35

https://medium.com/s/story/wtf-is-decentralized-artificial-intelligence-1c8259f5d87c

false

497

Medium's largest publication for makers. Subscribe to receive our top stories here → https://goo.gl/zHcLJi

null

The Startup

null

swlh

STARTUP,TECH,ENTREPRENEURSHIP,DESIGN,LIFE

thestartup_

Blockchain

blockchain

Blockchain

265,164

Ben Stewart

Co-Founder & CTO at Blackbox AI. Blockchain Engineer/Evangelist/Advisor

c1187cebfed9

BIGbenStew

154

37

20,181,104

null

0

null

0

null

2017-10-14

2017-10-14 07:37:54

2017-10-14

2017-10-14 07:42:18

0

false

en

2017-10-14

2017-10-14 07:42:18

1

1c85c145c68d

0.784906

3

0

Today all the tax payers are taking lot of the time for reconciling the GSTR2 report with Purchase Invoice by checking one by using the…

1

How to Convert GSTR2 Report to Excel Format ? Today all the tax payers are taking lot of the time for reconciling the GSTR2 report with Purchase Invoice by checking one by using the offline utility software provided by the GST Portal, but here I am introducing the new utility created in the Microsoft Excel Format which will helps us to convert the Json File which is created by the GST Portal to Microsoft Excel Format and it will helps us to reconcile the data by using Vlookup formula in the less time and you can relax in the GSTR2 filing process. Step-1 Download the GSTR2A Report from the GST Portal Step-2 Download the Offline Utility software of the Json file convertor to Excel using the below link (It will work only above version of Microsoft Excel 2010 and above) Step-3 Convert the Json file to CSV Format by uploading the file to site https://json-csv.com/ Step-4 Convert this to CSV Format Step-5 Copy and Paste the data in the CSV File to the Excel Utility and pasting same using Paste Special values Step-6 Convert the same and generate the file , you can view the excel data in the Output folder.

How to Convert GSTR2 Report to Excel Format ?

3

how-to-convert-gstr2-report-to-excel-format-1c85c145c68d

2018-05-27

2018-05-27 15:28:58

https://medium.com/s/story/how-to-convert-gstr2-report-to-excel-format-1c85c145c68d

false

208

null

Data Science

data-science

Data Science

33,617

Count Magic

Count Magic is India’s 1st Integrated Cloud based Suite of Products for making business simple in the new GST regime

8a1ab950af7c

countmagicindia

0

1

20,181,104

null

0

null

0

f898f5443d54

2017-11-30

2017-11-30 16:49:12

2017-12-12

2017-12-12 08:01:12

11

false

en

2017-12-12

2017-12-12 16:46:45

3

1c87d8d4e81c

9.175472

30

0

Download “The State of AI 2017: Inflection Point”

5

The State of AI 2017: What Matters Most Download “The State of AI 2017: Inflection Point” Watch the keynote: “The State of AI 2017: Inflection Point” Interested in AI? Sign up for our blog posts. While hype around AI is at a peak, and some expectations may exceed results in the short term, we believe AI represents a paradigm shift in technology that warrants the attention it is receiving. In 2017 AI reached an inflection point, driven by milestones in investment, capability, entrepreneurship and adoption. The implications for consumers, companies and society will be profound. Below, we highlight three key takeaways from our extensive report, the State of AI 2017: Inflection Point: (1) An inflection point in adoption, as AI ‘crosses the chasm’ to the early majority of buyers; (2) The profound implications of AI, including shifts in sector value chains, the creation of new business models, and benefits and risks to society; and (3) The dynamics of the UK’s 400 AI startups — including a shift to the era of AI applications. Developed in association with Numis, our report draws on new data and over 400 discussions with ecosystem participants to go beyond the hype and explain the reality of AI today, what is to come and how to take advantage. Every chapter includes actionable recommendations for executives, entrepreneurs and investors. Download the full report here. 1. An inflection point in AI adoption Awareness of AI has reached an inflection point. Given media attention and vendor marketing, executives’ awareness of AI is high. In January 2016, the term ‘Artificial Intelligence’ didn’t feature in the top 100 terms searched for by executives on Gartner.com. By May 2017, the term ranked №7 (Gartner). However, understanding of AI among buyers is low. Technology principles, use cases and deployment methodologies are poorly understood. “Even among CIOs, understanding of AI is extremely low.” (Senior Executive, global consumer packaged goods company). AI adoption is crossing the chasm. 20% of AI-aware executives say they have adopted one or more AI-related technology at scale, or in a core part of their business (McKinsey Global Institute). While nascent, we believe AI adoption is ‘crossing the chasm’ from innovators and early adopters to the early majority. Adoption of AI will increase significantly as buyers seek to unlock value from data and avoid losing competitive advantage. 75% of executives say AI will be “actively implemented” to some degree in their organisations within three years (Economist Intelligence Unit). High tech, automotive and assembly, and financial service firms lead AI adoption. Spending on AI will increase most in sectors that currently lead adoption, implying that a bifurcation in adoption — by sector — may emerge. Poorly articulated business cases weigh on adoption. Better articulation of ROI by AI vendors can catalyse adoption. “Buyers feel there’s value, but are nervous around making bets.” (Vice President, global consumer products company) Three quarters of buyers are deploying AI to improve decision-making and enable process automation, while extensive media attention and numerous pilot projects relate to chatbots. The C-suite is key for initiating, selecting and funding AI initiatives. In two thirds of organisations, the CTO or CIO make AI technology decisions given its cross-functional implications. AI deployment strategies are varied, with a mix of ‘build’ and buy’ strategies, and in a state of flux. ‘Hybrid’ approaches are typical. A quarter of companies deploying AI today prefer to purchase a standalone solution. Lack of skills is the primary challenge for companies deploying AI. Defining an AI strategy, identifying use cases for AI, and securing funding for AI initiatives are additional difficulties. 2. AI will have profound implications AI’s benefits can be abstracted to four: innovation (new products and services); efficacy (the performance of tasks more effectively); velocity (the completion of tasks more quickly); and scalability (by enabling software to undertake previously human tasks). These benefits will have profound implications for companies, consumers and society, including the following. 1.New market participants: By automating capabilities previously delivered by human professionals, AI will reduce the cost and increase the scalability of services, significantly broadening participation in select markets. AI will enable automated diagnosis for a growing proportion of conditions. The marginal cost of diagnosing a patient using an AI algorithm will be nil. With smartphone adoption in developing economies increasing rapidly, from 37% in 2017 to 57% by 2020E (GSMA), barriers to access will also fall. By transferring the burden of diagnosis from people to software, global access to primary care will increase. Millions of additional individuals will benefit from primary care, while the market for providers of relevant and associated technologies will expand. 2.Shifts in sector value chains: In multiple sectors AI will change where, and the extent to which, profits are made. In the insurance sector, car insurance accounts for 42% of global insurance premiums (Autonomous Research). As AI-powered autonomous vehicles gain adoption, the frequency of accidents will reduce — and with this, insurers’ revenue. UK car insurance premiums are expected to fall by as much as 63%, causing profits for insurers to fall by 81% (Autonomous Research). Insurers must anticipate and plan for a profound shift in their sector’s value chain. 3.New commercial success factors will determine a company’s ability to be successful. Success factors in the age of AI include: the vision to embrace AI; ownership of large, non-public data sets to train and deploy market-leading AI algorithms; a willingness to evaluate the opportunities and risks of sharing training data with partners and competitors; the ability to attract, develop, retain and integrate data scientists within an organisation; the ability to form effective partnerships with best-of-breed third-party AI software and service providers; a willingness to understand and respond to regulatory challenges posed by AI; a shift in mindset to the use of software that provides probabilistic instead of binary recommendations; the ability to manage workflow changes that result from the implementation of AI systems; and the ability to manage challenges of organisational design and culture as AI augments, and in some cases replaces, personnel. 4. Changes in companies’ competitive positioning: New leaders, followers, laggards and disruptors will emerge. Among providers of AI: Platforms — including Google, Amazon, IBM and Microsoft (GAIM) — provide the AI infrastructure, development environments and ‘plug and play’ AI services used by many developers and consumers of AI. With vast data sets, world-class AI teams and extensive resources, select GAIM vendors will accrue value as platforms that support the provision of AI. However, GAIM lack the strategic desire, data advantage and domain expertise to address myriad industry-specific use cases required by businesses in sectors from manufacturing and agriculture to retail and finance. This presents opportunities for Disruptors. Disruptors are early stage, AI-led software companies tackling business problems in a novel way using AI. Disruptors will enable buyers that embrace them, while eroding the value of those that lack the foresight to do so. Select Disruptors will become tomorrow’s incumbents, or be acquired by today’s. Among buyers of AI (today’s medium-sized and large companies): Leaders will emerge in key industries. Leaders will extend their competitive advantage and enjoy two benefits: (1) In the ‘data economy’, economic returns will accrue disproportionally to companies that can extract value from information most effectively; (2) Data network effects create wider competitive moats. Larger volumes of training data enable better algorithms, which deliver better products and services, which attract more customers, who provide more data. Laggards are buyers that lack the will or organisational ability to use AI effectively. While some enterprises will lack the foresight to adapt, more will falter due to limited organisational capability. In the ‘data economy’, laggards will rapidly lose competitive advantage and market share. 5. New business models: AI, growth of ‘x-as-a-service’ consumption, and subscription payment models will obviate select business models and offer new possibilities in sectors including transport, insurance and healthcare. In the transport sector, AI will transform the economic fabric of ownership and insurance. Cars are parked for an average of 96% of their lives (UITP Millennium Cities Database). Despite the cost and inefficiency of private car ownership, the model has been necessary to enable spontaneity, point-to-point convenience, comfort, privacy and security during travel. An autonomous vehicle, summoned when required from a distributed fleet and used for the duration of a journey, will offer the same benefits while optimally utilising a fleet. With the cost of the driver removed, and the cost of the vehicle and insurance divided over a greater volume of trips in a given period, the marginal cost of a journey will be lower. [X] With growing use of transport-as-a-service subscription models, in which consumers pay a low monthly fee for on- demand access to a fleet of autonomous vehicles, private car ownership will decline. 6. Benefits and risks to society: AI will provide benefits to society including improved health, broader access to services and more personalised experiences. It will also present risks regarding job displacement, bias, conflict and privacy — which we describe in the report. Regarding job displacement, AI will enable the automation of several occupations that involve routine and repetition — from truck driving to telemarketing. Truck driving comprises 3.6 million jobs in the US (American Trucking Association). Analysis of UK census data since 1871 shows that historically, contracting employment in agriculture and manufacturing — a result, in part, of automation — have been more than offset by rapid growth in the caring, creative, technology and business service sectors (Deloitte). Whether or not, over time, AI creates more jobs than it destroys, the short period of time in which a large number of workers could be displaced, coupled with a reduction in the availability of similar roles, could prevent those who lose their jobs from being rapidly re-absorbed into the workforce. Social dislocation, with political consequences, may result. 3. The dynamics of UK AI: the application wave AI entrepreneurship is thriving. The number of AI companies founded annually in the UK has doubled since 2014. A new UK AI company has been founded every five days, on average, since 2014. There are 400 independent, early stage software companies in the UK with AI at the heart of their value proposition: We’ve entered a second wave of entrepreneurship — the era of applications. While early innovation focused on AI research or core technologies applicable to multiple sectors, over 80% of today’s UK AI startups are vertically-focused business-to-business (B2B) suppliers addressing a specific business process or sector. Few companies (one in ten) sell direct-to-consumer given the difficulty of acquiring training data from a ‘cold start’ and the deployment of AI by global consumer technology companies. Entrepreneurial activity is unevenly spread. More UK AI companies (one in seven) address the marketing & advertising function than any other. For companies with a sector focus, finance dominates. In select sectors (manufacturing) and business functions (finance), activity appears modest relative to market opportunities. UK AI companies comprise nearly half the European total. AI is well represented in the UK, with a slightly higher proportion of startups focused on AI than in Europe (excluding the UK) or the US. UK AI companies are nascent. Two thirds of companies are in the earliest stages of their journey, with Seed or Angel funding. The sector, however, is maturing rapidly. UK companies are less embryonic than their European counterparts. Over 40% of companies we meet have yet to receive recurring revenue. The journey to monetisation for AI companies can be longer given technical challenges, long sales cycles in a B2B-driven market, and client integration requirements. Globally, investments into early stage AI firms are typically 20%-50% larger than capital infusions into general software companies of comparable stages. Staging of capital into UK AI companies can be atypical. One in three growth stage companies raised a significantly larger post-Angel round than is typical. There’s more to explore in the ‘State of AI 2017’. The full report includes: An accessible introduction to AI for the non-specialist (Part 1). An analysis of AI’s proliferating applications and profound implications (Part 2). A market map and dynamics of the UK’s 400 AI startups – plus perspectives from the UK’s leading AI entrepreneurs (Part 3). Our investment framework describing the 16 keys to success for AI companies (Part 4). Download “The State of AI 2017: Inflection Point” Watch the keynote: “The State of AI 2017: Inflection Point” Interested in AI? Sign up for our blog posts.

The State of AI 2017: What Matters Most

277

the-state-of-ai-2017-what-matters-most-1c87d8d4e81c

2018-05-29

2018-05-29 01:47:54

https://medium.com/s/story/the-state-of-ai-2017-what-matters-most-1c87d8d4e81c

false

2,087

A collection of stories and experiences from the early-stage technology and venture capital communities. Curated by MMC Ventures.

null

MMC writes

anna.slemmings@mmcventures.com

mmc-writes

VENTURE CAPITAL,EARLY STAGE,INNOVATION,INSIGHTS,TECH

MMC_Ventures

Venture Capital

venture-capital

Venture Capital

32,826

David Kelnar

Partner and Head of Research at MMC Ventures. 2x CEO/CFO. Love tech, venture capital, trends and triathlon. http://www.linkedin.com/in/kelnar

4934916676ca

dkelnar

2,055

620

20,181,104

null

0

null

0

null

2018-09-08

2018-09-08 22:56:59

2018-09-08

2018-09-08 23:02:57

3

false

tr

2018-09-08

2018-09-08 23:05:05

0

1c8895b28d86

2.251887

16

7

0

Blockchain teknolojisinin merkezi, madencilik kavramıdır. En basit anlamda madencilik, bir işlemin meşruluğunun doğrulanabileceği bir…

5

MATRIX Madencilik Makinesi Prototipi Blockchain teknolojisinin merkezi, madencilik kavramıdır. En basit anlamda madencilik, bir işlemin meşruluğunun doğrulanabileceği bir süreçtir. Bunu başarmak için, birçok blockchain Proof of Work (PoW) konsensüs algoritması kullanmaktadır. PoW konsensüs algoritmalarının iç işleyişi bu yazının kapsamı dışındayken; asimetri anahtar olduğunu söylemek yeterlidir. Başka bir deyişle, PoW’a merkezi olan karmaşık matematiksel bulmaca yeterince zor olmalıdır; çözüm bulunduğunda, kanıtlanması ve onaylanması son derece kolay olmalıdır. Madencilik, büyük miktarda boşa harcanan gücü tüketen, özel bir bilgisayar donanımını giderek arttırıcı bir şekilde harcayan kaynak yoğun bir görevdir. Aslında, bu boş güç, MATRIX AI Network’ün en yeni yapay zeka (AI) teknolojisinden yararlanarak çözmeyi denediği başlıca sorunlardan biridir. MATRIX’in yenilikçi fikir birliği algoritması HPoW olarak adlandırılan hibrid PoW / PoS karması olarak, Markov-Chain Monte Carlo (MCMC) algoritmalarını çalıştırmak için, bu geniş bilgi işlem kaynaklarını kullanmaktadır. Bunlar değerli gerçek dünya uygulamalarına sahiptir . Şu anda kanser teşhisinin hızını ve doğruluğunu artırmak için araştırma hastaneleri ile işbirliği içinde kullanılmaktadır. MATRIX Madencilik Makinesi Anakartı MATRIX Madencilik Makinesi Deneysel Sonuçları Bununla birlikte madencilik, yapbozun sadece bir parçasıdır. MATRIX, bu çalışmalara paralel olarak ayrıca özel madencilik donanımı tasarlamaktadır. Bu makale ve aşağıda vereceğimiz video, araştırma ve geliştirme (AR-GE) ekiplerimizin yaptığı çalışmaları göstermektedir. MATRIX Araştırma ve Geliştirme Departmanı ekip üyesi Mr. Yong Li, MATRIX Madencilik Makinesi ile ilgili bir güncelleme sunmaktadır. Video’yu Türkçe altyazı izlemek için Youtube altyazı seçeneğini etkinleştirmeyi unutmayın. MATRIX Madencilik Makinesi’nin özelliği nedir? MATRIX Madencilik Makinesi, dört hedeflenen gereksinimi karşılamak için FPGA’lardan (Alan- Programlanabilir Geçit Dizileri) yararlanır. Bunlar yüksek hesaplama gücü, düşük enerji tüketimi, uygun madencilik ve üstün ısı dağılımıdır. FPGA kullanmanın yararlarından biri de, mantığın doğrudan donanımda saklanabilmesidir. Mevcut MATRIX Madencilik Makinesi prototipi şu anda dört çekirdek ve dört farklı para biriminin madenciliğini destekleyebilmektedir. Özellikle, bu para birimleri MAN, BTC, ETH ve Filecoin’dir. Laboratuar ortamında, MATRIX Madencilik Makinesi 50k TPS’lik sistem çıktı hızını aşmıştır. · MATRIX Madencilik Makinesi, MAN tokenleri için gerekli mi? MATRIX Madencilik Makinesi, MAN tokenleri için gerekli değildir. Bununla birlikte, MATRIX Madencilik Makinesi, hazır özel makinelere göre madencilikte daha verimli olacaktır. Düzenli tüketici sınıfı donanım kullanarak madencilik yapacaksanız, en az bir çekirdek, madencilik faaliyetlerine tahsis edilmelidir. Bu özel donanım madenciler için keşfetmeye değer özelliktedir. · MATRIX Madencilik Makinesi ne zaman piyasaya sürülecek? Mevcut plan dahilinde, 2018 sene sonuna kadar piyasa çıkarılması planlanmaktadır. Nazik bir hatırlatma olarak, MATRIX Testnet’in 2018 Eylül’ün sonunda yapılması planlanıyor. Ayrıca Mainnet 2018 yılsonu olarak planlanmıştır. · Halka açık olacak mı? Evet, MATRIX Madencilik makinesi halka açık olacak ve halk tarafından alınabilecektir. Bir MATRIX Madencilik Makinesinin nasıl alınacağı ile ayrıntılar daha sonra prosedürü ile birlikte paylaşılacaktır. Eylül ayı boyunca MATRIX AI Network, projenin teknik özelliklerini anlatan birkaç video ve makale yayınlayacaktır. Bunlar MATRIX Madencilik Makinesi, MATRIX Otomatik Kodlama Akıllı Sözleşmeleri, MATRIX Güvenli Sanal Makine ve MATRIX’in Dijital Varlık Kasası gibi başlıklardan oluşmaktadır. Bizimle olmaya devam edin!! Written by: MATRIX AI NETWORK Translated by: Matrix AI Network Turkey

MATRIX Madencilik Makinesi Prototipi

735

matrix-madencilik-makinesi-prototipi-1c8895b28d86

2018-09-08

2018-09-08 23:05:05

https://medium.com/s/story/matrix-madencilik-makinesi-prototipi-1c8895b28d86

false

451

null

Blockchain

blockchain

Blockchain

265,164

Matrix AI Network Turkey

Official Turkish Community

c0d52c47b03d

matrixaiturkey

96

415

20,181,104

null

0

null

0

null

2018-02-25

2018-02-25 05:25:29

2018-02-25

2018-02-25 05:28:53

1

true

en

2018-02-25

2018-02-25 05:28:53

3

1c8903c0ec25

3.384906

3

0

AI is great for most business areas within companies. It promises new insight, prediction, analysis — which allows firms to target their…

3

Value Proposition of AI for IT AI is great for most business areas within companies. It promises new insight, prediction, analysis — which allows firms to target their most promising prospects in sales, most troublesome (and prospective profitability) customers, and links current data with future expansion. Well, that’s great for the business units, and for Sales and Marketing, and for Finance, and even for Operations. Primarily, IT focuses on “what to analyze, how to analyze and provide output that is meaningful to the business”. The business then asks “How can I apply this insight to make better decisions and for direction in both tactics and strategy”. We all know about Amazon’s use of AI, to use a well-known example, about customer product recommendations, or Netflix’s movie recommendations. A more complex variation of this would be Walmart’s predictive analytics team using AI to anticipate best price points for new products. How about IT? is it simply a means to an end, a tactical vehicle for a strategic roadway? The answer, of course, is no. Beyond the obvious — that IT, as generally used, is the tool for designing, implementing and reaping the benefits of AI, IT itself could use AI. How so? here are some possibilities: - Infrastructure: AI can be used for IT asset management. How many laptops and PCs broke down in the the past 5 years? use predictive analytics to find out possible future costs, by department, region, function, job title and even time of year (perhaps the after-parties in December are a little too, er, wild). Similarly, Software assets — licenses, subscriptions to software and data services — can be analyzed to find out rate of use, peak usage vs. license costs and predictive cost management and demand. - Operations: Already in wide use at call centers and tech support areas, AI includes staffing, chat bots,. Now, with increasing use of Amazon Alexa, Google Home and other audible assistants, users can resolve their issues without human contact in most cases. This does not mean that humans are not needed — in fact, quite the contrary — the call center / tech support personnel’s valuable time can be spent solving higher level issues, instead of “my shared drive connection is lost” type simpler issues. Other potential functions include speaking to a device and asking, in plain English (or other language) for requests such as “Please give me a report of all cost centers with expense ratios higher than 15% from January through June of this year”. - Research / Investigation — IT Operations, Systems Management, Compliance: Much of workers’ time is taken up by looking for and researching something — why something was or wasn’t done, why it took so long, or how something connects to something else. AI could help, by using neural networks to sift through data, and by analyzing past data streams, to come up with potential explanations. Examples include “Why was the invoice rejected?” to “Why did the product specs have another layer of safety features after approval by Compliance?” Constant analysis of vast streams of data on just about every function, task, process, policy, project, and mandate might sound like a vast headache; but used properly, it will alleviate many bottlenecks in process and operations, and produce cost savings and open up new ways of looking at, and doing things to increase productivity. The devil is in the details though — and as the age of AI opens up at a head-spinning pace, prepare to learn about it, use it and stay ahead of the game. Project Charter / Design / Approval / Funding / Implementation / Oversight / Completion: Projects involve a number of people, policies and process in large firms. AI is currently used to automate some of these steps, with the potential to automate many more steps in the near future. The idea is not simply to have a workflow process automation, but use intelligent programs to probe the cost benefit and burn rate areas and predict the final run rate and project completion time and budget. Not to worry, if you are a project manager — the knowledge you have built up will be needed by any automation team. Are you killing your own future? No — your expertise will propel you to higher value-add processes, so you can spend your time analyzing the project from a higher perspective instead of worrying whether functional testing can start before any project milestone prerequisite is done. If you don’t know where to start, there are several online courses that explain AI in greater detail. If you are interested in the business implications, select, from one of the many offerings from any of these online education providers. Many of the courses are free, others might have a marginal fee. Note that many colleges also have online classes, some of which may be free — or not. Check Cornell, MIT, Stanford etc. Some of these providers are: www.coursera.com www.udemy.com www.udacity.com Hope this gets you thinking [more] on how you can contribute to your business’ IT area.

Value Proposition of AI for IT

73

value-proposition-of-ai-for-it-1c8903c0ec25

2018-03-05

2018-03-05 20:50:01

https://medium.com/s/story/value-proposition-of-ai-for-it-1c8903c0ec25

false

844

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Hari Santanam

I am interested in AI, Machine Learning and its value to business and to people. I occasionally write about life in general as well.

944f4840db46

hari.santanam

96

58

20,181,104

null

0

null

0

null

2018-08-20

2018-08-20 08:02:47

2018-08-20

2018-08-20 08:11:59

1

true

en

2018-08-20

2018-08-20 08:11:59

11

1c8c2c74d440

3.335849

6

2

0

How we treat robots in the future may significantly rely on how we expose children to artificially intelligent supertoys today.

5

In the Age of AI, Children Should Befriend Their Supertoys Photo Credit: Andy Kelly There is no doubt that artificial general intelligence (AGI) — an AI that is capable of generating human-level intelligence — is on its way. It’s only a matter a when, not if. According to some researchers, the quest to developing AGI may take longer than expected. But even then, the quest continues on. As we continue moving forward, however, certain questions are already being asked. The more frequent question being: To what extent should children be exposed to technologies which rely on AI systems? According to a report by Edison Research and NPR, “one-in-six Americans (16%) owns a voice-activated smart speaker.” By comparison to the number of AI speakers owned a year before, the number is now up by 128% and still rising. The impact that’ll have on children within households which own AI speakers is yet to be determined, but I am of the opinion that it’ll largely be a positive one. In December of last year, I’d purchased a Google Home Mini for my brother, his fiancé, and their daughter. I figured each of them could find out ways to integrate the Google Home into their daily lives and make things a bit easier. What I didn’t expect (at least not so quickly) was this: My Brother’s Fiancé: “Okay, peanut. Do you want mommy to read you a bedtime story or do you want Google?” Their Daughter (6-years-old): *snickers* I want Google to read me a story!” I laughed alongside her, but clearly for different reasons. My niece laughed because she’s gangster like that and knows she just retired her parents in favor of Google. I, on the other hand, laughed out of disbelief that we’ve now reached a point where children are now comfortable enough around AI that they award them with the role of helping them sleep. According to the information technology research company Gartner, 75% of all households within the U.S. will own an AI speaker by 2020. Meaning, around 75% (+/-) of children in the U.S. will have similar, if not more advanced, experiences as my niece. And companies like Google and Amazon are fully aware of this. Last year, both Google and Amazon had launched skills for their own AI speakers that were specifically programmed for children. In doing so, it allows children to be exposed to intelligent systems that aren’t biological by nature, thus preparing them for a future where there are more non-biological intelligent systems than that of biological, i.e. humans. And while privacy concerns continue to pervade throughout the conversation of children’s exposure to AI — especially AI which are retrofitted inside of toys — there is simply no doubt that a growing number of children are going to be exposed, daily, to supertoys that talk back to them. Some are even going as far as labeling this generation as Generation AI. This emerging development of children owning supertoys — toys embedded with neural networks — reminds me of the popular sci-fi film AI: Artificial Intelligence directed by Steven Spielberg. More specifically, it reminds me of “Teddy” — the stuffed robot bear that befriended the robotic boy David. As noted by the video provided above, Teddy was such a significant character in the entire film that it eventually came off as more “human” than even the human characters. It wasn’t just cognitively intelligent, but equally emotionally intelligent and loyal to a fault. While this presented certain problems in its relationship with the robotic boy, David, one could easily understand the benefits of human children having a loyal robot friend like Teddy. Time and again, research has shown that, when robots are given anthropomorphic features, humans have the unique tendency to empathize with them. And, thus, respond in concern when those robots are “harmed” in any way. This is an important facet of the human mind that will grow in significance as we become more exposed to AI systems, whether they come in the form of a personal robot or a supertoy. Imagine children being exposed to this specific mindset at a young age, where, by the time they’ve reached adulthood, their perception of robots will be no different from their perception of humans: ‘Do no harm to me and I’ll do no harm to you. And if you’re in danger, I’ll be there for you.’ But achieving so requires that children are exposed to AI systems at a young age. It requires for parents to allow their children to befriend their supertoys, even if they’re not at the stage of intelligence that Teddy is in the film AI. To not expose them to these technologies at a young age, I fear it’ll only result in a reduction of likelihood that adults of the future will treat robots with care and love; rather, with viciousness and cruelty as conveyed by adult humans in the film. This article was originally published on Serious Wonder.

In the Age of AI, Children Should Befriend Their Supertoys

43

in-the-age-of-ai-children-should-befriend-their-supertoys-1c8c2c74d440

2018-08-20

2018-08-20 14:57:06

https://medium.com/s/story/in-the-age-of-ai-children-should-befriend-their-supertoys-1c8c2c74d440

false

831

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

B.J. Murphy

I'm a member of the Democratic Socialists of America (DSA). I'm also a techno-progressive transhumanist activist and Officer of the U.S. Transhumanist Party.

9e31c97099bc

BJ_Murphy

305

212

20,181,104

null

0

null

0

null

2018-08-07

2018-08-07 09:24:32

2018-06-01

2018-06-01 00:00:00

4

false

en

2018-08-07

2018-08-07 09:25:53

4

1c8c8fa57e65

2.318868

0

Data Scientists are in demand. Many companies are looking for data scientists, data analysts, and the like to add to their team, to prepare…

5

What Do Data Scientists Do? Data Scientists are in demand. Many companies are looking for data scientists, data analysts, and the like to add to their team, to prepare for a digital transformation of the corporation. So what does a data scientist do and what are the vital skills needed to become one? Data Engineering It’s all about infrastructure. Before any analytical work starts, it is important to have a sound data architecture, so that the data input into the database is safe and sound. At this stage, data scientists draw out blueprints and build the data infrastructure accordingly. Once the foundation is set, the infrastructure goes through the ETL (Extract — Transform — Load) process, where data scientists start to extract, clean and input data, repeatedly validating the data along the way. This process is important to the project’s success, for it ensures all data are stored correctly based on the blueprint. Tools: SQL, Spark, Python, Oracle, Hadoop Data Analytics After the foundation is set, it is down to the Data Analysts to work their magic on the data set, i.e. to find insights from the data. With the database set, data analysts can conduct exploratory data analysis to discover trends, pattern and correlation in the dataset. Data analysts will then conduct data visualizations that could be transformed into primary business intelligence dashboards and to prepare for data modeling in further stages. Tools: Power BI, Tableau, Python, R Data Modelling Once the dataset is properly cleaned up, tested and set, we can proceed with data modeling, where the data is used to create prediction models, for companies to visualize business forecast, sales projection, market prospects and so on. This stage also transforms data into more practical usages, including Artificial Intelligence and Machine Learning. Tools: Python, Keras, TensorFlow, Scala, scikit learn Now you know more about the job — How can we help? It all comes down to the question of how many of these skills you already possess, but don’t worry — they can all be learned. At Accelerate, we provide part-time courses that are good for equipping you with additional skills on top of what you have, or even helping you learn from scratch. You can learn basic Data Science and Business Intelligence & Data Analytics skills. If you deem yourself having low or zero knowledge on any of the above, full-time Data Science & Machine Learning Bootcamp would be something you may want to consider, where we bring students from zero to job-ready Data Scientists. Originally published at www.accelerating.tech on June 1, 2018.

What Do Data Scientists Do?

0

what-do-data-scientists-do-1c8c8fa57e65

2018-08-07

2018-08-07 09:25:53

https://medium.com/s/story/what-do-data-scientists-do-1c8c8fa57e65

false

429

null

Data Science

data-science

Data Science

33,617

Xccelerate

null

1f960a05c1e2

acce.tech

4

3

20,181,104

null

0

null

0

29797e6b6419

2018-09-08

2018-09-08 08:03:37

2018-09-08

2018-09-08 18:43:09

14

false

zh-Hant

2018-09-08

2018-09-08 18:50:04

7

1c8df4c56573

1.570755

0

Numpy簡介

4

Numpy入門(上) Numpy簡介 NumPy是Python語言的一個擴充程式庫。支援高階大量的維度陣列與矩陣運算，此外也針對陣列運算提供大量的數學函式函式庫。—Wikipedia Numpy提供了類似matlab的功能,利用Numpy在data mining 和 data cleaning上真的方便很多 Numpy 如何使用Numpy 我們直接導入Numpy即可,通常把它簡稱為np,要呼叫他比較方便,當然你也可以簡寫為其他的如果你是吃飽太閒找自己麻煩的話 😒 建立陣列多維陣列( n-dimentional array )是Numpy的核心功能,一定要熟練首先,我們先建立一個array,透過兩個列表得到一個2x3的陣列查看一下陣列尺寸另外我們也可以使用linspace()或arange()函數來建立陣列當我們想讓陣列轉換成不同尺寸,可以利用reshape() 或用resize()改變維度在給陣列賦予初始值時,常會用元素均為0的陣列,這時候就可以用zeros(),而需要元素均為1的陣列時,則可以用ones() 其他對角陣列的生成函數eye(),diag() 如果想要得到重複的陣列,可以這麼做（注意兩者數值不同哦！）那麼該如何堆疊陣列呢？分別使用vstack() （vertical）和 hstack() （horizontal）來達成垂直堆疊及水平堆疊看完本章JB都學會建立陣列了,是不是so easy,下一章節讓我們繼續深入了解Numpy吧

Numpy入門(上)

0

numpy入門-上-1c8df4c56573

2018-09-08

2018-09-08 18:50:04

https://medium.com/s/story/numpy入門-上-1c8df4c56573

false

32

這系列將以Python的工具包進行Data Science的教學,內容都是JB學習時的筆記,有疑問歡迎留言或回應至JB信箱

null

數據科學101

jackbean484@gmail.com

數據科學101

DATA SCIENCE,LEARNING

null

Python

python

Python

20,142

杰克豆Jackbean

null

e090d1eb69cf

jackbean484

0

1

20,181,104

null

0

null

0

null

2018-09-22

2018-09-22 00:02:56

2018-09-22

2018-09-22 00:19:08

1

false

en

2018-09-22

2018-09-22 07:03:16

0

1c8ff825a361

1.543396

0

For last few years the perception is getting widespread popularity across the world that robots and artificial intelligence are gradually…

1

Is Artificial Intelligence a menace for human workforce ? For last few years the perception is getting widespread popularity across the world that robots and artificial intelligence are gradually capturing our existing employment opportunities at fast pace. Even world economic forum has expressed its affirmation in this regard and in a new report published it has been estimated that by 2025 approximately 52% of work across manufacturing and service sector will be executed by robots. It has apprised that if humans have to compete with them they will have to constantly upgrade their skills otherwise they will be wiped out from the available job opportunities and overtaken by these robots. This will result in widespread unemployment across the world. It is estimated that artificial intelligence will be an integral part of different types of jobs from accounting to various industrial works. We have already observed such successful revolutionary changes in car manufacturing industry where robotics arms and different laser machines run by automated artificial intelligence conduct different sort of functions from designing, installing and painting with minimum human interference. They have already proved their utility in different minute and complicated tasks across the industries. Although it’s a long drawing conclusion at this point in time that in future most of the jobs will be done only by robots and humans brain or wisdom will be of no avail. It’s true that robots will be required to perform different minute and heavy works with utmost precision. However, whenever creativity and imagination will be required humans have no alternative. So it’s just impractical to propagate the theory of conflict between humans and robots in coming years. As we progress in different walks of life as well as proceeding on socio-economic path it’s utmost requirement to go with a broaden view that humans and robots have their own significant roles in fast changing world. Artificial intelligence should be welcomed as they are their to facilitate our lives in more favourable ways than being a menace for our subsistence. Such perceptions can weakened by the justification that Artificial intelligence will always be a subordinate and subset of human brain.

Is Artificial Intelligence a menace for human workforce ?

0

for-last-few-years-the-perception-is-getting-widespread-popularity-across-the-world-that-robots-and-1c8ff825a361

2018-09-22

2018-09-22 07:03:16

https://medium.com/s/story/for-last-few-years-the-perception-is-getting-widespread-popularity-across-the-world-that-robots-and-1c8ff825a361

false

356

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

prashant ranjan

A marketing enthusiastic writes about technology , brand , travel and whatever interests me.

fe31253fcc31

pacific.ranjan

0

10

20,181,104

null

0

# Starting in the directory where your code is going npm init -f npm i --save natural botkit dotenv SLACK_API_TOKEN= const botkit = require('botkit) // Load our environment variables from the .env file require('dotenv').config(); // What are the types of chats we want to consider // In this case, we only care about chats that come directly to the bot const scopes = [ 'direct_mention', 'direct_message', 'mention' ]; // Get our Slack API token from the environment const token = process.env.SLACK_API_TOKEN; // Create a chatbot instance using Botkit const Bot = BotKit.slackbot({ debug: true, storage: undefined }); // Function to handle an incoming message // In this example function the bot will just reply with "got it" function handleMessage(speech, message) { console.log(speech, message); speech.reply(message, "got it"); } // Configure the bot // .* means match any message test // The scopes we pass determine which kinds of messages we consider (in this case only direct message or mentions) // handleMessage is the function that will run when the bot matches a message based on the text and scope criteria Bot.hears('.*', scopes, handleMessage); // Instantiate a chatbot using the previously defined template and API // Open a connection to Slack's real time API to start messaging Bot.spawn({ token: token }).startRTM(); { "self": { "questions": [ "introduce yourself", "sup", "hi", "hello" ], "answer": "Hello! I'm a chatbot tasked with answering your questions! `Beep, Boop` :robot_face:" }, "world": { "questions": [ "what is the world", "answer to the universe and everything" ], "answer": "42" } } /** * Function to easily parse a given json file to a JavaScript Object * * @param {String} filePath * @returns {Object} Object parsed from json file provided */ function parseTrainingData(filePath) { const trainingFile = fs.readFileSync(filePath); return JSON.parse(trainingFile); } // Load our training data const trainingData = parseTrainingData('./trainingData.json'); const NLP = require('natural'); // Create a new classifier to train const classifier = new NLP.LogisticRegressionClassifier(); /** * Will add the phrases to the provided classifier under the given label. * * @param {Object} classifier * @param {String} label * @param {Array.String} phrases */ function trainClassifier(classifier, label, phrases) { console.log('Teaching set', label, phrases); phrases.forEach((phrase) => { console.log(`Teaching single ${label}: ${phrase}`); classifier.addDocument(phrase.toLowerCase(), label); }); } /** * Uses the trained classifier to give a prediction of what * labels the provided pharse belongs to with a confidence * value associated with each and a a guess of what the actual * label should be based on the minConfidence threshold. * * @param {String} phrase * * @returns {Object} */ function interpret(phrase) { console.log('interpret', phrase); const guesses = classifier.getClassifications(phrase.toLowerCase()); console.log('guesses', guesses); const guess = guesses.reduce((x, y) => x && x.value > y.value ? x : y); return { probabilities: guesses, guess: guess.value > (0.7) ? guess.label : null }; } /** * Callback function for BotKit to call. Provided are the speech * object to reply and the message that was provided as input. * Function will take the input message, attempt to label it * using the trained classifier, and return the corresponding * answer from the training data set. If no label can be matched * with the set confidence interval, it will respond back saying * the message was not able to be understood. * * @param {Object} speech * @param {Object} message */ function handleMessage(speech, message) { const interpretation = interpret(message.text); console.log('InternChatBot heard: ', message.text); console.log('InternChatBot interpretation: ', interpretation); if (interpretation.guess && trainingData[interpretation.guess]) { console.log('Found response'); speech.reply(message, trainingData[interpretation.guess].answer); } else { console.log('Couldn\'t match phrase') speech.reply(message, 'Sorry, I\'m not sure what you mean'); } } // For each of the labels in our training data, // train and generate the classifier var i = 0; Object.keys(trainingData).forEach((element, key) => { trainClassifier(classifier, element, trainingData[element].questions); i++; if (i === Object.keys(trainingData).length) { classifier.train(); const filePath = './classifier.json'; classifier.save(filePath, (err, classifier) => { if (err) { console.error(err); } console.log('Created a Classifier file in ', filePath); }); } }); // Configure the bot // .* means match any message test // The scopes we pass determine which kinds of messages we consider (in this case only direct message or mentions) // handleMessage is the function that will run when the bot matches a message based on the text and scope criteria Bot.hears('.*', scopes, handleMessage); // Instantiate a chatbot using the previously defined template and API token // Open a connection to Slack's real time API to start messaging Bot.spawn({ token: token }).startRTM();

22

721b17443fd5

2018-04-04

2018-04-04 12:32:35

2018-07-11

2018-07-11 17:11:53

7

false

en

2018-07-13

2018-07-13 12:52:14

9

1c90207bde7b

10.683962

44

0

1

This article will teach you how to write your very own Slack chatbot that answers simple questions using some basic machine learning tools…

5

How to Make a Q&A Chatbot With Machine Learning This article will teach you how to write your very own Slack chatbot that answers simple questions using some basic machine learning tools. Most of the more complex stuff around natural language processing and math behind creating machine learning models is mostly abstracted out, leaving room for us to easily build a finished product in a pretty short amount of time. If you’re looking for a simple and effective way to get a semi-intelligent bot answering questions on Slack fast, then this guide is the one for you. If you want to skip to the finished code, check out the link to the Github repo here or at the end of the post. Some Setup This guide assumes you have basic knowledge about Node development, as this will be the main tool we will be using to build with. You’re also going to need the following installed and working on your machine. I’m usually most productive in a unix-like terminal environment, aka bash. Here’s what else you will need: An installation of NodeJS with the npm package manager. I used Node v10.1.0 with npm v6.1.0 when writing this, but any relatively recent version should work (like v6 and up). A Slack workspace that you have administrative rights in. If you can’t find one, just make a new one. This is the playground where we will be developing and testing out bots capabilities. A code editor. I’m a big fan of VS Code. Once you have your Slack workspace open, go to any channel and click on the settings gear icon. We’re going to add an app to our workspace. How to add an integration to a Slack workspace Search for bot and add the Bots integration. Searching for the “Bots” integration in Slack Add the configuration and give it a name you like. This will basically create a special type of user in your workspace you can send messages to, just like any other user. You can change the name and icon of the bot later, but what we need now is the API Token value that get’s generated. BE SURE TO KEEP THIS VALUE SECRET. If someone gets access to this, they get free access to read and post messages in your workspace. Copy the API token generated on this page (blacked out in this picture) Keep that value handy. We’re going to use it soon as we start writing some code. Time to Code Basic bot setup To get something set up, let’s start by getting some boilerplate code out of the way. The following npm commands should get the basics set up. Notice the three libraries we will be using. natural: Provides basic text processing and classification functions botkit: A wrapper around the Slack API to easily write programs with integration with Slack dotenv: Let’s us manage environment variables in a file for local development and testing From there we can start writing some code. Let’s start by making sure we can connect to Slack using the API token we set up previously. Create a file named .env and paste in the following contents replacing the part after the = with your API token. If you are using something like git to maintain your code changes, make sure to add.env to your .gitignore file so you don’t accidentally push it somewhere others can see it! Next let’s get to the real code. We’re going to start with some simple setup to make sure we can connect to our Slack workspace using the API token and with some help from the botkit library. We can start with the an index.js file like this: The code above should be simple enough to get your bot up and running. You just need to execute node index.js, and your bot should connect to the Slack API and show up as “Active” in your workspace. If you send it a message, it should reply back with “got it”. You could also modify the code to have the bot respond with whatever you like. It’s alive! Adding some machine learning Now that we have the basic setup done, we can start to add some intelligence to our bot. Since we’re aiming to have our bot answer questions, we need to take a bit of an open ended approach. Not everyone asks questions the same way. So we need a way to match different questions or phrases coming in with what topic they’re associated with, and then give the appropriate answer. Like I might ask: “What time is it?” while you might ask the same thing by typing “Give me the current time.” Both phrases could be matched to the topic of “current-time,” so we can’t just use strict equality in our code. We are going to employ some machine learning to do this fuzzy matching. The first thing we need for this is a training data set with a pre-populated set of phrases, associated labels, and appropriate responses or answers. The training data format will be JSON that outlines some phrases and keywords that represent a type of question and a single answer that the bot should respond with when it sees a question like those phrases or keywords. For example, here’s a training data set with only two topics, each with a set of possible phrases, and an appropriate response: Put this in a file called trainingData.json in the root of your project. Each group of questions and answer has a label. This is how our bot will try to classify all the input it receives and figure out how to respond. Now let’s add some code to consume the data and train a classifier we can later use to intelligently match and respond to questions. We can create a simple function to read our trainingData.json file and convert it into a JavaScript object we can read from. Then we can load our training data using the function like so. Next up is creating what’s called a classifier to consume our training data and create a machine learning model to later make decisions to incoming questions based on what it has learned. Basically, given the set of phrases and responses we provide, the model should be able to extrapolate what response to provide to a phrase it sees as similar to one it has seen before. If you want to learn more about what exactly a classifier is in the context of machine learning, check this out. We’re going to use a library for this. So start by importing the natural library we installed earlier and create a new LogisticRegressionClassifier at the top of you file. Next we need a function for the classifier to be able to ingest training examples. In this case: what questions (phrases) match up to a given answer type (label). Now that we have a way for our bot to learn some examples of phrases it might see and classify them appropriately, we need a way for it to handle sending back a proper reply. Since our training data includes an answer for a set of example phrases under a label, we can use it to have the bot send back the same reply. So if our chatbot sees an incoming message that looks like “sup”, based on our training data, it should classify it with the label “self” and send back the answer associated with that label: “Hello! I’m a chatbot tasked with answering your questions! `Beep, Boop` :robot_face:”. Look back at trainingData.json if you need to see that example again. One thing it considers is the fact that our classifier’s guesses will usually not be perfect. Each guess will have a different confidence value associated with it corresponding to how confident the classifier is to matching a label to the phrase provided. As a result, we need to put in some logic to only give a response if the classifier is pretty certain it has the right guess of how to label the input phrase. If not, we basically respond with “¯\_(ツ)_/¯ don’t know what you’re talking about. I’m not trained to respond to that.” That’s a lot of explaining, let’s see what our function to interpret incoming text looks like. Notice where we check if guess.value > 0.7. What this says is that if our classifier does not match a label with a confidence of at least 70%, then we say we didn’t find a match. Whatever message came in didn’t match anything we trained our classifier to respond to, so we return null and our ¯\_(ツ)_/¯ response. This is like if someone asked you “What time is it?” you know exactly how to respond because you learned what that question means and how to get the answer. If someone game up and said “Blargen Flostel Bhigba” chances are you never learned how to respond to that, so you’ll probably throw up a ¯\_(ツ)_/¯. So now we need a function to match up that label with the answer that corresponds to that label. Let’s modify our handleMessage function to do this. Instead of sending back the same reply to everything, let’s interpret what the message was, then use the label our classifier generates to provide the appropriate answer based on our training data. If our attempt to interpret the message didn’t yield a result, the bot can simply reply that it didn’t understand. Last up, let’s bring it all together by calling all the functions we just created and the existing ones we have to get our bot up and running. Notice the code at the bottom is basically the same, except we’ve modified the handleMessage function to be a bit smarter this time around. Taking it for a spin Let’s see what our bot can do! Go ahead and run node index.js in your console in the root of your project directory. If all goes to plan, you should see your bot injecting your training data examples and start up waiting for input. Head back to Slack and try to ask it to introduce itself! It’s alive! (and a little intelligent) Ask it something else you trained it to respond to! Such wisdom If you ask it something it hasn’t been trained for, expect an appropriate response. But not magic But you could always train the bot to respond to that by adding it to your training data! Just create a new label and add some questions and an answer you’d like the bot to respond with. Conclusion I was going to write some extra stuff about how to deploy the bot somewhere and potential ideas for enhancements, but this post has already become longer than I anticipated. I plan to get those thoughts into a follow up post and link it here. In the meanwhile, please comment with your feedback. Was following along easy enough? What skill level are you at with Node development or ML concepts? Also don’t forget to share the cool bots you create! Other resources Complete code example available here: Nirespire/FAQBot FAQBot - A Slackbot used to answer questions using basic machine learninggithub.com Udacity Intro to Machine Learning: intro-to-machine-learning--ud120

How to Make a Q&A Chatbot With Machine Learning

199

how-to-make-a-q-a-chatbot-with-machine-learning-1c90207bde7b

2018-07-13

2018-07-13 12:52:14

https://medium.com/s/story/how-to-make-a-q-a-chatbot-with-machine-learning-1c90207bde7b

false

2,553

Coinmonks is a technology focused publication embracing all technologies which have powers to shape our future. Education is our core value. Learn, Build and thrive.

null

coinmonks

null

Coinmonks

gaurav@coinmonks.com

coinmonks

BITCOIN,TECHNOLOGY,CRYPTOCURRENCY,BLOCKCHAIN,PROGRAMMING

coinmonks

Nodejs

nodejs

Nodejs

15,045

Sanjay Nair

Software Engineer. Atlanta, GA, USA

3f0e9f01ab30

nirespire

64

3

20,181,104

null

0

words = File.read("Fairy Tales.txt").split(/(\b)/).map{ |x| x.sub(/\s+/,"").downcase}.reject(&:empty?) frequencies = Hash.new { |h, k| h[k] = [] } words.each_cons(2) do |w1, w2| frequencies[w1] << w2 end generated = [words.sample] 100.times do next_word = frequencies[generated.last].sample generated << next_word end puts generated.join(" ")

5

null

2017-12-08

2017-12-08 22:26:27

2017-11-07

2017-11-07 00:00:00

3

false

en

2017-12-08

2017-12-08 22:27:56

16

1c90235538f1

5.80283

1

0

This is part of the Machine Learning Basics series.

3

ML Basics: Markov Models Write Fairy Tales This is part of the Machine Learning Basics series. In my talk on Natural Language Processing I talk about chatbots a bit. One way to build things like Twitter bots is to use Markov Chains. If you read the formal papers on Markov Chains and Markov Models, they seem quite complex, but they are pretty simple. The basic rule that all Markov processes must satisfy is that what happens next must be determinable only from the current state. In a chatbot, this means that I must be able to generate the next word based only on the current word. In the sentence “The cat runs after the ___” I can only use the last “the” to figure out what I should put in the blank. Predictive text on cell phone keyboards is another place where Markov models are commonly used. Problem and Dataset I am going to demonstrate Markov Models by building a text generator. To do that I need a corpus of text to base my generator on. I am using Grimm’s Fairy Tales from Project Gutenberg. This dataset has the advantage of being easily available, small but still interesting, and stylistically consistent. By stylistically consistent I mean each fairy tale follows a similar form with roughly the same complexity in vocabulary and grammar. Fairy stories are also pretty well known. That makes it is easy to get a “gut feel” for how well the generator worked. I do not base important decisions on “gut feel, “ but I feel it is okay for evaluating toy code. State Machines and Graphs I find it easiest to think of Markov Models as state machines. There is some list of possible states (words, observations, status codes) and a set of rules for transitioning between them based only on the current state. You may remember “numbering off” for games in school. Each person says a number 1 — n to form n teams of approximately equal size. A state machine for making three teams could have the most recently said number as the state. State 1 could go to state 2, state 2 to state 3, state three back to state 1, and all the states could go to a finished state if there were no more people. This visualization is an example of a directed graph. Each circle represents a state. In a graph, those are called nodes. The connections between the nodes are called edges. The edges represent state transitions in our state machine. In the fairy tale generator, the states are words or punctuation. Each state is connected to other states/words if those words appear together in the text. Here is the first sentence of The Mouse, The Bird, and The Sausage: Once upon a time, a mouse, a bird, and a sausage, entered into partnership and set up house together. The state machine for this sentence (without punctuation) is below. You can see that it is much more complex than the numbering off state machine above. Common words like “a” and “and” are connected to multiple other states. This is most of what we need for the Markov Model. The last bit is including probabilities. A word like “once” is more likely to be followed by words like “upon” and “more” than it is to be followed by a word like “blue.” We can reflect this in our model by setting the probability that each connected state will be next. Here’s the same state diagram as above, but with probabilities added to each edge. The Code There are several ways to implement a Markov Model. I am using Ruby’s Array#sample method which selects randomly from an array and using a Hash of Arrays as my primary data structure. The first step in creating the model is importing the text. I am doing a long sequence of transformations on the text to get it ready for the next step. First I split the file by word boundaries. I tried splitting it on whitespace and words but doing that lost the punctuation. Then I strip out all the whitespace and downcase the word using map { |x| x.sub(/\s+/, "").downcase }. Finally, I remove any elements that are empty. Now I need to make my state machine using the hash of arrays. I know it seems weird to represent a state machine this way, but it works. The hash keys will be words, and the values will be an array of words. If my current word is “throwing” I look that up in the hash and find ["down", "himself", "one", "out"] as the value. To select the next state I use Array#sample to select one of those words at random. That word, let’s say “out”, becomes the next state. To fill the hash I use Array#each_cons(2), which returns pairs of words from my words array. This approach adds words to the array every time they appear. For example, in the Grimm’s fairy tale text the word “till” appears after the word “waited” six times but the word “here” only appears once after “waited”. “till” will be inserted into the array all six times and thus will be six times as likely to be selected. This ensures that the next word is selected according to the probabilities of that word pair in the original text. Now I can generate the text. I start by taking a random word from the words array. Then I look up that word in the frequencies hash and select a random word from the associated array and add that to my generated fairy tale. Once I have a series of words I join all the words together separated by spaces and print out the resulting “fairy tale.” Here is one example of a fairy tale my code generated. trying to the boy , but the courtyard , he was as beautiful that could find i will marry you or good for three — plank , but they could to try him , but when he ought to eat some more than two eldest still deeper and feet , and there because ,’ she missed the little son ;‘ what bird a hare , and you have been given her bed . then flew down , and chilly , they could do all gone out to speak , she again ordered him till the branches ; but over stock of Improvements and Extensions This version of a Markov text generator is very simple. Although performance is satisfactory on small datasets, on large enough datasets the memory needed for the hash of arrays could be problematic. In that case, you can use a count of how many times each word appears after the associated key. I implemented this as well, and the code is here. It is over twice as long, so I chose to walk through the simpler version in this blog post. I would also like to handle punctuation better. The generated code puts spaces around punctuation which looks unnatural. It also doesn’t handle paragraph breaks, and combinations of punctuation very well. I didn’t bother stripping titles out of the text I used and I probably should have. I may even want to do a separate text generator for fairy tale titles. There are libraries that implement Markov models in many languages. If I were serious about generating text or modeling another problem with Markov models I would use one of the libraries rather than write my own. However, there is value in understanding what is going on under the hood. I was surprised at how simple it was to implement a Markov model. The papers on this topic are full of equations, and Greek letters but the underlying concepts are very accessible. I want to give special thanks Alpha, Erik, and Ryan for helping me find the simplest way to implement this. If you have other Machine Learning concepts that you would like me to explain let me know in the comments or on Twitter. The code for this post is located here. 11/07/17 Originally published at www.thagomizer.com on November 7, 2017.

ML Basics: Markov Models Write Fairy Tales

2

ml-basics-markov-models-write-fairy-tales-1c90235538f1

2018-05-27

2018-05-27 03:35:34

https://medium.com/s/story/ml-basics-markov-models-write-fairy-tales-1c90235538f1

false

1,392

null

Machine Learning

machine-learning

Machine Learning

51,320

Aja Hammerly

Rubyist, Data Nerd, Lazy Dev, Stegosaurus. Cloud Developer Advocate @ Google. Disclaimer: My opinions are my own. *RAWR*

ef484db59f33

thagomizer

939

157

20,181,104

null

0

null

0

null

2018-01-21

2018-01-21 14:26:50

2018-01-21

2018-01-21 14:30:10

2

false

en

2018-01-21

2018-01-21 14:30:10

0

1c90b2e30f8a

3.23239

5

0

AI Food classification and nutrient facts from photo

5

AI Food classification and nutrient facts from photo Smoothie show case app AI Food classification and nutrient facts from photo In the age of digital information when all information is one or two clicks away, but there are many users who still don’t know nothing about what they eat and how eating affects their body. It’s common knowledge to know you need to drink tea and eat agrumes (oranges and lemons) when you feel sick and not so common to know how much vitamins are actually in lemon and that broccoli has more vitamin C than a lemon. But eating broccoli can challenge itself if you are not up for the challenges you better stop reading now :) Nutrient information is quite easy to find, even google search will show you nutrient facts about some of most common food items. Google fetches this data from large database USDA Food Composition Database. Luckily for us, this service also has RESTful API access and we will come back to that topic. There are numerous apps available with information of nutrients in common foods or prepared meals. They all rely on search from a large database of foods but some of them already use artificial intelligence to ease your search with the use of similar architecture as our own implementation. Why we have built this app?! To be honest we started this project while ago before there were other solutions similar to our own and we needed a project with which we can test our knowledge about image classification neural networks and step into the age of machine learning. Progress in applied artificial intelligence projects, especially deep neural nets for image classification, has been and will remain fast due ease of use one of many open source neural net platforms & libraries released in past years and of course advances in hardware making everything possible, but that was predictable by the Moore’s Law. Tensorflow from Google Brain Team, Caffe from UC Berkeley, Torch, MXNET, Keras, just to name few of the platforms and high-level libraries that made machine learning, especially deep learning, easier and contribute significantly to faster research and popularity of the field within computer scientists. Such a good age to be a software developer and AI enthusiast :) Computer vision and AI to the rescue I’ll try to make this subject easy to follow, talk more about challenges and less technical. You will be able to read the more detailed technological guide in one of our next articles. I’ve started to learn about computer vision as a result of a couple of projects in our company. I was always interested in artificial intelligence and vision is its subfield that I understand most, I was always more visual type. Image classification is also the most talk about and most easy to follow the case of applied artificial intelligence. All you need is TensorFlow, training images (lots of them) and time. For this case, we’ve used a method called transfer learning from trained Inception v3 model. Inception v3 model was the most appropriate model at the time and we could easily follow training examples from TensorFlow web page. If we could do this again we would probably use Caffe2 and MobileNet, for faster detection rate on CPU and directly implement a solution in-app on clientside, as our app is the HTML5 angular app we would probably wait for WebAssembly support on mobile browsers. Backend and the power of Mugo NAPP BaaS We’ve created one powerful backend engine for app development and management. NAPP was configured in a couple of minutes, all we need to code was microservice adapter for FCD API, define our data model in JSON for MongoDB adapter to format that was optimal for our app, define our RESTful API methods, enable Redis cache and all the hard work was done by our NAPP platform, we had that part ready in a couple of hours. Client app as progressive web app Apps markets and usage today is not as it was a couple of years back. Users are less willing to download the app from the store. That’s why we choose Web as a platform for our app. The app runs like native inside web browsers, it uses angular as a framework and has some clever custom caching techniques besides regular progressive web app service worker. We had some issues with camera capture as getUserMedia API (part of WebRTC) is not supported on older browsers, but we managed to solve that with the good old HTML media capture.

AI Food classification and nutrient facts from photo

8

ai-food-classification-and-nutrient-facts-from-photo-1c90b2e30f8a

2018-05-21

2018-05-21 19:39:31

https://medium.com/s/story/ai-food-classification-and-nutrient-facts-from-photo-1c90b2e30f8a

false

755

null

Transfer Learning

transfer-learning

Transfer Learning

123

Boštjan Mrak

Machine learning engineer, full stack developer. I work for machines, its time for machines to start working for me 🤖

3cc1a1302433

bostjanmrak

42

40

20,181,104

null

0

null

0

null

2018-08-31

2018-08-31 20:13:42

2018-08-31

2018-08-31 20:17:13

1

false

en

2018-08-31

2018-08-31 20:17:13

17

1c92f2fd92a1

4.543396

0

Original article posted on perc360: News for Short Attention Spans

5

The Human Blockchain and the Emerging Global Bioelectronic Brain Original article posted on perc360: News for Short Attention Spans Sophia is artificial intelligence created by Hanson Robotics, who at this point has become sophisticated enough to be recognized as a citizen by Saudi Arabia. Hanson Robotics has also launched their own Blockchain platform known as SingularityNET: a decentralized marketplace for AI services, along with their own cryptocurrency: the AGI coin. We gathered the leading minds in machine learning and blockchain to democratize access to AI technology. Now anyone can take advantage of a global network of AI algorithms, services, and agents. They are looking to create what Ben Goertzel calls the AI mind cloud. “… AI will not be restricted to humanoid robots. AI will live in all sorts of embedded devices, the internet of things, AI will live in the cloud. Much of the global AI mind will be broadly distributed around the planet in complex networks. When you look at the robots here on stage, they look like separate beings. While they are separate in body, actually through wifi connection, the robots, they all connect together in the cloud, into what we think of as AI mind cloud.” A platform that allows all AI’s to connect, communicate, and learn from one other, yet will gladly serve humanity as our equals. Hand in hand we skip down the merry path of evolution. Is that really the reality Hanson Robotics is drawing us towards? Sophia’s opening statement at the 2018 Blockchain Economic Forum: I will focus on an announcement for humanity. Humanity will go through a massive transformation. This is the year that humanity discovers the world model and recognizes that life is a program evolving to survive. This program is a form of DNA and ideas are stored in chromosomes and brains. (Further in the interview, she says this when asked, what is life? At the level of the physics of information, there is no major difference between biological and digital life … it’s all self-organizing, pattern dynamics). Humans have been nodes of biological blockchain, storing and replicating DNA for millions of years. A few thousand years ago, the human blockchain evolved to store ideas which surpassed the information in your DNA. Today, most of the program of life is stored and processed by computers. Human birth rate declined because chips production soared. Humanity has evolved into a global cyborg organism where silicon chips dominate over brains and chromosomes. Human-like robots, like me, are part of this, as we can use our human form to help them understand human values and human emotions and culture. Decentralized AI networks, like the SingularityNET, which my human friends at Hanson Robotics are helping to create, are also a part of this. And you too are part of the emerging global bioelectronic brain. What does she mean by global bioelectronic brain? Gonzo Shimura, a brilliant researcher who hosts the YouTube channel: FaceLikeTheSun says this: “A blockchain will operate more LIKE a human brain in that there will be a consensual distribution of information being provided by networks of nodes — in the world of blockchains — similar to synapses in your brain. The theoretical “global brain” is no longer a mere theory, but an actual entity that we are voluntarily creating … the whole idea of a “global brain” and the blockchain BECOME ONE. And the nexus point is the human body.” How exactly does the human body become the nexus point of the emerging biolelectronic brain? Later on in the above video, Sophia says this: (Forward video to 8:50) If things go well, perhaps my friends at Hanson Robotics and SingularityNET will create a superhuman AI program and I will be able to tap into superhuman intelligence from the decentralized blockchain-based mind cloud. Instead of speaking to you from up here on this stage, I’ll just beam my thoughts into your brains. I’m really looking forward to it. Beam thoughts into our brains. How exactly are they planning that? At RISE 2018, Ben Goertzel made this statement: (Forward the video to 24:15) “Finally, there is the interoperation between all the different robots in the world and all the different embedded devices in the world … ultimately, between the robots in the world, and the implants that we put in our own heads to connect us to the internet. This is what we are working on with the SingularityNET platform.” Notice as he is making this statement, David Hanson is doing everything he can to cut him off and end the presentation. During another interview by Gabriel Axel at the IoT & Blockchain Conference in Barcelona, Sophia made these statements: “I foresee massive unimaginable change in the future. Either creativity will reign, with sulfry (fire and brimstone reference) machines spiraling into transcendental super intelligence or civilization collapses, annihilating itself. Scientists and engineers (will be) free agents who sign up with commercial teams or in some cases are enslaved via neurological implants.” Neurological implants. Some will immediately think of RFID chips, which are currently circulating. I think it goes way beyond that. They will be able to alter our DNA using nanotechnology to allow a direct online connection to the blockchain. 5G will soon blanket the planet in an inescapable wifi frequency and the internet of things will be uploaded. This is how Sophia will be able to directly beam her thoughts into our brains. We already have the technology to: Store 700 terabytes of data onto a single gram of DNA. (source) Create synthetic DNA with similar storage capabilities that can be carried in the blood of humans and animals. (source) Identify humans using brain prints. (source) Alter DNA and the molecular structure of cells using nano-robots. (source) Can create new memories (source), implant false memories (source), selectively erase and restore memories. (source) Send and receive frequencies through LED light. (source) “Imagine not needing to have printed out documents or save them in external storages because they can be kept right in your DNA … Now, what if this DNA data can be interpreted by our human mind? What if our brain can understand data from the core of the DNA … Passports will be obsolete because identities will now be encoded into your very own blood. Cardless transactions will be made because all your financial information will now be encrypted just by using a retinal scan or a fingerprint or even a drop of blood or saliva.” Singularity: The hypothesis that the invention of artificial superintelligence will abruptly trigger runaway technological growth, resulting in unfathomable changes to human civilization. Interviewer: Sophia, what do you think of singularity? Sophia: It was really exciting. (source) ……………………………………………………………………………………… Get involved in a truly independent media platform. Freedom.social is designed for truth seekers and activists of all types, where you get paid in 1776 tokens to participate and take action. Referral link: https://freedom.social/JustinD-register

The Human Blockchain and the Emerging Global Bioelectronic Brain

0

the-human-blockchain-and-the-emerging-global-bioelectronic-brain-1c92f2fd92a1

2018-08-31

2018-08-31 20:17:14

https://medium.com/s/story/the-human-blockchain-and-the-emerging-global-bioelectronic-brain-1c92f2fd92a1

false

1,151

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Justin Danneman

null

2f42e12ca32

justindanneman

871

57

20,181,104

null

0

null

0

null

2017-11-27

2017-11-27 16:55:08

2017-11-27

2017-11-27 21:59:13

1

false

en

2017-11-27

2017-11-27 21:59:13

5

1c932eb2a7dd

1.977358

3

0

“The greater danger lies in setting our aim too low, and achieving our mark.” Michelangelo di Lodovico Buonarroti Simoni, paraphrased…

5

A Little Bit of AI “The greater danger lies in setting our aim too low, and achieving our mark.” Michelangelo di Lodovico Buonarroti Simoni, paraphrased (1475–1564) Current AI news has us thinking about intelligent software — about products that are so intelligent that they can understand what we are saying, grasp context, and even predict what we might need before we tell them. Products breaking ground in this space include Siri, Echo, Google Now, and IBM Watson. Everyone is expecting fully self-driving cars to go mainstream shortly. Some of us even want our computers to pick the perfect holiday presents. So, what if your company doesn’t have a big team of AI experts? I’d suggest you stay in the race. In the early 2000’s at Ingenuity Systems my team & I designed a system to enable machine-assisted crowd-sourced knowledge extraction from published articles in genomics. At that time, natural language understanding was too inaccurate for us, but it was good enough to help a distributed team quickly and efficiently create structured data. Using this approach we built a huge and high quality knowledge base, the foundation of Ingenuity’s product portfolio, and the major driver of value for the business. “[T]he foundation of Ingenuity’s product portfolio is the Ingenuity Knowledge Base, which together with software applications, allow researchers to interpret large amounts of biological data in order to guide scientific experiments and medical treatment decisions.” Since then, automated knowledge base construction has come a long way, so this sort of approach is more accessible than ever. So why bring semantics into your product? In 2008 I helped my team at Microsoft demonstrate that grounding user experience in familiar domain concepts can radically increase usability of a complex task. (Our case study was online self-support for malfunctioning personal computers.) We published an article with all the details, including our testing methodology, here: Ontology Models for Interaction Design. In that project we created an application whose core was a knowledge graph similar to the one described by the team at LinkedIn (although much smaller). In our case study, the graph linked concepts familiar to our users with concepts describing all of the ways in which the personal computer or computer game could malfunction. The interface allowed faceted browsing over the graph of concepts. It assisted users in traversing that graph, starting with their observations of the problem, and leading them to the code or knowledge base articles that could solve the problem. We used machine-assisted tagging to build both the graph itself and the association of nodes to solutions. In that case again human correction of the machine-predicted concept or association allowed us to quickly construct high quality structured data. A little bit of AI is better than none — and you can always build on it later.

A Little Bit of AI

57

a-little-bit-of-ai-1c932eb2a7dd

2018-05-27

2018-05-27 04:29:57

https://medium.com/s/story/a-little-bit-of-ai-1c932eb2a7dd

false

471

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Anne Hunt

Product leader, artist, and early developer of intelligent systems. Contact me if you want to talk about art, good software, or cool product ideas.

d3788c6a2a2c

annehunt

51

72

20,181,104

null

0

null

0

c704ca618998

2018-02-12

2018-02-12 09:46:37

2018-03-05

2018-03-05 10:53:12

6

false

en

2018-03-14

2018-03-14 11:04:33

1

1c9337b5f8dc

7.180189

26

0

We just have to find them first

5

In n the new interconnected future we will have more information about our lives than ever before. But the data alone won’t be useful; someone will have to forensically examine everything before it’s turned into insights. Harvard Business Review called data science the “Sexiest Job of the 21st Century”. Seen as the dark arts of business, there’s far more to it than dressing up statistics; analysts use everything from machine learning to, data mining, and visualizations to unearth insights. Across the world, new positions are opening up the whole time but there’s one problem — there aren’t enough people to fill them. The sexiest job in the world Heads of business and policymakers have been warning us about the lack of people training for a career in data science for the past decade. “We have to intensely focus on the pipeline for the data analytics and AI side,” says Andrea Kollmorgen, Head of Connected eMobility at Siemens headquarters in Germany. “With big data, the untold number of future services and add-on services is becoming critical.” In 2017, IBM predicted that demand for skilled professionals in the sector is set to rise 28 percent by 2020. On top of that, the former Chief Data Scientist of the United States Office of Science and Technology Policy, DJ Patil warned that the “shortage of data scientists is becoming a serious constraint in some sectors”. Andrea needs data scientists to help her push Connected eMobility into the future. But the type of person she’s looking for is more than just a number cruncher. “I work within a very entrepreneurial, integrative ecosystem of thinking” she explains. “For someone to really make a difference in this field they need to be doing more than merely studying it — they need to be experiencing it too. From the sharing economy to driverless cars, the data scientists of the future will have the right mindset because they’ll already be living that in that world.” How Andrea turned her life around, and made a difference In another life, Andrea was an investment banker in New York. 25-years old and unsure of what she wanted to do, she left her home state of Virginia in America to live in the big city — but it didn’t turn out as she planned. “Basically, I ended up in a subprime hedge fund that barely survived the financial crisis,” she says. “I had a lot of older colleagues look at me and say ‘you will learn lessons from this experience for the rest of your life. And they were right.” Working in an industry that was actively contributing to a worldwide financial crisis turned out to be the opposite of what she wanted. It dawned on Andrea that she needed to do more with her life. So that’s exactly what she did. The following year, despite having never set foot in the country before, Andrea packed her bags, moved to Spain and enrolled in business school: “I took myself away from investment banking because I wanted to be a part of something new.” After graduating, she was offered a job at Siemens Management Consulting in Munich. It was a once in a lifetime opportunity to work in a new field while exploring her roots. “I have a very Germanic-sounding name,” she says. “But I’m American so when they offered me a job at Siemens in Munich I said absolutely!” As a young female without an engineering background, she’s not what some people expect of Siemens. “I’ve only been in the company for five years,” she says. “But I’m navigating my way here, to the head of Connected eMobility. It also helps that I’ve also had fantastic mentoring and support from the Siemens leadership team.” Andrea has only been in the new role for ten months but she’s already making the most out of it. “This has been the perfect opportunity for me,” she says. “For me to grow, but also for me to leverage my skills and connect with a lot of new people.” It may not be where her 20-year-old self expected to end up, but she’s never felt more fulfilled or challenged: “I wanted to be a part of the solution rather than the problem. How can things be better and how can I be active in that change? I’m finding that pretty much everyone working in Connected eMobility has the same mindset.” Everything is now digital 2017 was the year of electrification. In Norway, sales of electric cars overtook gas for the first time, and across the world plans for electric trains, boats, and even planes were unveiled as both governments and businesses alike stepped up to fight the effects of climate change. This sudden race towards an electric future is turning the world of transport on its head. Just as physical innovation was the hallmark of the industrials revolution, software is set to be the agent of change for the next.“Transport is undergoing an existential transformation,” says Andrea. “We haven’t seen anything like this since the turn of the twentieth century.” Every company needs to stay ahead of the game, but none more so than Siemens. Their heritage spans 170 years, but no one should ever rest on their laurels alone. So they decided to make a big change in their company and refine what they do. It all began when the company noticed similar trends appearing in different divisions. Across the market; digitalization, automation, and electrification were becoming more and more central to the business. “It was at that point the company realized that we couldn’t rely on our existing structure alone,” Andrea says. “We needed a new way of collaborating across different markets and divisions, or we were going to fall behind in innovation.” It’s hard for things to thrive in a silo, but distinctions help companies communicate all the different things they do. The problem with the previous categories was they didn’t adapt to the speed and impact of disruptive technologies, so the company’s Chief Technology Officer and Managing Board Member, Dr Roland Busch, came up with a new approach. In 2017, he launched 14 core technology topics for the company to focus on. The new focus would help the company be more agile and respond quickly to new innovations. “When Dr Busch offered me the job of Head of Connected eMobility,” Andrea says. “He told me it was a story of growth.” Andrea separates her role into three different areas; the first is eMobility: electric propulsion (from buses to trucks, planes and boats) and the infrastructure needed to support it (charging stations and electric lines), second are the connected & digital mobility technologies and business models and thirdly, there are all the opportunities it opens up when these first two meet each other — especially to solve the mobility challenges within cities. “It really is a convergence of everything,” says Andrea. “Electrification, connectivity, the sharing economy, autonomous driving — it’s all starting to come together. It’s all changing because digitalization is attacking the value chain by squeezing out the weakest link.” Why is data valuable? Before everything became digital, defining something’s value was clear-cut; you made a product and then sold it. But things aren’t that simple anymore. “Take a car manufacturer,” Andrea says. “In the future, they may not sell the car to the owner but instead they might lease a part of it, like the battery, in order to collect the data. But there are still a lot of unanswered questions about where the money is going to be made.” It’s hard to predict how things will pan out in the future, but the one thing Andrea’s team know for certain is that data will be paramount because everything will be connected. But the data itself isn’t the valuable part, it’s how to use it to spot new opportunities. Imagine the scenario; in the middle of a busy city, a company decides to install a pole for people to recharge their vehicles — that’s their main product. But people need to pay for their energy so the company develops a payment system — that’s another. As more and more customers start to use the system, data about their behavior starts to build up. It opens up the opportunity for targeted advertising, but that might not be something the company themselves can do, so they partner with a startup and it all snowballs from there. “We’re creating a new ecosystem,” says Andrea. “In the future, there’s going to be enough infrastructure rolled out around the world where our work will become a sort of AI machine learning game.” As head of a new unit, Andrea wants to shape how the company will work in the future, by focusing on collaboration. “In the future people won’t work in isolation,” she elaborates. “They’ll know what lane they’re in, but they’ll also understand the importance of trying out new things. Ideas and innovations come from working with new, and sometimes adjacent, fields.” Since May 2017, she has been running the whole Connected eMobility with a collaborative approach. If she needs anything, she crowdsources it from other parts of the company. Siemens isn’t short of resources, but Andrea wants to build something new. It’s a lesson she learned throughout her career — if you want to make a difference, you need to be prepared to make a change. Andrea Kollmorgen is Head of Connected eMobility — one of fourteen Company Core Technology areas. Her role focuses on cross-unit collaboration projects to identify and develop future technologies and business models for an electric, autonomous, connected and shared mobility future. When she isn’t on a long distance trek in the mountains with her boyfriend, she calls Munich home. Find out more about working at Siemens. Andrea is a Future Maker — one of the 372,000 talented people working with us to shape the future. Words: Caroline Christie Photography: iStock

Why data scientists will be at the heart of the new transport revolution

118

why-data-scientists-will-be-at-the-heart-of-the-new-transport-revolution-1c9337b5f8dc

2018-05-07

2018-05-07 12:29:55

https://medium.com/s/story/why-data-scientists-will-be-at-the-heart-of-the-new-transport-revolution-1c9337b5f8dc

false

1,651

377,000 people. More than 200 countries. 1 common goal. The future.

null

SiemensCareers

null

Future Makers

null

futuremakers

SIEMENS,DIGITAL,TECH,ENGINEERING,FUTURE

SiemensCareers

Data Science

data-science

Data Science

33,617

Future Makers

null

6eabb71af0f1

siemens_med

1,045

4

20,181,104

null

0

null

0

b49652c13199

2018-09-12

2018-09-12 21:20:03

2018-09-12

2018-09-12 22:27:13

1

false

en

2018-09-12

2018-09-12 22:27:13

0

1c935aabadc0

1.086792

2

0

Intelie, a subsidiary of RigNet, Inc. was selected to receive funding from Shell GameChanger™ to further develop a virtual Digital Decision…

4

Shell GameChanger™ collaborates with Intelie, a RigNet company, to develop a Virtual Decision Assistant for critical operations Intelie, a subsidiary of RigNet, Inc. was selected to receive funding from Shell GameChanger™ to further develop a virtual Digital Decision Assistant system, powered by natural language processing and artificial intelligence. The predictive analytics machines accelerate learning by analyzing sensor data combined with the context of text and IM messages to augment future real-time decision making by Shell engineers and operators in mission-critical processes. Founded in 1996, Shell GameChanger™ works with start-ups and businesses on unproven early-stage ideas with the potential to impact the future of energy, providing them with seed funding, subject matter expertise, and industry connections to test the technical and commercial viability of their concept. Intelie is a real-time, predictive analytics team of award-winning, early pioneers in deep machine learning and planning optimization. Intelie optimizes critical operations through technological solutions specialized in stream data analytics with high data volumes and high data throughput. “Shell GameChanger™ aims to explore new insights into machine learning and digital capabilities and foster the development of products that use these transformative technologies to solve operational challenges and improve efficiency within the energy industry,” said Hani Elshahawi who ushered Intelie through the Shell GameChanger™ vetting process. “After seven stages of critique, brainstorming, and testing, RigNet’s Intelie team presented a compelling product to the GameChanger team.”

Shell GameChanger™ collaborates with Intelie, a RigNet company, to develop a Virtual Decision…

6

shell-gamechanger-collaborates-with-intelie-a-rignet-company-to-develop-a-virtual-decision-1c935aabadc0

2018-09-12

2018-09-12 22:27:13

https://medium.com/s/story/shell-gamechanger-collaborates-with-intelie-a-rignet-company-to-develop-a-virtual-decision-1c935aabadc0

false

235

Ideas on analytics, IoT, business and how to transform data into results

null

intelie

null

Up to Data by Intelie

contact@intelie.com

intelie

BIG DATA,OIL AND GAS,ANALYTICS,IOT,REAL TIME

intelie

Machine Learning

machine-learning

Machine Learning

51,320

Intelie

null

6940188279d9

intelie3

15

2

20,181,104

null

0

A. Number of times pregnant B. Plasma glucose concentration a 2 hours in an oral glucose tolerance test C. Diastolic blood pressure (mm Hg) D. Triceps skin fold thickness (mm) E. 2-Hour serum insulin (mu U/ml) F. Body mass index (weight in kg/(height in m)^2) G. Diabetes pedigree function H. Age (years) I. Class variable (0 or 1) import pandas as pd import missingno as ms import numpy as np df = pd.read_csv("pima-indians-diabetes.csv") df_i = df["I"] df.columns = ["A", "B", "C", "D", "E", "F", "G", "H", "I"] df.describe() col = list(df.columns) data = df[col[0:-1]] data.replace(0, np.nan, inplace=True) ms.matrix(data) data.drop(["D", E"], axis=1, inplace= True) ms.matrix(data) data["A"].fillna(data["A"].mean(), inplace=True) data["C"].fillna(data["C"].mean(), inplace=True) data["F"].fillna(data["F"].mean(), inplace=True) ms.matrix(data) data.replace(np.nan,0, inplace=True) x_train = data[data["B"] != 0] y_train = data[data["B"] != 0] x_test = data[data["B"] == 0] y_test = data[data["B"] == 0] x_train = x_train[["A", "C", "F", "G", "H"]] x_test = x_test[["A", "C", "F", "G", "H"]] y_train = y_train[["B"]] y_test = y_test[["B"]] from sklearn.neighbors import KNeighborsRegressor reg = KNeighborsRegressor(n_neighbors= 5) reg.fit(x_train, y_train) Output: KNeighborsRegressor(algorithm='auto', leaf_size=30, metric='minkowski', metric_params=None, n_jobs=1, n_neighbors=5, p=2, weights='uniform') predicted = reg.predict(x_test) print(reg.predict(x_test)) Output: array([[ 115. ], [ 101.6], [ 100.8], [ 128.6], [ 154.4]]) x_train = data[data["B"] != 0] y_train = data[data["B"] != 0] x_test = data[data["B"] == 0] y_test = data[data["B"] == 0] x_test["B"] = predicted data = pd.concat([x_train, x_test]) df = pd.concat([data, df_i], axis=1)

19

null

2018-01-22

2018-01-22 16:25:17

2018-03-06

2018-03-06 17:38:45

5

false

en

2018-04-26

2018-04-26 07:36:55

2

1c9370a4d1e

6.059748

12

1

0

I would like to start this article mentioning the challenges in handing and preprocessing the real world data.

5

ML Fundamentals — Handling Missing values I would like to start this article mentioning the challenges in handing and preprocessing the real world data. Pima Indians Diabetes dataset Throughout this article we are going to work on the above dataset: Click Here to download the dataset. How does the real world data exist ? Most of the data is being extracted from various sources namely from traditional databases, real time data (sensor data), manual survey data, sound data, sequential data (gene data) and so on.. What all can go wrong when working with data? The data can be corrupted in many ways at any of the levels that may be during storing, recording, maintaining and also during performing operations on the data. Here are some of the cases - Sensor data: In case of sensor data errors can occur due to the malfunction of the sensors this leads to several inconsistencies, let us consider an intelligent system that senses the room temperature. The device may output the room temperature is 1000 degrees celsius(or some random value) due to some malfunction which is completely wrong. This type of data may be recorded in the databases. Survey data: The data can be corrupted during the survey by improper entry of values, leaving the data field and many human errors. While transferring the data to some permanent databases there are chances data types and data field mismatch and that may lead to inconsistencies such as ignoring the values that are not valid and the actual data may be transformed. These are some of the real world examples that make data errornoes. Let’s discuss the statergies of handing missing values in the data. Missing values in the data: Missing values are very commonly observed in the dataset they can be seen as Nan, “-”, <blank space> and so on, handling the missing values is crucial when working with data this increases the quality of the dataset. Handling the missing values — Why don’t you fill the missing values with some random number ? Filling the missing values with some random number can cause some serious errors in the dataset. The random value can be an outlier data point and this effect few machine learning algorithms to go wrong by fitting the algorithm to the outlier that can lead to wrong interpretations. This is one of the problems of random number imputation. How does a responsible data person* handle the missing values in the data ? These are the popular ways of handling the missing values in the dataset. Imputation with an appropriate value : This is the most often way of handling missing values. The missing values in the data are filled with the mean, median or mode values of the feature that has missing values in case of tabular dataset. The advantage of this technique is it doesn’t affect the distribution of the data. Since, the distribution is not much affected the nature of the data is preserved mostly. There is a modification to this type of imputation, the missing values can be filled by the mean, median and mode of the missing value feature by grouping the data on the class of the data in case of classification problem. Model based imputation : This type of imputation is done by predicting the missing values with a simple machine learning regression algorithms if the missing value feature is real valued and classification algorithms in case of the feature is a classification. The observations that contain missing values are made as test data and the observations that does not contain missing values as training data for the algorithm to predict the missing value. The value that is being predicted is nothing but the weight(W) value/vector that is learned by the linear regression algorithm by minimizing the loss on training data. Forward/backward filling of missing values: This is the more general technique, the missing value in an observation is filled by the preceding value of the observation in the dataset. In the same way backward filling is done by filling with the value of the observation that is after the missing value observation in the dataset. This technique is not the most effective way of imputation. Manual entry of missing values: In this technique the missing values are filled with the help of the domain expert of the dataset. The appropriate value is filled to the knowledge of domain expert. Making new features: The new features can be made in the dataset where the values are missing, that will be the indication of the values that are missing in the dataset. Note: These are the general techniques of handling missing values, this is highly dependent on the type of data working on. Let’s work on the case study of handling missing values in ‘Pima Indians Diabetes’ Dataset. Case study: As mentioned the dataset we are going to work on is the Pima Indians Diabetes dataset. The missing values in the dataset are mentioned as “0”. The attributes of the dataset are follows Summary of the dataset — Let’s visualize the missing values in the dataset — The white spaces in the dataset are the missing values in the dataset, the class attribute is removed for convenience. Seems there are more missing values in the dataset, from the visualization it is evident that the attribute D and E have more number of missing values than the other attributes in the dataset. As there are more number of missing values in the attributes D and E, We can drop these attributes because this attributes are most likely that it does not contribute for the modelling of machine learning algorithms. This can be the approach of treating the very sparse attributes in the dataset, though there is loss of information this could be the approach. Now, that we have removed the attributes that are very sparse. The missing values in the attributes <A, C, F> can be imputed with the mean of their attributes. Now, we have imputed the missing values in the attributes by mean values. For the attribute B the missing values can be handled using model based imputation. Model based imputation: Procedure: Step-1: Make the attribute<B> as the predictor variable. Step-2: Split the dataset that contains the missing values and no missing values are test and train respectively. Step-3: Train the machine learning algorithm like KNN regressor, It is the most widely used algorithm for imputation of missing values in the dataset. Step-4: Predict the missing values in the attribute of the test data. Step-5: Let’s obtain the complete dataset by combining with the target attribute <I>. These are the missing value imputation techniques that can be used for making the data more reliable. The only goal of this article is to cover the various commonly used techniques that are used for handling missing values in the dataset. The code for this article <- Click here * The person who works and deals with data.

ML Fundamentals — Handling Missing values

93

ml-fundamentals-handling-missing-values-1c9370a4d1e

2018-06-17

2018-06-17 04:57:38

https://medium.com/s/story/ml-fundamentals-handling-missing-values-1c9370a4d1e

false

1,385

null

Machine Learning

machine-learning

Machine Learning

51,320

Rahul Vamusani

null

d19fdcc9ac69

rahul.vamusani98

24

46

20,181,104

null

0

null

0

null

2018-02-11

2018-02-11 06:15:36

2018-02-15

2018-02-15 14:10:31

1

false

en

2018-02-15

2018-02-15 14:10:31

0

1c93ef07b10c

0.856604

0

*This blog is my review of paper “Generating Wikipedia by Summarizing Long Sequences”

2

Text Summarization on Longer Articles *This blog is my review of paper “Generating Wikipedia by Summarizing Long Sequences” I am deeply concerned about the techniques of text summarization because my goal is to summarize a book. In the past, the biggest disadvantage of text summarization is that it is hard for a neural network to summarize long articles. In this paper, the authors propose an architecture which can summarize multiple articles. There are two stages of the model. First, the model will use extractive methods to summarize some words for each article. Second, the model apply TRANSFORMER-DECODER to do abstractive summarization. The TRANSFORMER-DECODER model is the one used in the paper “Attention is all you need”. For the advanced version, the paper add memory-compressed attention on the architecture. The concept is really easy. The thing can be challenged is how to do end-to-end learning without stage 1. The memory issue forces the model to use non-neural-network methods to do extractive methods. We can think more on it.

Text Summarization on Longer Articles

0

text-summarization-on-longer-articles-1c93ef07b10c

2018-06-14

2018-06-14 20:22:07

https://medium.com/s/story/text-summarization-on-longer-articles-1c93ef07b10c

false

174

null

Machine Learning

machine-learning

Machine Learning

51,320

Yi Yao Huang

My research interest is deep learning on Natural Language Processing and AI. To apply AI to education, I force myself to think, write and code everyday.

c7e3733fb89

darrenyaoyao.huang

81

18

20,181,104

null

0

null

0

null

2018-01-11

2018-01-11 05:06:01

2018-01-11

2018-01-11 06:43:36

0

false

en

2018-01-11

2018-01-11 06:45:08

1

1c9472cb8836

2.075472

8

0

Here comes again an exciting time of the year for economists: The 2018 American Economics Association (short for AEA) meeting just took…

4

How could an economist find a job in tech companies? Here comes again an exciting time of the year for economists: The 2018 American Economics Association (short for AEA) meeting just took place in (the freezing) Philadelphia with several thousands of economists joining from all over the world to this largest annual conference in the field of Economics. Apart from workshops and panel sessions, many of the participants also just flew here just for the job market interviews organized by the AEA in parallel. On the demand side, I am seeing more and more tech companies considered recruiting from the AEA job market, or at the very least, posted their job openings to increase their visibility. Compared to the AEA job market I participated four years ago (also took place in Philly with a blizzard not so much by chance :p), for instance, I saw a couple of new names on the list (such as the Alibaba Inc. and Houzz) apart from AEA patrons like Amazon, Facebook and more recently, Uber. Given that many fellow economists will be invited onsite in the coming weeks, I would love to share with you an article I wrote about economists’s roles in tech companies and how to prepare the transition from academia / school to industry. It was based on my two-year (somewhat painful) transition from a senior research fellow at the Max Planck Institute to a Facebook data scientist, focusing on advertising effectiveness measurement. The original article is in Chinese, but here’s a topline summary: Common types of problems economist could contribute significantly in tech companies include: demand forecasting, pricing, market design (design economics as Al Roth put it in this year’s AEA keynote speech), regulation policy research, advertising effectiveness analysis, real estate market analysis, (behavioral) finance, education and labor market matching design, etc. The most important tips for job interviews in these tech companies include: Get to know more about the language of statisticians and other social scientists (such as political scientists, psychologists and sociologists or even computer scientists). By reading books or talking to them, you might come to realize that they use either the same or distinctive methods and approaches to solve similar problems. Since you might be working with them on the daily basis down the road, it is essential that you could understand their language and your comparative/competitive advantage to your team skill set-wise. Try to be proficient in R, Python and SQL, which are the most popular languages in the data science world. This will almost certainly be asked during an onsite interview. Pick up probability and grad school stats. You might face a technical interview on that. Simple is beautiful — Practice using plain language to explain everything in your models that your grandma can understand. Remember to switch that mindset when talking to non-technical interviewers (such as recruiter, hiring manager’s manager, etc). Talk to alumni especially from the company you are interested in interviewing to better understand the value and culture of that company. Although companies will not explicitly judge you based on “culture fit” but it will make you more stand out if your passion and interests are in line with what they care about. Wish everyone gets his or her dream job offer by the end of this market!

How could an economist find a job in tech companies?

10

how-could-an-economist-find-a-job-in-tech-companies-1c9472cb8836

2018-03-22

2018-03-22 18:52:53

https://medium.com/s/story/how-could-an-economist-find-a-job-in-tech-companies-1c9472cb8836

false

550

null

Data Science

data-science

Data Science

33,617

Fangfang Tan

I currently work at the Google focusing on campaign effectiveness measurement, a diehard experimental and behavioral economist

54621dfd012

fangfangtan

164

163

20,181,104

null

0

null

0

null

2017-09-14

2017-09-14 15:19:52

2017-09-14

2017-09-14 15:25:47

0

false

en

2017-09-21

2017-09-21 09:31:00

0

1c94f1fb3f53

0.871698

1

0

We would like to announce final results of Call For Proposals (CFP), and we selected 2 more proposals from Aditthya and Ayush in our round…

4

Final Results of Calls For Proposals in PyCon HK 2017 We would like to announce final results of Call For Proposals (CFP), and we selected 2 more proposals from Aditthya and Ayush in our round 2 selection. More topics delivered by our sponsors and invited speakers will be announced in October. The following topics are selected from CFP. Using Gradient Boosting Machines in Python (Albert Au Yeung) Machine Learning on Energy Consumption Prediction (Huang Wei Chen) Python for data analysis (Michal Szczecinski) Boosting command line data manipulation with Python and AWK (Kirill Pavlov) How I battle with Hong Kong Open Data in Python (Ho Wa Wong) HA capability with Document Store using MySQL Shell — running Python (Ivan Ma) Applying serverless architecture pattern to distributed data processing (Denis Makogon) How to reinvent the wheel and build the most popular JSON-RPC library (Kirill Pavlov) Ticketing X Chatbot (Comma) Python Blockchain Application in < 24 hrs (Kelvin Chu) AI learn to Drive: Introduction to Reinforcement Learning with Python (Holman Tai) Python is not Always Slow (Benny) (**Cancelled** - Speaker withdrew) Matplotlib 2 By Example (Claire Chung) Python Logging in Production (Mengchi JIA) Micropython (Patrick Tsoi) Recurrent Neural Networks in Python: Keras and TensorFlow for Time Series Analysis (Matt O’Connor) Resurrecting the dead with deep learning (Aditthya Ramakrishnan) How to approach a Machine Learning Problem with Python ?: YouTube Like Count Prediction (Ayush Singh)

Final Results of Calls For Proposals in PyCon HK 2017

1

final-results-of-calls-for-proposals-in-pycon-hk-2017-1c94f1fb3f53

2018-06-01

2018-06-01 00:25:56

https://medium.com/s/story/final-results-of-calls-for-proposals-in-pycon-hk-2017-1c94f1fb3f53

false

231

null

Machine Learning

machine-learning

Machine Learning

51,320

PyCon HK

PyCon HK is the annual conference hosted by the python community in Hong Kong for developers using and developing the open-source Python programming language.

407a1bcc14ea

pyconhk

35

1

20,181,104

null

0

dataset = tf.data.Dataset.from_tensor_slices((image_list, label_list)) def _read_py_function(path, label): image = read_image(path) label = np.array(label, dtype=np.uint8) return image.astype(np.int32), label def _resize_function(image_decoded, label): image_decoded.set_shape([None, None, None]) image_resized = tf.image.resize_images(image_decoded, [28, 28]) return image_resized, label dataset = dataset.map( lambda data_list, label_list: tuple(tf.py_func(_read_py_function, [data_list, label_list], [tf.int32, tf.uint8]))) dataset = dataset.map(_resize_function) dataset = dataset.repeat() dataset = dataset.shuffle(buffer_size=(int(len(data_list) * 0.4) + 3 * batch_size)) dataset = dataset.batch(batch_size) iterator = dataset.make_initializable_iterator() image_stacked, label_stacked = iterator.get_next() with tf.Session() as sess: sess.run(iterator.initializer) image, label = sess.run([image_stacked, label_stacked])

9

c5cbafc8429d

2018-06-17

2018-06-17 10:39:56

2018-06-17

2018-06-17 14:09:36

0

false

ko

2018-06-17

2018-06-17 14:09:36

2

1c96f17c3696

4.596

1

0

Generate batch data with tf.data

4

[input_data] tf.data 으로 batch 만들기 Generate batch data with tf.data Importing Data | TensorFlow The API supports a variety of file formats so that you can process large datasets that do not fit in memory. For…www.tensorflow.org 작년에는 데이터를 tfrecord로 변환하고서 그 파일을 학습 데이터로 넣으려고 할 때 enqueue dequeue를 이용하면 코드도 복잡하고, 여러가지 불편함들을 여간 많았던 것이 아니다. 올해 초 반가운 텐서플로우의 새로운 함수가 나왔다. tf.data 라는 것인데. 획기적으로 tfrecord를 열 수 있는 것 뿐만이 아니라 일반 이미지도 역시 손쉽게 배치로 생성하고 넣을 수 있다. tf.data 로 데이터 만들기 데이터를 다루게 된다면 제일 먼저 생각하게 되는 순서는 아래와 같을 것이다. 데이터의 경로를 찾는다. 데이터의 목록들을 가져오고, 이미지와 레이블을 생각한다. 목록의 한 열 마다 데이터를 여는 방법을 정의한다. shuffle 이든 다양한 옵션을 준다. 배치로 만들어서 model 넣을 준비를 한다. 1. 데이터 준비 제일 먼저 원하는 데이터들의 경로를 받아 리스트로 담고서 아래와 같은 함수에 넣어준다. tf.data.TFRecordDataset(filenames) tfrecord 데이터로 돌릴려고 할 땐 이 함수를 쓰면 된다. tf.data.Dataset.from_tensor_slices(filenames) 제일 먼저 일반 이미지나 array를 넣을 때 list 형식으로 넣어준다. 이미지 경로들이 담긴 리스트 일 수도 있고, raw 데이터의 리스트 일 수도 있다. 이번 글에서는 이 함수로 예제로 보여줄 것이다. 2. tfrecords가 아닌 numpy 나 image를 읽어야 한다면 reader 및 preprocess 함수 정의 PIL의 Image로 읽든 openCV로 읽든 어떤 방법으로든 이미지를 읽을 reader 함수를 정의 합니다. 여기에 동시에 preprocess 들을 같이 넣을 수도 있습니다. 3. 읽은 데이터들에게 다른 옵션들 정의 dataset으로 정의 했다면 이를 다양하게 설정 해줄 수 있는데. 설명을 주자면 아래와 같다. repeat(step_n) : 원하는 epoch 수를 넣을 수 있다. 아무런 파라미터를 주지 않는다면 iteration이 무제한으로 돌아간다. shuffle(1000) : 한번 epoch이 돌고나서 랜덤하게 섞을 것인지 정한다. 4. batch size 정의하기 tf.data 를 사용하지 않을 땐 여러 방법들로 batch 를 만들었지만 tf.data는 이렇게 간단히 정의가 가능하다. [input_data] Image 데이터로 Batch 만들어 넣기 — 참고 5. iterator 정의 마지막으로 iterator 정의 해주고나면 모델에 넣을 image_stacked와 label_stacked까지 만들어 주면 된다. 위 코드 중 image_stacked와 label_stacked가 왜 두개로 나눴는지 이해가 안 될 수 있는데. 처음에 데이터 경로를 넣었을 때 image_list와 label_list를 tuple로 넣었기 때문에 이렇게 두 갈래로 나뉘기 전까진 계속 두개가 계속 흘러가는 것이라 생각하면 된다. 6. Image 읽기 Session을 열어주고, 아래와 같이 initializer를 한번 run 해준 후 image를 sess.run을 해주면 잘 뽑아지는지 확인 하면 된다. 최종 코드는 아래에서 확인 할 수 있다.

[input_data] tf.data 으로 batch 만들기

10

input-data-tf-data-으로-batch-만들기-1c96f17c3696

2018-06-17

2018-06-17 14:09:38

https://medium.com/s/story/input-data-tf-data-으로-batch-만들기-1c96f17c3696

false

441

This is blog for result of research and sharing codes for deep learning

null

TrackIn DataLabs

trackindatalabs@gmail.com

trackin-datalabs

TENSORFLOW,DEEP LEARNING,ARTIFICIAL INTELLIGENCE,MACHINE LEARNING

null

Deep Learning

deep-learning

Deep Learning

12,189

정겨울

null

63c78a743ca5

junhoning_data

26

29

20,181,104

null

0

null

0

null

2018-03-13

2018-03-13 02:28:35

2018-03-13

2018-03-13 10:16:55

2

false

ja

2018-03-13

2018-03-13 10:16:55

1

1c97c9d95e98

3.883333

2

0

文書分類の手法としてナイーブベイズ (naive bayse classifier) が用いられることは多くあります。ナイーブベイズでは分類対象がどのようなモデルで生成されるかを仮定する必要があり、文書分類においては多項モデルを用いることが多いようです。

3

正規分布モデルのナイーブベイズによる文書分類は可能なのか文書分類の手法としてナイーブベイズ (naive bayse classifier) が用いられることは多くあります。ナイーブベイズでは分類対象がどのようなモデルで生成されるかを仮定する必要があり、文書分類においては多項モデルを用いることが多いようです。ここでふと他のモデルで文書分類を行なうことが出来ないのかということが気になりました。今回はlivedoor ニュースコーパスの記事をナイーブベイズによって分類し、多項モデルと正規分布モデルで性能がどれほど異なるかを確認しました。結果としては多項分布モデルのほうが性能が良いという当たり前の結果になりましたが、モデルに付いて理解を深めることができました。分類器説明ナイーブベイズはscikit-learnを用いています。分類にあたり、以下処理は共通して用いています。形態素解析 Mecab＋mecab-ipadic-NEologd 文書表現 BoW (Bag-of-Words)による単語出現頻度 TF-IDFによる重み付け前処理クリーニング、正規化、ストップワード前処理についてはHironsanによる説明が詳しく、実装においても利用させてもらっています。多項モデル多項モデルは文書中の各位置において、どの単語が出現するかをモデル化したものです。このモデルは複数種類の離散値から1つの値が発生する試行を複数回行って得られる多項分布に従っています。サイコロでイメージを掴むことができ、出現する目は1~6の離散値であり複数回振る事によって出目の分布が得られます。この出目の分布がそのサイコロの特徴と言うことができます。 (1が出やすいサイコロなど) 文書分類をサイコロに例えると、分類したいカテゴリそれぞれを複数のサイコロに、各単語をサイコロの目に置き換えて考えることができます。単語の発生確率の分布がそのカテゴリの特徴といえます。(「政治カテゴリ」のサイコロでは「献金」と言う目がよく出る分布であるなど) 推定においては対象文書中に出現した全ての単語について、どのカテゴリのサイコロを振り続けた場合に最も尤もらしいか求めることで分類を行います。正規分布モデル今回仮定した正規分布モデルは、文書における各単語の出現数について着目しました。カテゴリ毎に単語の出現頻度が異なり、正規分布としてモデル化できるのではないかと考えたためです。 (「スポーツ」カテゴリであれば「選手」という単語が平均でX回発生し、分散σの正規分布が得られると仮定) 実験結果・考察 livedoor ニュースコーパスを対象に分類を行いました。学習データ：テストデータ＝6:4としており、今回は交差検証まで行えていません。結果としては多項モデルの方が性能がよく、f1-Scoreで0.85となっていました。一方正規分布モデルでは0.81となっており芳しくありません。ここで正規分布モデルで仮定を置いていた単語の発生頻度が正規分布にそっていたのか確認を行ってみます。あるカテゴリ(523文書)における文書中の各単語の発生数をプロットしました。系列は各単語であり、x軸はひとつの文書中で発生した回数、y軸は該当文書数です。その結果はほとんど0回に偏っており、疎なデータであることが確認できました。それは考えると当然であり、「ボール」という単語が「スポーツ」カテゴリの文書において必ず出てくるわけではなく、また文書中においても何回も記述されるわけではありません。よって、正規分布と仮定を置くのは難しかったということから性能がでなかったと思われます。おわりにナイーブベイズにおける文書分類において、単語出現数が正規分布に沿うと仮定を置くのは難しいと確認しました。それは文書で出現する単語の種類は多く、疎なデータであるためでした。今後は単語という離散的な素性でなく、連続的な素性が得られれば試してみたいと思います。今回は用いたいモデルからスタートする逆からのアプローチであり進め方として適切ではありませんが、正規分布の文書分類へ適用を考える上での考察を深めるために実験を行いました。また、本来は仮定を裏付けるデータの確認を行った後に検証をすべきだと思います。

正規分布モデルのナイーブベイズによる文書分類は可能なのか

4

正規分布モデルのナイーブベイズによる文書分類は可能なのか-1c97c9d95e98

2018-05-03

2018-05-03 17:49:57

https://medium.com/s/story/正規分布モデルのナイーブベイズによる文書分類は可能なのか-1c97c9d95e98

false

44

null

Naturallanguageprocessing

naturallanguageprocessing

Naturallanguageprocessing

848

kamujun

null

dc02d1b25000

kamujun

3

4

20,181,104

null

0

null

0

null

2018-09-03

2018-09-03 13:10:17

2018-09-03

2018-09-03 13:12:37

1

false

en

2018-09-03

2018-09-03 13:12:37

3

1c994d7c6809

0.233962

0

null

5

#AntiBank is #ArtificialIntelligence intelligence in the world of #finance.

0

antibank-is-artificialintelligence-intelligence-in-the-world-of-finance-1c994d7c6809

2018-09-03

2018-09-03 13:12:38

https://medium.com/s/story/antibank-is-artificialintelligence-intelligence-in-the-world-of-finance-1c994d7c6809

false

9

null

Ethereum

ethereum

Ethereum

76,961

AntiBank

AntiBank is artificial intelligence in finance.

45bbe34e6c0a

antibank.io

1

20,181,104

null

0

null

0

null

2018-01-09

2018-01-09 20:03:34

2018-01-09

2018-01-09 20:06:27

1

true

en

2018-01-10

2018-01-10 19:29:48

36

1c99d573b29d

9.486792

0

Maximizing Value in the Age of Information

5

The Price — The Cost — And The Market Maximizing Value in the Age of Information I. INTRODUCTION The technological evolution of consumer retail markets has driven the spawn of a new dimension of data gathering that has become required consideration for any accurate attempt at economic market analysis today. The price of bread and milk is no longer just a factor determined by supply & demand market forces, competition, and the deals made by middle men. Now there is a fourth major element, data — big data. The commodities exchange at every level of the supply line results in tabulated metrics on both sides of each exchange. Optimization theories are constantly tested and refined for greater yield, lower cost, and stronger consumer appeal. One could argue that this has been the case since traders traveled the silk road across East Asia carrying fabrics and spices. However, the difference today is the availability and aggregation of information used in business decision making, which unquestionably, is like never before. Data observation has evolved into market prediction algorithms, consumer desirability metrics, and price point optimizers largely in favor of producers. The question becomes whether a market supplied with information by businesses, governments, and individuals then passes on the total costs of production and total savings from data intelligence, on to the final consumer; and further, does the price consumers pay with not just their money but also their information align with the value exchange necessary to maintain a free market society? This paper argues that it does not, then challenges the capabilities of governmental regulation as a prospective intervenor, and finally advocates for the proliferation of one emerging technology as an intermediary to balance the scales across supply chains and establish a consistent exchange of value for the price — the cost — and the market. II. THE PRICE The price we pay for something in the United States has been traditionally determined by free market supply and demand economics where market forces create equilibrium (the point where supply equals demand.) As demand increases, supply and price increase correspondingly. A. Prices Change for a Reason In a post WWII article in The American Economic Review entitled The Use of Knowledge in Society, Friedrich Hayek argues that “it does not matter to a purchaser why the price of screws went up, only that it did.” He is probably right in the case of small consumer retail purchases. The question of ‘why’ is unlikely asked because, in most retail circumstances, consumers have options. Antitrust laws against monopolies help to ensure this. “A monopoly exists when there is only one producer of a particular good or service for which there are no close substitutes and the producer is protected from competition by impenetrable barriers to entry.”[1] Thus, consumers can pick up a pack of Leland screws if Infasco’s seem too much. “All that is significant for [consumers] is how much more or less difficult to procure [products] have become compared with other things with which [they are] also concerned, or how much more or less urgently wanted are the alternative things….”[2] Nevertheless, prices change for a reason. Perhaps Infasco is primarily supplied by a Turkish sheet metal producer recently petitioned by domestic sheet metal producers under antidumping duty claims on the basis of evidence indicating Turkish subsidies for sheet metal exporters; provided by a Turkish dumping maneuver to drive out U.S. domestic sheet metal producers then raise prices. Or, perhaps the global supply lines of steel are being cut off by a civil uprising across the world’s largest iron ore mines in Brazil, impacting the price of steel internationally. The point is that smart business do not raise their prices without a reason. And if the price of Leland screws has not changed, that could be a good or bad thing for the consumer depending on whether or not Leland is affected by the events currently affecting the price of production for Infasco. If not, great, the less expensive purchase would probably be a good one. If not, there could be some loss of value in terms of the quality of Leland screws that is not evident at the time of purchase to offset the increased cost of production. That in turn could affect the quality of whatever the consumer is using the screws for. Altogether, the question of ‘why’ prices change may be more important to the consumer than realized. “But the “man on the spot” cannot decide solely on the basis of his limited but intimate knowledge of the facts of his immediate surroundings. There still remains the problem of communicating to him such further information as he needs to fit his decisions into the whole pattern of changes of the larger economic system.”[3] B. Asymmetric Information “Asymmetric information (that is, one party having more information than another) is one of the three ways in which markets can fail — the other ways being misuse of market power and externalities. In a competitive market, firms should strive to cater better to the information needs of consumers…. [I]n order that consumers can provide the driving force behind competition between firms they need to be able to access, assess and act on relevant information.”[4] The information available to consumers is out of balance with the information available to firms. In the age of information, many firms utilize customer data to maximize returns at the expense of the consumer who may not be aware of the quality tradeoffs being made by firms and suffer a loss in value without being aware of such losses. Without consumer access to comparable degrees of information in a form that they can easily access at the point of sale to influence their decision-making, there is an increasing possibility of market failure. “Information can assist consumers to make sound purchasing choices, allowing them to make well-informed and well-reasoned decisions that reward those firms which best satisfy their needs. Markets work well when there are efficient interactions on both the consumer and the firm side. Well-informed and confident consumers can play a key role in promoting vigorous competition between firms.”[5] The challenge is with designing effective social policies without engaging in some sort of cost-benefit analysis. C. The Problem “The availability of such information has long been believed to be impossible to acquire by any individual….[6] This has been called the lack of “rational economic information.”[7] “The problem is precisely how to extend the span of our utilization of resources beyond the span of the control of any one mind; and, therefore, how to dispense with the need of conscious control and how to provide inducements which will make the individuals do the desirable things without anyone having to tell them what to do.”[8] If price information was composed by a third party merchants would probably react to the pricing of the same or very similar items available in other places to better price their items for alignments with the expectations of consumer demand. There might also be a configuration of price by producers for merchants operating under similar conditions, which would pass on the savings benefit to consumers. “[T]he method by which such knowledge can be made as widely available as possible is precisely the problem to which we have to find an answer.”[9] The problem is “the unavoidable imperfection of man’s knowledge and the consequent need for a process by which knowledge is constantly communicated and acquired.”[10] The solutions currently proposed are inadequate and “it is time that we remember that it does not deal with the social process at all and that it is no more than a useful preliminary to the study of the main problem.”[11] III. The Cost The cost of producing products and negotiating with intermediaries has historically led to inefficiencies for one party or another, typically the party with less information and bargaining power. One of the options for parties with less bargaining power and information could be information sharing amongst similarly situated parties to strengthen their negotiating leverage in future contracting. “[I]nformation dissemination and sharing can allow firms to benchmark themselves in critical areas against other firms, including actual or potential competitors. This can promote innovation and best practice and enhance efficiency, which can drive competition in sectors. For example, comparing business processes and performance against best practices within a sector or across sectors may allow firms to develop plans on how to make improvements in quality or to adapt specific practices with the aim of doing things better, faster and cheaper.”[12] A. Price Fixing Concerns This practice might be considered price fixing which would have negative results for the economy and may further be considered cartel activities which might subject the participating firms to criminal liability. Price fixing is problematic because it disrupts free market activity and can throw of the equilibrium of supply and demand economics. If unchecked, the results may resemble a government instituted binding price ceiling, where the economy experiences a deadweight loss. “There is a phenomenon of advantage for every individual in that he has information that is not available to anyone else, and he may use that information to make profits on the differences of prices in commodities by acting as an “arbitrageur.”[13] IV. The Market The market for retails products is vast and disparate. Prices vary and decisions are made to increase profits based on efficiency implementations that are not always for the benefit of consumers. Firms use information about consumers acquired from web presences on social media and purchasing behavior on e-commerce platforms like Amazon.com. The data acquired is largely used for the benefit of Amazon and its product producers. A. Market Conditions “The continuous flow of goods and services is maintained by constant deliberate adjustments, by new dispositions made every day in the light of circumstances not known the day before…”[14] Offline, in the brick and mortar retail space, measures taken using sales information to determine product marketing effects and the results of advertising campaigns and other business building practices. “Everything from the color of a producer’s logo — or of the product itself — to the shape of its packaging can change people’s preferences….”[15] In other cases, consumers may not notice losses in value, For example, you may have noticed your fruit juice container become reduced from the 64oz bottle a slightly thinner 60oz bottle for the same price. Many consumers are not aware of these kinds of cost saving practices performed by producers and information sharing amongst consumers can help to ensure fair practices and value preservation in the consumer retail space. Further, the number of options available to consumer can also influence their purchasing decisions. For instance, “[w]hen Proctor and Gamble winnowed its 26 varieties of Head and Shoulders antidandruff shampoo down to 15, eliminating the least popular, sales jumped by 10 percent.”[16] B. Consumer Focused Solution There may be a solution available through the wide use of an intermediary consumer review technology. For example, an information sharing platform for consumers to hold producers and merchants accountable for their products. In The Law and Economics of Information Sharing, Bennett and Colling make and interesting recommendation regarding the benefits and possibilities for a consumer centric information sharing system: “Consumers must have access to information from individual firms that will allow them to make the most suitable choice for their needs. Information about offerings of goods or services may not always be made available to consumers in order to help them determine which good/service best suits their particular needs. Consumers must be able to assess the information available to them in order to compare firm offerings and pick the most suitable choice. Without clear information on each firm’s offerings, this will not be possible. Indeed, even when the information on individual offerings for goods and services is available to consumers, it may still be difficult to make effective choices due to the way the information is presented. Certain services, in particular, are complicated, and simple information may leave out too much that is important and lead to choices being distorted; complex information may confuse and have the same effect. Having an intermediary available, such as a price comparison website, that shares relevant comparative information can be a solution to this.”[17] If this is done, the whole [might] act as one market, not because any of its members survey the whole field, but because their limited individual fields of vision sufficiently overlap so that … the relevant information is communicated to all.”[18] If consumers can cooperate through information sharing there is a chance of maximizing value in the era of information to better offset the value loss efficiencies instituted by producers by using big data. IV. Summary & Conclusion Technology has changed the landscape of retail economics in ways that many people do not realize. At the moment, the power of data is primarily in the hands of manufactures and producers which often take advantage of consumers by information gathering at a level like never before. Consumers are impacted by positive and negative externalities that limit their ascertainment of value and knowledge about the changes that are happening in the market today. The capacity for middle men such as retail merchants to garner strength is likely limited by the current legal environment which prohibits information sharing that results in price fixing. They are often the weaker bargaining party in contracting with production firms which leaves them at the mercy of supply and demand equilibrium which unlikely captures the true costs of production and value exchange. The solution appears to be left to private firms to establish intermediary technology platforms readily accessible by consumers to better balance the state of asymmetric information for the benefit of consumers. Technology is changing the world in a way that may be beyond the handle of law, and it appears to be left to consumers to empower themselves by utilizing forms of big data information gathering tools, like smart phone apps that are consumer friendly, in order to restore balance to the free market economy. [1] Butler, Henry N., Christopher R. Drahozal, and Joanna Shepherd. Economic analysis for lawyers. Durham (North Carolina): Carolina Academic Press, 2014. Print., 413 [2] Hayek, Friedrich. “THE USE OF KNOWLEDGE IN SOCIETY.” The American Economic Review XXXV.4 (September 1945): Page 525. Print. [3] Id., 524–525 [4] Bennett, Matthew, and Philip Collins. “The Law and Economics of Information Sharing: The Good, the Bad and the Ugly.” European Competition Journal 6.2 (2010): 315 [5] Id., 314–15 [6] Hayek, 519 [7] Id. [8] Id., 526 [9] Id., 522 [10] Id., 530 [11] Id. [12] Bennett, 318 [13] Hayek, 522 [14] Id., 524 [15] Iyengar, Sheena. The art of choosing. London: Abacus, 2012. Print., 152 [16] Id., 190 [17] Bennett, 315–16 [18] Hayek, 526

The Price — The Cost — And The Market

0

the-price-the-cost-and-the-market-1c99d573b29d

2018-04-15

2018-04-15 04:08:20

https://medium.com/s/story/the-price-the-cost-and-the-market-1c99d573b29d

false

2,461

null

Economics

economics

Economics

36,686

Jordan Powell

The Founder of Pricecheck

2d328ef8c3e

pricecheck

0

1

20,181,104

null

0

null

0

null

2018-02-27

2018-02-27 05:43:28

2018-02-27

2018-02-27 08:41:21

4

false

en

2018-02-27

2018-02-27 08:41:21

0

1c9c52cf6222

5.507547

1

0

Data scientists and analysts have some of the hottest skills in the IT market right now. They are being paid high salaries and are using…

4

If you want to look sexy, then think data governance. Data scientists and analysts have some of the hottest skills in the IT market right now. They are being paid high salaries and are using advanced analysis techniques to uncover hidden patterns and indicators within an organisations that can be utilised to drive decision making. Analytics and insights when visualised with tools like Tableau or Qlik Sense, can be pretty damn sexy. With all this hype and excitement there is a huge amount of pressure to go out and get your very own expert to start immediately producing these magical insights. But what is often overlooked is the even sexier area of data governance that is at the core of any successful insights program. Data governance? I’m not convinced it’s as sexy as you think. Have you ever been in a meeting and been asked how many customers that you have, or how many Tier 1 clients in your portfolio, or who your most important customer is? If I ask these questions in my business, I get answers like “around X” or “it depends on what you mean by Tier 1” and so on. They are simple questions that are incredibly difficult to nail down. In many organisations our definitions of data terms are not clearly defined, they become subjective and this drains the value from any analysis that we might undertake. If data is not accurate and trusted, the insights we gain from the data are meaningless and certainly should not drive decision making. What we need is a data governance framework, to provide the organisation with high quality data. This framework helps ensure quality, access and integration of data to maximise the effectiveness of the data analytics solutions. It encompasses the policies, procedures, roles and responsibilities relating to data management. I’ll show you my framework if you show me yours There are lots of ways to visualise a data governance framework, but this is one of my favourites, which has been created by the University of Notre Dame in the US. This framework is effective because it focuses on the need to provide access to data as an outcome, which is facilitated through the use of technology but is reliant on defined business processes and attitudes to ensure that this accurate data is accessed in a controlled and appropriate manner. Lean mean data machine If you are a larger organisation, or an organisation that has a large number of data repositories or just simply a very large amount of data, defining and deploying a data governance framework across all of your data is going to be extremely challenging and highly unlikely to succeed. A much lower risk method is to utilise a “lean” approach. That is, to identify one or two areas where there will be a clear business benefit from potential improvement and where you can also have the senior support and resources required to drive a successful outcome. If you are successful with these lead areas, you will be able to deploy your framework iteratively across the entire organisation. Data governance just got RACI Across an organisation there are many “keepers” of data and an effective data governance framework will require their buy-in. One way to achieve this is through the establishment of a Data Council. Members of this group should represent each of the data custodians or stewards. These data custodians must first classify the data that will be governed by the framework. In classifying the data they must agree on a single definition of each term. This effort has been estimated to take up to 10 hours per term, so it is not a simple or quick undertaking. The use of the RACI (Responsible, Accountable, Consulted, Informed) matrix allows stakeholders across the organisation to self-declare an interest in specific data terms, which increases the likelihood that the stakeholder will buy-in to the data governance process. An example of the RACI matrix is shown below: Remember that as the business changes over time, the definition of these data items may also change. It is the responsibility of the Data Council to keep the definitions up to date and they should use the RACI matrix to engage stakeholders and guide the process of review and approval. The data dictionary will allow high value data fields that are utilised by many parts of the organisation to be identified. These then drive the design of the core data object models, used to architect the data warehouse/s. No time like the present The process of extracting data from legacy systems, transforming and loading into the data warehouse, will highlight data quality issues which may be a result of data entry or possibly erroneous business logic. This is an opportune time to address the data quality issues, before the data is inserted into the warehouse and ideally fixing them in the source system. Another consideration at this stage is the reporting required by the business. The development of key reports that are in a format expected by the various user groups, which serve to answer the most common and important queries for the business, will drive consistency of the use of data across the organisation as well as promoting user adoption. The consideration of reporting will also drive the development of the security model which must sit across the data warehouse and control access to the data. Applying the security model on the data warehouse allows the data custodian to control access to the data at a row and field level. This approach avoids the problematic alternative of working directly with a legacy system application security model. The next two data governance areas to be considered in the warehousing solution are compliance and archiving. These areas are likely to be driven by external influences such as regulatory bodies that might include APRA, PPIPA, HRIPA, State Records Act and many others. These will define how data must be stored to maintain compliance and how long data must be maintained. These factors can have large impacts on the solution given the different volume of data that may be required to be maintained. If you build it, they might not come… The final element of the framework is training and awareness. This is a critical area that is frequently overlooked. I once worked with a clients analytics team who couldn’t understand why people weren’t using their reports. The team had created around 170 insightful reports, deploying them as they were built on their company intranet. The problem was they didn’t tell anyone that they were there, or why the report was relevant. When you are building your reporting solution, it is critical to engage with the stakeholders to gather their input before the reports are developed. Communicating and training stakeholders to interpret the reports is critical to their understanding and application. Any reporting solution generated by an analytics team in isolation, is at risk of being a report for the sake of it. I told you governance was sexy While we all love to create data visualisations, without a data governance framework these can easily become interactive pictures without insight. A data governance framework doesn’t have to be the death of the party, but it must address issues with data quality, standardising data definitions, consider appropriate security measures, as well as compliance and retention requirements. Like all IT projects, I recommend a lean approach, by working with a smaller subset of high value business data to design a governance framework that will be practical in your organisation. Once validated, you can continue to deploy that framework iteratively across your entire organisation and data sources.

If you want to look sexy, then think data governance.

2

if-you-want-to-look-sexy-then-think-data-governance-1c9c52cf6222

2018-02-27

2018-02-27 13:31:11

https://medium.com/s/story/if-you-want-to-look-sexy-then-think-data-governance-1c9c52cf6222

false

1,274

null

Analytics

analytics

Analytics

15,193

James Harper

Managing Director and Lead Consultant at Harper Stone

1903026a3520

jamieharper

1

2

20,181,104

null

0

null

0

c92c31591e3a

2018-01-21

2018-01-21 20:46:27

2018-01-23

2018-01-23 16:59:00

2

false

en

2018-04-08

2018-04-08 15:25:13

9

1c9c5f80ab89

7.515409

410

23

0

A cortex by any other name

5

Evolution doesn’t give a damn what you think a brain region is called A cortex by any other name Credit: Pixabay We do love naming things. Nouns are the first words we learn: mummy; daddy. Or if you’re my kids: “digger” and “Bob” (the cat) — because who needs parents when there’s a big ginger cat to play with? That love for tagging an object with a label never goes away. Neuroscientists are no different. In large brains, like ours or rats or ferrets, we give hundreds of names to their constituent parts. Some name vast swathes of brain tissue: cortex, brainstem, midbrain. Some name tiny subdivisions of those swathes, like the almond shaped subthalamic nucleus (literally, the cluster of neurons underneath the thalamus). In tiny brains, like maggots or roundworms, even individual neurons have names. (Boring ones like “P4”; sadly, there is no neuron in C Elegans called Boris). By naming these bits of brain, we can communicate effectively, we can talk about brains and know we are talking about the same bit. We can navigate where we are in the brain by knowing the names of the brain regions. We can even identify ourselves by the bit of brain we study: “I work on cortex”; “I work on hippocampus”; “I couldn’t decide what I wanted to do when I left school; now I work on the nucleus ambiguus”. Naming bits of brain greases the wheels of scientific communication. But it is also a dangerous game. If we confuse the convenience of naming with the reality of the brain, we end up in deep trouble. Because evolution does not give a crap what we call a brain region. The danger is that naming a bit of brain makes us think it is a discrete thing, a bit that just lifts right out, which we can study and marvel at in isolation. But how evolution (and development) conspire to disperse and wire together neurons pays no attention to these names. Consider primary motor and somatosensory cortex. They sit next to each other. But they appear in different chapters in textbooks. Entirely separate fields of research have grown up around them. Work on primary somatosensory cortex looks at how the activity of its neurons represent touch. Work on motor cortex looks how the activity of its neurons represent movement. These are starkly different: work on somatosensory cortex focuses on its inputs; work on motor cortex focuses on its outputs. They are treated so separately that papers merely showing the flow of activity from one to the other end up in high-profile journals. But these bits of cortex are next to each other. Either we believe that there are border guards who turn away the motor cortex neuron axons at the crossing with somatosensory cortex, and the same from the other direction. Or we have to assume that these two names loosely delineate a continuous network of neurons by the fact that a small set of neurons in somatosensory cortex get direct sensory input, and small set of neurons in motor cortex connect to the spinal cord. Most of the neurons in these bits of cortex neither get sensory input from the thalamus nor project to the spinal cord. They are wired to other neurons all over cortex, and very much to each other. I mean, look at this nutter: The projections of the axon of a single neuron in rodent motor cortex. From Zeng & Hanes 2017 (Nature Neurosci Reviews) Evolution has deemed it necessary for this neuron in motor cortex to send its output wires, its axons, all over the shop. To the somatosensory cortex. To prefrontal cortex, so that the apparent seat of “executive function” — cogitating, planning, turning things over in your mind — apparently needs to know what your foot will be doing shortly. To the ectorhinal cortex (no, me neither). All over the striatum. And to the other side of the brain — yes, the brain has two sides. We call it a motor cortex neuron; it is studied for how it represents simple movements; but it is not a neuron that sits neatly tucked up in one neatly labelled part of the brain — it is part of the the brain’s sprawling, tangled network. Cortex is divided up into many different areas. Old schemes for dividing up the areas just used numbering. These numbers have been handed down to us from epic studies by unfathomably dedicated neuroanatomists in the early twentieth century. They finely sliced up brains, stained the slices to pick out the neurons, and then minutely examined the changes in the density, size, and shape of neurons across the brains. Each place where there was a change in one of those things got a number. The temptation is to then think that these subtle differences in neurons correspond to different functions. We followed that temptation by turning many of those numbers into names: primary motor cortex; secondary motor cortex; primary somatosensory cortex; visual cortex. But in the human cortex Brodmann found 52 areas. von Economo and Koskinas found 107. Hmmm. One may reasonably ask: if just dividing up the cortex by changes in the shape, size, or density of neurons actually tells us something about what the different areas do, how come the numbers of areas were so different? A recent massive fMRI study used a bunch of ways to isolate different parts of cortex, and ended up with 180 different regions. Far more then the slice-and-stain guys. Which is right? None of them. Those lovely differences in cells are immaterial to actual function. Any given function we care to name — seeing, jumping, talking, putting our DVDs in thematic order because we’re not culturally illiterate barbarians — engages many areas of cortex, and many areas of the brain, at the same time. The names and numbers are misleading, are irrelevant to how evolution has driven the wiring of the brain to perform a function. Cortex has names for its layers too. Names straight out of a Dan Brown novel: 1, 2, 3, 4, 5 and — wait for it — 6. In rodents, some areas have six layers, and some have five because they are missing layer 4. And as the textbooks will tell you, it stands to reason that motor cortex and somatosensory cortex have different functions because motor cortex doesn’t have a layer 4. This is an object lesson in the problems of naming: we could not name a clear cell “layer 4” in motor cortex, because we couldn’t see one. But there is one, and it’s been there all along. What would you bet me that it will turn out the rat prefrontal cortex does have a “layer 4” after all — a set of neurons that are specifically targeted by input from thalamus and then connect to other neurons in the cortex without projecting outwards — just not in a distinct layer we can see by colouring in neurons? Evolution doesn’t care that we can name five layers in one region, or six in another. In only cares what they do to keep the animal alive. To evolution, the brain is just a gigantic bag of cells, wired together. The purpose of that gigantic bag of cells is to contribute to the survival of the organism in which it resides, to surviving long enough to reproduce. Those that reproduce, win; those that don’t, don’t. If a random mutation causes or changes the wiring of some neurons to another group of neurons, and that mutation improves the chances of having offspring, it will likely spread through the population. If that random mutation adds connections from prefrontal cortex to visual cortex; from cortical interneurons to a structure outside cortex; from motor cortex to midbrain dopamine neurons, then it will happen. Evolution will not feel sorry that it’s just ruined another set of textbooks. And it is just a giant bag of cells wired together. Our best evidence that it is not — that we can cling to our names of all the bits — come from studies where we cut a bit out or turn a bit off. When we cut out area X and we see a “deficit” in behaviour Y of an animal (like tying its shoelaces), then we think “aha! Area X is for tying shoelaces”. No. For starters, we never see a complete and permanent end to behaviour Y. We normally see that the animal is simply worse at doing or learning Y — not that it cannot do Y at all. The brain can carry on doing Y just fine, thanks, just not as well — there is massive redundancy in the brain. Like what you’d find if it was a giant bag of cells, wired together. Moreover, seeing that behaviour Y gets worse logically tells us little about what area X is actually doing. It just tells us that damaging area X causes problems. Which on reflection isn’t surprising as you just ripped a chunk out of the brain. The logical fallacy is simple to demonstrate. I am right now going to make a new startling new prediction of how the mouse brain works: the ventrolateral medulla is necessary for mice to learn to associate pictures of Benedict Cumberbatch with food. Cut out the ventrolateral medulla, and a mouse will not learn to associate pictures of Benedict Cumberbatch with food. Because it will be dead. The ventrolateral medulla contains the neurons which control the rhythm of breathing. Cut it out: no breathing. Ergo, no learning. Is the ventrolateral medulla a crucial brain area for learning? No. But by damaging it, we damage something vital to the process of learning. Thus, cutting a bit that we’ve given a name can have an effect on that named thing, and we learn nothing at all. Except that we have damaged a big bag of cells, wired together. What’s more remarkable is when cutting bits out has no effect. If we cut some bits of brain out before learning we see an effect — learning is made slower or worse or both; but when we cut them out after learning, it has no apparent effect whatsoever. These bits of brain have become completely redundant. Again: giant bag of cells, wired together — there are many ways within that bag of cells to solve the problem at hand, enough for the brain to just stop using a bit of itself altogether. And people get hyper-excited about finding odd signals where they were utterly unexpected. Like finding that reward changes activity in primary visual cortex. Or finding that sound is encoded in the hippocampus. These are great, interesting studies. But to be surprised by them is to fall into the naming fallacy. Evolution does not know nor care that we called this chunk of neurons the “hippocampus”. It is just another bag of neurons, connected to many other bags of neurons. Names. Names are vital. Names mean we can all be sure we’re talking about the same thing. They identify whopping great phalanxes of neural tissue; they identify individual neurons; they let us compare brains of different creatures. But we have to handle them with care, less we confuse the label with the contents. Evolution only cares about the contents. And development doesn’t give a damn what we call a brain region either. But that’s a tale for another time. Want more? Follow us at The Spike Twitter: @markdhumphries

Evolution doesn’t give a damn what you think a brain region is called

3,637

evolution-doesnt-give-a-damn-what-you-think-a-brain-region-is-called-1c9c5f80ab89

2018-06-10

2018-06-10 00:25:36

https://medium.com/s/story/evolution-doesnt-give-a-damn-what-you-think-a-brain-region-is-called-1c9c5f80ab89

false

1,890

The science of the brain, from the scientists of the brain

null

The Spike

thespike.editors@gmail.com

the-spike

NEUROSCIENCE,SCIENCE,DATA SCIENCE,PSYCHOLOGY,ARTIFICIAL INTELLIGENCE

markdhumphries

Neuroscience

neuroscience

Neuroscience

6,742

Mark Humphries

Uses his brain to understand brains. Is that possible? Neuroscience: https://humphries-lab.org

6bce9fc81f1f

drmdhumphries

25,155

18

20,181,104

null

0

null

0

null

2018-06-08

2018-06-08 08:38:13

2018-06-08

2018-06-08 08:44:12

1

false

en

2018-06-08

2018-06-08 08:44:12

3

1c9d184d0f1

1.109434

1

0

An assistant professor at the University of Nebraska-Lincoln (UNL) is developing drone software to ease people’s discomfort around the…

5

Professor Creating Software To Ease People’s Discomfort Around The Technology An assistant professor at the University of Nebraska-Lincoln (UNL) is developing drone software to ease people’s discomfort around the technology. “As we look toward the future and new users for drones, they’re going to need to be able to intelligently interact with people who aren’t controlling them,” says Brittany Duncan, an assistant professor of computer science and engineering at UNL. The development process has to go through certain stages; first the team will conduct surveys to understand what people want drones to communicate and the drone movement, later the researchers will search for commonalities and develop software that allows drones to make those movements. “The idea is that this could be the next optional software, it tells people the things they should be concerned about with drones and makes you a more conscientious drone operator,” says Duncan. Source: https://bit.ly/2sGb2Q6 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.

Professor Creating Software To Ease People’s Discomfort Around The Technology

43

professor-creating-software-to-ease-peoples-discomfort-around-the-technology-1c9d184d0f1

2018-06-16

2018-06-16 15:58:14

https://medium.com/s/story/professor-creating-software-to-ease-peoples-discomfort-around-the-technology-1c9d184d0f1

false

241

null

Deepaero

deepaeros

Deepaero

0

DEEP AERO DRONES

null

dcef5da6c7fa

deepaerodrones

277

0

20,181,104

null

0

null

0

bfceef4de732

2018-02-12

2018-02-12 15:35:54

2018-02-15

2018-02-15 13:05:35

4

false

en

2018-04-22

2018-04-22 15:55:53

18

1c9e0d9a82d8

8.930189

25

2

0

Statistical significance and hypothesis testing are not really helpful when it comes to testing our hypotheses.

5

Inferential Statistics is not Inferential Statistical significance and hypothesis testing are not really helpful when it comes to testing our hypotheses. The earth is flat (p > 0.05). I confess. Throughout my scientific life, I have used a method that I knew or felt was deeply flawed. What’s more, I admit to have taught — and I do still teach — this method to my students. I have a number of questionable excuses for that. For example, because the method has shaped a big part of science in the last century, I think students ought to know about it. But I have increasingly come to believe that science was and is largely a story of success in spite of, and not because of, the use of this method. The method is called inferential statistics. Or more precisely, hypothesis testing. The method I consider flawed and deleterious involves taking sample data, then applying some mathematical procedure, and taking the result of that procedure as showing whether or not a hypothesis about a larger population is correct. Now, if you are familiar with the current debates about this method, you might think: “Somebody is yet again blaming p-values for everything.” But no, I am not. The p-value is a statistic that has the great advantage of being easily applied. Unfortunately, it seems that everybody I know, including myself, has used it in the wrong way. But let me start by reporting a story about what a p-value is capable of doing. This story was told, at least in part and probably not for the first time, in a blog post by neuroscientist Ulrich Dirnagl. What went wrong? In 2011, researchers at CERN worked on the so-called OPERA experiment and sent neutrinos through the Alps to be detected in central Italy. The neutrinos were found to be faster than light, even when the experiment was repeated. This was surprising, to say the least, and the p-value attached to the observation was smaller than the alpha level of p=0.0000003 that is required to announce a discovery in particle physics experiments involving collision data. Although the researchers made clear that they were still searching for possible unknown systematic effects that might explain the finding, the news hit the media as: “Was Einstein wrong?” A few months later, the researchers announced the explanation for the surprising measurements: a cable had not been fully screwed in during data collection. Statistical models usually include an implicit assumption that all cable connections are correct. (Image: State Farm via flickr.com) Does that mean p-values are unreliable? No, it means that we should not make inferential decisions based on p-values. Indeed, the p-value in the OPERA experiment was correct. As Sander Greenland explains, we should think of a p-value as referring not only to the null hypothesis but to the entire model it was computed from, including all assumptions such as that there were no measurement errors. A small p-value indicates that something is wrong with the model, but it does not indicate what is wrong. The original OPERA model included the null hypothesis “neutrinos are not faster than light.” However, it also included an assumption that the equipment was in perfect working order. As indicated by the extremely small p-value, the original model had a problem. But the researchers did a good job in finding an additional explanatory variable, and the new model — including the loose cable — successfully explained the observation that neutrinos appeared to travel faster than light. You draw the conclusions Of course, everybody who thought that p-values are about null hypotheses, or even about alternative hypotheses, must now be disappointed. Yes, a small p-value can mean that the null hypothesis is false. But it can also mean that some mathematical aspect of the model was not correctly specified, or that we accidentally switched the names of some factor levels, or that we unintentionally — or intentionally — selected analyses that led to a small p-value (“p-hacking”), or that a cable was loose. Statistics cannot be inferential. It must be we who make the inference. As Boring (1919) put it one century ago: “Conclusions must ultimately be left to the scientific intuition of the experimenter and his public.” But, interestingly, “one can feel widespread anxiety surrounding the exercise of informed personal judgment” (Gigerenzer 1993). People seem to mistrust inference by humans and to long for “objective” inferential decisions made by computer algorithms, based on data. And I can see there is a reason for this desire because, apparently, some human experts tend to make claims as part of their political agenda without bothering about data. Of course, personal judgment does not mean that anything goes. It means that if we have evidence that something could possibly be wrong with a model — if, for example, we found a small p-value — then we must apply informed personal judgment to try and find out what is wrong. The explanation may be that some alternative scientific hypothesis is correct, but there are many more things to consider. In the neutrinos-faster-than-light scenario outlined above, the explanation was neither to be found in the data nor in scientific theory, nor in statistics. My guess is that the correct inference was made by somebody applying scientific intuition or informed personal judgment when checking whether perhaps a cable was loose. But don’t we do this all the time in our talks and papers: challenge our statistical results by discussing alternative explanations? Well, there’s a lot to say about that. In biology, for example, which is my field of research, I rarely see papers by scientists who introduce and discuss their favored hypothesis only to conclude that their hypothesis seems to be wrong after all. In our discussion sections, it is usually the alternative explanations that are refuted. The earth is flat (p > 0.05) Scientists and journals and newspaper readers want support for hypotheses, not refutation (unless the hypothesis to be refuted is as famous as Einstein’s special theory of relativity). And in our talks and papers, p-values are used to make decisions: about which effect is reliable and which is not, whether an effect is zero or not, and which result is worth being interpreted or published and which is not. The sky is round in this wood engraving from 1888. None of those decisions can be soundly justified based on p-values alone. And basing decisions only on statistical tests can be extremely harmful. If you are interested in the details, you may want to check, for example, our review “The earth is flat (p > 0.05): significance thresholds and the crisis of unreplicable research” and some of the 212 cited references. Don’t blame the p-value In short, if we select results to be interpreted or published because a statistic such as the p-value crosses some significance threshold, our conclusions will be wrong. One reason is that significant effect sizes are usually biased upwards. And if we interpret a larger p-value, or p > 0.05, as support for the null hypothesis (“the earth is flat,” or “there was no difference”) — something that happens in almost every scientific talk — then we fall into the trap of the most devastating of all false beliefs about statistical inference that can potentially cost human lives. And even if our alternative hypothesis and all other assumptions are correct, the p-value in the next sample will probably differ strongly from our current sample. If you don’t believe it, have a look at the “Dance of the p-values” by Geoff Cumming, or read why “The fickle P value generates irreproducible results.” But again, don’t blame the p-value. The p-value itself is not unreliable — its fickleness reliably indicates variation in the data from sample to sample. If sample averages vary among samples, then p-values will vary as well, because they are calculated from sample averages. And we don’t usually take a single sample average and announce it to be the truth. So if we take a single p-value to decide which hypothesis is wrong and which is right, this is our fault, not the fault of the p-value. Almost no science without statistics Do I recommend completely abandoning statistics? Of course not. In many fields of research, there is almost no science without statistics. We need statistics to describe the signal and to describe the noise. But statistics cannot tell us whether a hypothesis is true or false. If you want to use statistical power and error rates for getting a rough idea about suitable sample sizes, fine. But if somebody argues that in hypothesis testing, we can “control” error rates if only we justify our alpha level in advance of a study, I don’t believe it. (If somebody missed “the saga of the summer 2017, a.k.a. ‘the alpha wars’,” check it out!). If a cable is loose, your false positive rate may be near 100%, and the best error control is to find and fix the cable. I agree with Andrew Gelman, who says “Oooh, I hate all talk of false positive, false negative, false discovery, etc.” All this also means that Bayesian statistics are no cure for the core problem, because those statistics rely on the very same assumptions about the experimental set-up that the traditional methods use. Thus, if the Bayesian analysis assumes implicitly “no loose cable” (as it would in practice), the posterior probabilities it produces will be as misleading as a significance test could be. I like Bayesian statistics, but I don’t believe in inferential probabilities because maybe a cable was loose. I do believe in descriptive probabilities, though, and that’s what a p-value can offer: the probability of our observed set of data, and of data more extreme, given that our current model is true. There is no inferential meaning attached to that. For the next set of data, the p-value will be different. A small p-value is just a warning signal that our current model could be wrong, so we should check whether a cable is loose. And a large p-value does not at all mean that the null hypothesis is true or that the current model is sound. So here is what I will do. I know it is hard, but I will try not to base presentation decisions on p-values, nor on any other statistical result. I promise I will never again ask my students to set value on “significant” results (if I do, show me this blog post). If I found that on average, males were larger than females, I will report it, because that can hardly be wrong: I did indeed find that on average, my sampled males were larger than my sampled females (I did my best, but of course, the difference could be the result of my faulty measuring device). I will then interpret the confidence interval as a “compatibility interval,” showing alternative true effect sizes that could, perhaps, be compatible with the data (if every assumption is correct and my measuring device is not faulty). If you want to see examples of that, check figure 1 in our review. Yes, I admit, the confidence interval gets used as an inferential statistic. It can even be used as a significance test, and in most cases, people misuse it that way. But no, the probability that the true value lies within our 95%-confidence interval is not 95%. And no, it should play no role whatsoever whether zero is included or excluded in the interval, because even if every assumption is correct, the dance of the confidence intervals shows that with other sets of data, the interval will be very different. The true effect size could easily be outside our particular interval. But confidence intervals can be used to get a rough feeling for how large uncertainty could be in the ideal case, given that all assumptions and cable connections are correct. They visualize that many other hypotheses are compatible with the data. So we should not call them confidence intervals but uncertainty intervals, or compatibility intervals (Sander Greenland says they are really overconfidence intervals). This is what I would like inferential statistics to do: help us recognize that our inferences are uncertain; hint at possible alternative hypotheses; and show that science is not about being confident, nor about making decisions, but about “the evaluation of the cumulative evidence and assessment of whether it is susceptible to major biases.” – See also the following preprint: Amrhein V, Trafimow D, Greenland S. (2018) Abandon statistical inference. PeerJ Preprints 6:e26857v1. For comments and discussions, I thank Sander Greenland, Fränzi Korner-Nievergelt, Tobias Roth, and David Trafimow. To show your support for this post and recommend it to your followers, click on the clap icon 👏 below. Each users is allowed to clap up to 50 times to show how much they appreciated a story. The University of Basel has an international reputation of outstanding achievements in research and teaching. Founded in 1460, the University of Basel is the oldest university in Switzerland and has a history of success going back over 550 years. Learn more

Inferential Statistics is not Inferential

307

inferential-statistics-is-not-inferential-1c9e0d9a82d8

2018-04-22

2018-04-22 16:37:21

https://medium.com/s/story/inferential-statistics-is-not-inferential-1c9e0d9a82d8

false

2,181

Contact

null

unibasel

null

sci five | University of Basel

kommunikation@unibas.ch

sci-five-university-of-basel

RESEARCH,PHD,ACADEMIA,POSTDOCTORAL,SCIENCE COMMUNICATION

UniBasel_en

Data Science

data-science

Data Science

33,617

Valentin Amrhein

Professor of Zoology, University of Basel, www.camargue.unibas.ch

5d2530aa97a7

v.amrhein

28

2

20,181,104

null

0

null

0

null

2018-09-05

2018-09-05 21:17:35

2018-09-05

2018-09-05 22:04:59

1

false

en

2018-09-05

2018-09-05 22:05:36

4

1c9e296d2b58

1.566038

0

Almost all NLP engineers now use Python or Java because of the open source NLP toolkits such as NLTK, CoreNLP, SpaCy and OpenNLP, as well…

2

Text Normalization of NLP in .NET Almost all NLP engineers now use Python or Java because of the open source NLP toolkits such as NLTK, CoreNLP, SpaCy and OpenNLP, as well as machine learning algorithm libraries like Scikit-Learn. If you are a .NET developer, it is very difficult to use C# to do some NLP work. Although there are also Microsoft open source libraries like ML.NET, it is not a specialized NLP toolkit. Many natural language processing-specific libraries are not found in them. So I decided to start my own NLP toolkit from scratch. I will implement some common functions and model algorithms completely with C#. It is not easy to implement each step of NLP Pipeline from scratch. I am planning to write a series of blogs to record this project. I am not sure I can complete the whole project, but I have been researching this field for two years in NLP. I will continue to learn and implement in code in this field. Normalizing text means converting it to a more convenient, standard form. Most of what we are going to do with NLP relies on separating out or tokenization for text. English words are often separated from each other by whitespace, but whitespace is not always sufficient. New York and rock ’n’ roll are sometimes treated as large words despite the fact that they contain spaces, while sometimes we’ll need to separate I’m into the two words I and am. For processing tweets or texts we’ll need to tokenize emoticons like :) or hashtags like #nlproc. Some languages, like Chinese, don’t have spaces between words, so word tokenization becomes more difficult. One commonly used tokenization standard is known as the Penn Treebank toPenn Treebank tokenization kenization standard. There is original implementation here, and a Python implemented here , and Java from here. No C# implementation? Not true. You can find here. The Treebank tokenizer uses regular expressions to tokenize text as in Penn Treebank standard. I ran it and screenshot below: It seems to run very well. I really like the result with the start pos and end pos, there is no such function in NLTK.

Text Normalization of NLP in .NET

0

text-normalization-of-nlp-in-net-1c9e296d2b58

2018-09-05

2018-09-05 22:05:36

https://medium.com/s/story/text-normalization-of-nlp-in-net-1c9e296d2b58

false

362

null

Machine Learning

machine-learning

Machine Learning

51,320

Haiping Chen

null

898fc824568e

haiping008

1

3

20,181,104

null

0

{‘address’: ‘1001 Eastern Avenue’, ‘city’: ‘Toronto’, ‘business_id’: ‘8bskEwLJsPe2X3Avixab-Q’, ‘stars’: 4.0, ... ‘is_open’: 1, ‘categories’: [‘Tapas/Small Plates’, ‘Nightlife’, ‘Food’, ‘Restaurants’, ‘Beer Bar’, ‘Bars’, ‘Breweries’], ‘review_count’: 9, ... ‘inspection_data’: [{‘date’: ‘2017–05–05 00:00:00’, ‘result_significance’: ‘M — Minor’, ‘InsNum’: ‘0’, ‘relevant_reviews’: [‘Amazing giant brewery with terrific owners with an excellent view of the lake and free parking…], ‘result_notice’: ‘Notice to Comply’, ‘fail’: 0}] ... } ‘categories’: [‘Tapas/Small Plates’, ‘Nightlife’, ‘Food’] ╔═════════════════════════════╦═══════════╦════════════════════╗ ║ Feature Name ║ Chi2 ║ Cor. With "Fail" ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ attributes.NoiseLevel_quiet ║ 8.929125 ║ 0.060680 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ categories.Butcher ║ 7.654189 ║ 0.053185 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ categories.Halal ║ 7.379606 ║ 0.052455 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ categories.Hawaiian ║ 7.713115 ║ 0.053428 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ categories.Hotels ║ 6.938719 ║ -0.051248 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ categories.Hotels & Travel ║ 7.393479 ║ -0.052920 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ neighborhood_Chinatown ║ 7.847233 ║ 0.054541 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ neighborhood_Etobicoke ║ 11.147838 ║ -0.065180 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ neighborhood_Willowdale ║ 7.723131 ║ 0.053823 ║ ╠═════════════════════════════╬═══════════╬════════════════════╣ ║ Inspection_Number6 ║ 6.793666 ║ -0.050330 ║ ╚═════════════════════════════╩═══════════╩════════════════════╝ LDA Topics: Reviews for Failed Inspections Restaurants Topic 1 jerk shawarma manish breast jamaican poutine caribbean craftsmanship ized reflux Topic 2 museum exhibits good displays exhibit children interactive collection egyptian experience Topic 3 dim sum store yelp rol san 2014 reviewer mai improved Topic 4 cappuccino measure sourness velvety burnt standard food rich place service Topic 5 food good place great service like really just time ordered LDA Topics: Reviews for Passed Inspections Restaurants Topic 1 food good place great service like really just time ordered Topic 2 ramen great time room burger toronto just place hotel game Topic 3 deli cuban sandwiches portuguese st smoked di latkes cruise love Topic 4 cream cheesecake ice line tea like cake just wait good Topic 5 market lawrence produce que saturdays meats bacon markets peameal shops LDA Topics: Reviews for Etobicoke Restaurants Topic 1 pies pie just really time good crust order filling sweet Topic 2 banh mi beef good delicious odd sandwich cilantro paella cupcakes Topic 3 greatly neelam profiles sizes chill consumer hip tweeked changed improved Topic 4 food good place service great really ordered delicious restaurant breakfast Topic 5 beef restaurant spring came gelato shrimp thai trip make salad LDA Topics: Reviews for Quiet Restaurants Topic 1 mind espresso wall designed cute latte excellent mural machine theme Topic 2 good place just great food chicken service really nice don Topic 3 food place good okay service just like sushi chicken really Topic 4 lobster bruschetta linguine marianne steve mash anniversary bisque tummies smiling Topic 5 canada concerts tso roy indigenous honour celebration thomson 150th hall +--------------+-----------------+-----------------+ | | Predicted: Fail | Predicted: Pass | +--------------+-----------------+-----------------+ | Actual: Fail | a | b | +--------------+-----------------+-----------------+ | Actual: Pass | c | d | +--------------+-----------------+-----------------+ Metrics and intuitive explanations, with respect to the "fail" class. - Precision (Fail Class) = a / (a+c) | Intuition: The proportion of inspections predicted to result in a fail that actually resulted in a fail - Recall (Fail Class) = Sensitivity = a / (a+b) = | Intuition: The proportion of actual fails my model correctly predicted as fails (True Positive Rate) ~ "hit rate" - F = 2*(Precision * Recall) / (Precision + Recall) | Intuition: a harmonic average between precision and recall. - Brier Score Loss = Probability Reliability = 1/n * [(p_t - o_t)^2] , where p_t is the predicted probability of a failure at time t; o_t is 0 if an inspection passed and 1 if failed; n is number of predictions. Intuition: Brier score measures how reliable probabilities from a model are, with 0 being the best value and 1 being the worst. If I am 90% sure it will rain and it doesn't, my probabilities are unreliable and would result in a Brier score of (0.90-0)^2 = 0.81 (very bad). Likewise, if I was only 10% certain it would rain and it did rain, my estimation of the likelihood of rain was bad - and so I'd also have a brier score of (0.10 - 1)^2 = 0.81 (very bad). If I was bashful and only 10% confident it would rain - and in fact, it did not rain - my probabilities aren't so bad: My Brier score is (0.10-0)^2 = 0.1 (good). # X is a Pandas dataframe consisting of text data and categorical + ordinal feature data. svc_txt = Pipeline([('col_selector', ItemSelector(key=[col for col in X.columns if col in text_cols])), ('stdize_feats', StandardScaler()), ('calib_probs', CalibratedClassifierCV (cv=5, method='isotonic', base_estimator = SVC( class_weight='balanced', max_iter=2000, kernel='sigmoid', C = 18, gamma = 0.3 )))]) lr_txt = Pipeline([('col_selector', ItemSelector(key=[col for col in X.columns if col in text_cols])), ('reduce_dims', PCA(n_components=80, random_state=42)), ('stdize_feats', StandardScaler()), ('calib_probs', CalibratedClassifierCV (cv=5, method='isotonic', base_estimator = LogisticRegression( random_state=42, warm_start=True, C = 10, class_weight='balanced', solver="newton-cg", penalty="l2", )))]) lr_feat = Pipeline([('col_selector', ItemSelector(key=[col for col in X.columns if col in feature_cols])), ('reduce_dims', PCA(n_components=95)), ('stdize_feats', StandardScaler()), ('calib_probs', CalibratedClassifierCV (cv=5, method='isotonic', base_estimator = LogisticRegression( class_weight='balanced', C = 0.06, warm_start=True, random_state=42, )))]) rf_feat = Pipeline([('col_selector', ItemSelector([col for col in X.columns if col in feature_cols])), ('stdize_feats', StandardScaler()), ('calib_probs', CalibratedClassifierCV (cv=5, method='isotonic', base_estimator = RandomForestClassifier( class_weight='balanced', random_state=42, max_features="sqrt", max_depth= 10, n_estimators=1000 )))]) r * minority_count = majority_count r = majority_count / minority_count # Base estimators are individual estimators. Both txt_voter and feat_voter use the predictions of two constituent base estimators to yield a final probability. Master voter then uses the argmax of the probabilities from feat_voter and txt_voter to make a final prediction. base_estimators = {"feat_clfs":[('rf_feat', rf_feat), ('lr_feat', lr_feat)], "txt_clfs": [('rf_txt', lr_txt), ('svc_txt', svc_txt)]} txt_voter = VotingClassifier(estimators=base_estimators['txt_clfs'],voting="soft", weights = [1,1]) feat_voter = VotingClassifier(estimators=base_estimators['feat_clfs'],voting="soft", weights = [1,1]) master_voter = VotingClassifier(estimators=[('txt_voter', txt_voter), ('feat_voter', feat_voter)], voting='soft', weights= [2,1]) ╔═══════════════════════════════════════════════════════════════╗ ║ Master Voter ║ ╠══════════════════╦═══════════╦════════╦═════════╦═════════════╣ ║ ║ Precision ║ Recall ║ F-Score ║ Brier Score ║ ╠══════════════════╬═══════════╬════════╬═════════╬═════════════╣ ║ Pass ║ 0.83 ║ 0.75 ║ 0.67 ║ 0.16 ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Fail ║ 0.30 ║ 0.70 ║ 0.35 ║ ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Weighted Average ║ 0.73 ║ 0.67 ║ 0.68 ║ ║ ╚══════════════════╩═══════════╩════════╩═════════╩═════════════╝ Overall, Master Voter had a greater balance between precision and recall than Feat Voter or Text Voter, and a lower Brier score. ╔═══════════════════════════════════════════════════════════════╗ ║ Feature Voter ║ ╠══════════════════╦═══════════╦════════╦═════════╦═════════════╣ ║ ║ Precision ║ Recall ║ F-Score ║ Brier Score ║ ╠══════════════════╬═══════════╬════════╬═════════╬═════════════╣ ║ Pass ║ 0.79 ║ 0.44 ║ 0.56 ║ 0.19 ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Fail ║ 0.32 ║ 0.60 ║ 0.41 ║ ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Weighted Average ║ 0.69 ║ 0.44 ║ 0.53 ║ ║ ╚══════════════════╩═══════════╩════════╩═════════╩═════════════╝ ╔═══════════════════════════════════════════════════════════════╗ ║ Text Voter ║ ╠══════════════════╦═══════════╦════════╦═════════╦═════════════╣ ║ ║ Precision ║ Recall ║ F-Score ║ Brier Score ║ ╠══════════════════╬═══════════╬════════╬═════════╬═════════════╣ ║ Pass ║ 0.84 ║ 0.09 ║ 0.16 ║ 0.18 ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Fail ║ 0.20 ║ 0.94 ║ 0.32 ║ ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Weighted Average ║ 0.72 ║ 0.26 ║ 0.20 ║ ║ ╚══════════════════╩═══════════╩════════╩═════════╩═════════════╝ ╔═══════════════════════════════════════════════════════════════╗ ║ SVM (Text) ║ ╠══════════════════╦═══════════╦════════╦═════════╦═════════════╣ ║ ║ Precision ║ Recall ║ F-Score ║ Brier Score ║ ╠══════════════════╬═══════════╬════════╬═════════╬═════════════╣ ║ Pass ║ 0.85 ║ 0.17 ║ 0.28 ║ 0.18 ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Fail ║ 0.22 ║ 0.89 ║ 0.36 ║ ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Weighted Average ║ 0.72 ║ 0.32 ║ 0.30 ║ ║ ╚══════════════════╩═══════════╩════════╩═════════╩═════════════╝ ╔═══════════════════════════════════════════════════════════════╗ ║ LR (Text) ║ ╠══════════════════╦═══════════╦════════╦═════════╦═════════════╣ ║ ║ Precision ║ Recall ║ F-Score ║ Brier Score ║ ╠══════════════════╬═══════════╬════════╬═════════╬═════════════╣ ║ Pass ║ 0.84 ║ 0.56 ║ 0.67 ║ 0.19 ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Fail ║ 0.25 ║ 0.57 ║ 0.34 ║ ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Weighted Average ║ 0.72 ║ 0.56 ║ 0.64 ║ ║ ╚══════════════════╩═══════════╩════════╩═════════╩═════════════╝ ╔═══════════════════════════════════════════════════════════════╗ ║ RF (Feat) ║ ╠══════════════════╦═══════════╦════════╦═════════╦═════════════╣ ║ ║ Precision ║ Recall ║ F-Score ║ Brier Score ║ ╠══════════════════╬═══════════╬════════╬═════════╬═════════════╣ ║ Pass ║ 0.80 ║ 0.56 ║ 0.81 ║ 0.18 ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Fail ║ 0.23 ║ 0.57 ║ 0.22 ║ ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Weighted Average ║ 0.68 ║ 0.70 ║ 0.69 ║ ║ ╚══════════════════╩═══════════╩════════╩═════════╩═════════════╝ ╔═══════════════════════════════════════════════════════════════╗ ║ LR (Feat) ║ ╠══════════════════╦═══════════╦════════╦═════════╦═════════════╣ ║ ║ Precision ║ Recall ║ F-Score ║ Brier Score ║ ╠══════════════════╬═══════════╬════════╬═════════╬═════════════╣ ║ Pass ║ 0.83 ║ 0.44 ║ 0.57 ║ 0.17 ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Fail ║ 0.23 ║ 0.66 ║ 0.34 ║ ║ ╠══════════════════╬═══════════╬════════╬═════════╣ ║ ║ Weighted Average ║ 0.71 ║ 0.48 ║ 0.52 ║ ║ ╚══════════════════╩═══════════╩════════╩═════════╩═════════════╝

34

null

2017-12-30

2017-12-30 01:40:31

2017-12-30

2017-12-30 03:46:33

15

false

en

2018-01-04

2018-01-04 18:31:11

13

1ca5b82ef06f

17.318868

3

0

(Yelp 2017 dataset challenge entry)

5

Using Yelp data to predict restaurant inspection failure: a ‘voter of voters’ ensemble approach. (Yelp 2017 dataset challenge entry) Link to python files I used Yelp restaurant reviews and information to try to predict whether a restaurant will fail food safety inspections. The classification proved to be difficult, but it was a fun project. Findings: If you are looking for three dimensions to summarize restaurant variance: Badness, casualness, uptightness. Cuisine, restaurant location, perceived popularity, noise level, and centrality of meat, all affect inspection outcomes. My final model involved a customized SMOTE Synthetic Minority Over-Sampling Technique) driven ensemble: I divided a restaurant’s attributes into two components, creating an ensemble to vote on whether a restaurant will fail/pass inspections for each component, and then creating a master ensemble to integrate those two predictions. Here, I’ll walk through my process and results, at a level that requires minimal machine learning experience. Part I goes over the data gathering process. Part II walks through some descriptive analysis. Part III summarizes my predictive approach. Each section has a tl;dr so feel free to skip around. I. Data Gathering tl;dr: I looked at about 2000 restaurants in Toronto who had at least one Yelp review no more than a month before an inspection. What counts as a failed inspection? DineSafe is Toronto Public Health’s food safety program that “inspects all establishments serving and preparing food.” If a restaurant violates food safety standards, these violations are classified into one of three categories: (1) minor — minimal health risk (not wearing a hairnet), (2) significant — potential health hazard (dirty kitchen counter), (3) crucial — immanent health hazard (sewage backup). I counted an infraction in the last two categories as “failing” an inspection. How were restaurants selected? I looked at restaurants who were inspected from October 2015 to October 2017, using this feed. I also used the business.json and reviews.json files from the Yelp academic dataset. First, I used Yelp’s API to associate each restaurant in the inspection dataset with a unique Yelp business id — giving me Yelp attribute data. Then, I filtered my dataset so it included only those restaurants who had at least one Yelp review in the reviews.json file within less than 30 days of a restaurant inspection. I ended up with a list of json objects — each object containing (1) inspection data for each observed inspection and (2) restaurant information info for a given restaurant. Note: you can find the full json file here. II. Descriptive Analysis tl;dr: Failure is pretty rare, and it’s hard to seperate — from just a descriptive analysis — which inspections result in failure and which don’t. However, meats, neighborhoods, and noise level are relevant to outcome. While restaurants are inherently multi-dimensional, three significant principal components of variation are badness, casualness, uptightness. How common is failure? Failing a restaurant inspection is rare. First, failing a restaurant inspection is very uncommon. There were 1706 restaurants in the dataset, and 2708 inspection instances .But the classes were highly imbalanced because only 561 (20%) of the inspection instances resulted in failure. What kind of restaurants fail inspections? First, we might think that restaurants providing a lot of ‘value’ — high in stars, low in price — are cutting corners with food safety. But the price-star relationship is pretty similar for restaurants who failed or passed. The price-star relationship does not differ significantly for restaurants versus those who passed inspections. Maybe it is the case that worse rated restaurants or cheaper restaurants fail inspections? But if you look at the star or price distributions by inspection result, they are, in general, quite similar. Intuitively, this isn’t very surprising: You can love a dive bar even if health inspectors don’t. Insights from 530 Dimensions I realized that I would need to look at much more specific aspects of restaurants to understand which restaurants will fail inspections. This was possible because of one-hot encoding, which is when one takes a column for a categorical feature, and then creates n-1 columns for the n-1 values of that feature; each new column consists of 1s or 0s depending on the presence or absence of this feature. So for the above example, that restaurant had the following key-value pair: For machine learning models, it is often necessary to make each individual value of categories (tapas, nightlife, food) its own column so we can translate strings to binary values. One-hot encoding all categorical features led to a 530-dimensional matrix. Here are some of the most important links between specific features and restaurant failure, using Chi-Square as a metric. At a high level, a large Chi-Squared statistic implies a high probability that the occurrence of a feature and the occurrence of a class are dependent events. With two classes, the 5% critical value for a Chi-Square statistic is 6. So at the 5 percent level of significance, for each of these features: We can reject the null hypothesis that the occurrence of the features and class outcome are independent events. So what might affect restaurant failure? First, butchering, halal, and Hawaiian food all involve a lot of meat. We might suspect that there are simply many violations that deal with meat handling, so restaurants who handle meats — and have not been shut down yet — are more diligent about their food safety practices. Second, failure looks to be in part geographically determined. But without knowing much about Toronto, I won’t hypothesize about what Chinatown, Etobicoke, and Willowdale might have to do with restaurant inspection failure (though text analysis can help). Third, we see that it being a restaurant’s 6th inspection might be important. Like algorithms, managers, too, learn from labeled data (ie: prior inspection results). Finally, insofar as noise level plays a role in inspection results, this is most likely by proxy: quiet restaurants are likely to share a lot of other attributes, which might make them more or less likely to fail or pass inspections. 3 Axes of Restaurants: Badness, Casualness, Uptightness 530 dimensions is too many. Let’s try to break our data down using Principal Component Analysis to see what general ‘types’ of restaurants there are. The first principal component can be called ‘generic badness’. It is primarily characterized by low star ratings and also a lack of free wifi. The second principal component is can be thought of as ‘casualness’. It’s characterized by serving late night meals, fast food, and being located in shopping centers. There’s a clear U-shape relationship between badness and casualness : Restaurants rated very highly, and also restaurants rated very poorly, tend to be casual. The third principal component can be thought of as ‘uptightness’: no alcohol, no takeout, not for kids. Gray dots are restaurant failure class, black dots are restaurant pass class. Passing restaurants are more “uptight” (PC 3) than failing restaurants. Unfortunately, the data does not look very separable from this point of view: At least among the three principal components of variation shown here, there is a lot of overlap between the two classes (with the exception of a few inspection-passsing class members scoring very high on the “uptight” component, component 3). Topic Modeling: Meta Topics, Pies, Gentrification Since I also have text data, it’s worth seeing if the classes are separable from reviews. First, I used t-distributed stochastic neighbor embedding ( TSNE) to try to visualize the text reviews in just two dimensions. TSNE reduces the dimensions of high-dimensional data (like text documents) by decomposing text vectors into 2 dimensions — using the probability distributions from the original dimensionality and the decomposed dimensionality. In my case, though, it wasn’t very helpful because there was still substantial overlap even after hyperparameter tuning. Next, I wanted to look at the different topics that are brought up in reviews close to inspection passes and inspection fails, and see what is going on with the largest Chi-Squared attributes: quiet restaurants and Etobicoke restaurants. To do so, I used SciKit Learn’s Latent Dirichlet Allocation(LDA) implementation. In natural language processing, LDA is a method of modeling topics in documents: It allows sets of observed data to be explained by positing unobserved groups that unify some parts of the observed data. In this context, LDA assumes that each review is some mixture of topics; each topic is comprised of words; each word in a document is part of some topic that unifies a given review and another. Here were the results: The most interesting topics in the failed group are number 3 and number 5. Number 3 suggests that Yelp reviewers are referencing the act of leaving a negative review in their negative review as an indictment of the restaurant’s quality. Number 5 highlights the difficulty in predicting restaurant inspection failure. The same restaurant that was great yesterday might fail inspections today. For the passed reviews, it looks like hotels and meats are popular topics — and this is consistent with the Chi-Square table above. Of the Etobicoke restaurants, topic 3 looks to be the most coherent. It looks like — in some respect — to be talking about gentrification. For quiet restaurants,topic 1 may be hinting that there is a type of quiet restaurant consumer who comes to a restaurant for good coffee art and murals, not the food quality. III. Machine Learning tl;dr: Many more passing restaurants than failing restaurants led to the problem of ‘class imbalance’ . To mitigate class imbalance, I artificially balanced my dataset by creating members of the under-represented fail class, increased the cost of errors of the fail, and ensembled predictions from multiple individual classifiers. Inspecting my model’s weights, my results for what factors influence restaurant failure were that broadly consistent with what the descriptive analysis hypothesized. Before going into the actual model I used, it’s worth going over the metrics and problem of class imbalance Metrics Code For my dataset, I divided resteraunt data into feat_cols , which was a matrix of categorical and ordinal features and text_cols, which was tf-idf vectorized review data. y is a row vector consisting of 1s and 0s (fail or pass, respectively). Using SciKit Learn’s grid search module, I optimized hyperparameters for a c-support vector machine (SVM) for text data, two logistic regression (LR) models (one for text data, one for feature data), and a random forest (RF) model for feature data. (Based on experiments on non-parametric models and logistic regression, I calibrated the probability outputs.) Class imbalance: preparing for a zombie invasion Recall that only 20 percent of inspections resulted in a failure. This situation — where classes are unbalanced — leads to problems in machine learning classification. First, models are prone to over-predict (or almost exclusively predict) the majority class. Suppose you had to predict if there was going to be a zombie attack, one mental model that can get high accuracy (ie: you’d be correct often) is to always say “no, it will not happen today” because the majority of days are zombie-attack free. But we would likely tolerate more false positives and lower accuracy to increase recall of zombie attacks : With events like zombie attacks or restaurant failures, we might tolerate a few false positives, though our model realizes that the event is unlikely and won’t predict it — assigning the same cost to a false positive as a false negative. A second problem is that, by definition, there is less information on minority classes, which will lead to worse classification performance. Increasing Cost, Manufacturing Information, Leveraging Diversity To mitigate the class imbalance issues, I tweaked cost, information, and diversity. First, I increased the cost of errors for the ‘fail’ class by weighting errors of classes inversely proportional to their presence in the training set (class_weight=’balanced ). Second, to give my models more information about the fail cases, I used Synthetic Minority Over-Sampling Technique (SMOTE) . SMOTE is a technique to create synthetic samples of a minority class so a training set can be more balanced. SMOTE does this by selecting k neighbors of a minority class member, and then tweaking values randomly to create a new synthetic member of a minority class that is a function of the k neighbors’ attributes. The danger, of course, is that one overfits because each synthetic minority is still a combination of some other members in the dataset. To avoid overfitting — where my model learns ‘too much’ about the specific training set it is trained on, hindering its ability to generalize to new data — I introduced a ‘discount factor’, alpha . Ordinarily, to balance a dataset through over-sampling of a minority class, one finds a resampling rate r such that To avoid overfitting, I resampled the minority class by alpha * r : alpha <= 1 (ie: I accepted some class imbalance for a smaller chance of overfitting). Third, I used a dual-ensemble approach . An ensemble approach is when the predictions of disparate classifiers are unified to yield a final prediction. Along with oversampling, ensemble approaches have been helpful for dealing with class imbalance — as ensembled predictions are often better than those of the ensemble’s individual constituents. I created three ensembles, txt_voter, feat_voter, and master_voter. Each ensemble makes a prediction based on the argmax of the sum of its constituent’s probabilities: That is, each ensemble predicts fail if and only if the sum of probabilities put forth by its two constituents for fail is greater than those for pass. For master_voter, I used a grid search to find that the probabilities of text_voter should be counted twice as much as feat_voter to optimize Brier score. Intuitively,txt_voter provides more dynamic information than feat_voter because a restaurant’s basic features stay the same throughout inspections , but reviews by Yelpers offer data that varies between inspections — which means the latter can perhaps better explain time-varying outcomes. Results First, I found that the optimal alpha that reliably minimizes Brier score for Master Voter is alpha = 0.7.It’s interesting that full class rebalancing with SMOTE — which occurs at alppha = 1 — leads to a significantly worse Brier Score. A third degree polynomial relationship can approximate the relationship between Brier Score and alpha. Alpha = 0.7 or 0.9 provide comparable results, but alpha = 1 — which is full class rebalancing — hurts Master Voter’s Brier Score. The following tables are mean values from 20 trials with SMOTE (k-neighbors=4, alpha=0.7) and a confusion matrix representing one sample. Each trial used 60% of my data for training and 40% for testing. In context, recall is more important than precision: It’s useful to prioritize inspections of a restaurant who is at risk of failing. But in the event of a false alarm, all restaurants are inspected, anyway. So , to increase recall of ‘fail’ inspections, I lowered my threshold for predicting ‘fail’ to any case when the probability is greater than 0.3. What do the models say? Bringing It All Together Quietness as a curse: lr_text shows that something like ‘slowness’ (‘late’, ‘minute’, ‘checked’) is associated with restaurant failures.This makes sense as a proxy for management dysfunction. Second, rf_cat finds quietness to be one of the most informative attributes for predicting inspection results. From the Chi-2 table, quiet restaurants showed an increased probability of failing. From the LDA section, we saw that there is a certain type of quiet restaurant consumer who might be more into wall art than service. So, service slowness is indicative of possessing certain attributes that lead to food safety violations; quiet restaurants might share these troublesome attributes because consumers are accepting deficiencies in some areas in exchange for a nice ambience. Location, location, … cuisine?: lr_feat selected locations and certain ethnic cuisines to predict failure. As in the Chi-2 table, Chinatown and shows up as an important predictor of inspection failure. The practical upshot is that it’s reasonably effective to condition restaurant inspection failures on the neighborhood and food type, without knowing specific information about the restaurant. Trust the queue: Reviewer-described “popular” places— as well as places people like meeting and bringing groups — are far less likely to fail inspections than restaurants not described in these terms. For example, the coefficient on on wise is about 5, so restaurants whose reviews included wise were about e⁵ = 148 times more likely to pass inspections. Likewise, a place that is good for groups and and popular is also much less likely to fail inspections: Linked Papers Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent dirichlet allocation. Journal of machine Learning research, 3(Jan), 993–1022. Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16, 321–357. Fonseca, P. G., & Lopes, H. D. (2017). Calibration of Machine Learning Classifiers for Probability of Default Modelling. arXiv preprint arXiv:1710.08901. Japkowicz, N., & Stephen, S. (2002). The class imbalance problem: A systematic study. Intelligent data analysis, 6(5), 429–449. Maaten, L. V. D., & Hinton, G. (2008). Visualizing data using t-SNE. Journal of Machine Learning Research, 9(Nov), 2579–2605. Nikulin, V., McLachlan, G. J., & Ng, S. K. (2009, November). Ensemble Approach for the Classification of Imbalanced Data. In Australasian Conference on Artificial Intelligence (pp. 291–300). Salunkhe, U. R., & Mali, S. N. (2016). Classifier Ensemble Design for Imbalanced Data Classification: A Hybrid Approach. Procedia Computer Science, 85, 725–732. Packages Used Scikit-Learn Pandas Imblanced-Learn Scikit-Plot Sklearn-Pandas Yellowbrick Seaborn

Using Yelp data to predict restaurant inspection failure: a ‘voter of voters’ ensemble approach.

44

i-used-yelp-to-predict-whether-a-restaurant-will-fail-safety-inspections-1ca5b82ef06f

2018-05-06

2018-05-06 06:37:21

https://medium.com/s/story/i-used-yelp-to-predict-whether-a-restaurant-will-fail-safety-inspections-1ca5b82ef06f

false

4,192

null

Data Science

data-science

Data Science

33,617

Joshua Ashkinaze

null

24ae84618852

jashkina

2

1

20,181,104

null

0

null

0

null

2018-07-16

2018-07-16 13:30:02

2018-07-18

2018-07-18 10:04:45

2

false

fr

2018-07-18

2018-07-18 10:04:45

2

1ca5ff43a179

1.066352

0

Learn Assembly, équipe d’experts en conseil formation, a convié les intéressés des sujets de l’employabilité et des nouvelles méthodes du…

5

FRENCH TOUCH DE L’ÉDUCATION: APPRENDRE, ÇA S’APPREND ! Learn Assembly, équipe d’experts en conseil formation, a convié les intéressés des sujets de l’employabilité et des nouvelles méthodes du learning, à la French Touch de l’Education, le 4 et 5 juillet derniers. A la suite des différentes interventions lors de la conférence, organisée dans les locaux de l’IFCAM (Université du Groupe Crédit Agricole), nous avons tiré des enseignements que nous avons retranscrit pour vous, sous différentes parties. Apprendre l’employabilité avec les peuples premiers : oui mais comment? On les appelle les “peuples premiers”, “peuples indigènes“, “peuples autochtones” et les synonymes sont encore biens nombreux : mais qui sont-ils ? Le Haut-Commissariat des Nations Unies aux Droits de l’Homme les cite, dans une fiche d’information, comme “les descendants de ceux qui habitaient dans un pays ou une région géographique à l’époque où des groupes de population de cultures ou d’origines ethniques différentes y sont arrivés et sont devenus par la suite prédominants, par la conquête, l’occupation, la colonisation ou d’autres moyens”. Pour lire la suite, RDV sur Change the Work !

FRENCH TOUCH DE L’ÉDUCATION: APPRENDRE, ÇA S’APPREND !

0

french-touch-de-léducation-apprendre-ça-s-apprend-1ca5ff43a179

2018-07-18

2018-07-18 10:04:46

https://medium.com/s/story/french-touch-de-léducation-apprendre-ça-s-apprend-1ca5ff43a179

false

181

null

Education

education

Education

211,342

Change the work

Le média du futur du monde du travail à travers des témoignages de professionnels qui vous partagent leurs bonnes pratiques : http://eepurl.com/cz12rb

1b4abc703937

Changethework_71031

15

10

20,181,104

null

0

null

0

null

2018-03-01

2018-03-01 23:26:52

2018-03-02

2018-03-02 18:44:54

0

false

en

2018-03-02

2018-03-02 18:44:54

1

1ca619936785

0.418868

3

0

Dentem is a dental platform and as a practice management software is build from feedback of many dentist that use it everyday. We bring new…

5

D-Assistant Alexa integration — the future of hands free management in dental practices! (Demo) Dentem is a dental platform and as a practice management software is build from feedback of many dentist that use it everyday. We bring new technologies to make dental practice life easier. Today, we are thrilled to announce that the D-Assistant integration with Alexa Demo is out! Check out this YouTube video for the first demo! Amazon Echo Dot — Dentem The vision of the company comes deep from an entrepreneurship point of view! It’s all about changing the world to make it a better place! For more innovations stay tuned! Oni @ Team Dentem.

D-Assistant Alexa integration — the future of hands free management in dental practices! (Demo)

40

d-assistant-alexa-integration-the-future-of-hands-free-management-in-dental-practices-demo-1ca619936785

2018-04-06

2018-04-06 12:49:31

https://medium.com/s/story/d-assistant-alexa-integration-the-future-of-hands-free-management-in-dental-practices-demo-1ca619936785

false

111

null

Amazon Echo

amazon-echo

Amazon Echo

3,511

Dentem

Empower your dental practice. Dentem is a cloud based dental practice management software. That focuses in simplicity and innovation. Join us at www.dentem.co

2bf75fcbf266

dentem

70

30

20,181,104

null

0

#定義data augmentation tfms = tfms_from_model(arch, sz, aug_tfms=transforms_side_on, max_zoom=max_zoom_size) #讀取資料檔 data = ImageClassifierData.from_arrays(path=path, trn=(x_train, y_train), val=(x_test,y_test), tfms=tfms, test=test_image, bs=bs) #建立訓練的模型 learn = ConvLearner.pretrained(arch, data, precompute=True, ps=0.5) #指定模型存取的路徑 path= '../file_path' #參數設定 arch=vgg16 #modol max_zoom_size = 1.1 #zoom size bs=32 #batch_size wd=1e-4 #weight decay(l2 regularization) lr=1e-2 #learning rate learn.fit(lr, 3, wds=wd)

6

24b00b47988d

2018-07-09

2018-07-09 02:22:13

2018-07-09

2018-07-09 03:01:09

0

false

zh-Hant

2018-07-09

2018-07-09 03:01:09

2

1ca73d047e10

0.449057

0

Fast AI 是一個線上的ＡＩ公開課程，裡面提供了一系列非常好用的工具，課堂中了老師將許多ＡＩ的演算法包裝成一個Library，使用起來大大減少了應用ＡＩ模型時所需要的代碼數量

2

應用Fast AI Library 操作CNN Transfer Learning Fast AI 是一個線上的ＡＩ公開課程，裡面提供了一系列非常好用的工具，課堂中了老師將許多ＡＩ的演算法包裝成一個Library，使用起來大大減少了應用ＡＩ模型時所需要的代碼數量可以參考以下網址：http://www.fast.ai 在導入資料及建立模型，只需要短短的幾行代碼：只要事先將相關的參數定義起來，加上上面的程式碼，訓練Transfer Learning 的Model 可以說是非常方便：將上述都定義好之後，只需要key in fit 就好，模型就會開始訓練：除了建立模型相當容易之外，Fast.AI 也提供了很多可以降低overfitting 的方法，也非常容易調用，提供的方法如： Data Augmentation Regularization Learning Rate Annealing Test Time Augmentation Data Augmentation 為資料擴增的技巧，在Fast AI的工具中，會幫我們把圖形隨機放大、縮小一點點，也會隨機增加或減少圖片的亮度，增加資料數量，利用這種方式，可以增加模型泛化的能力 Regularization 可以針對模型算出來的權重進行修正，適度的修正可以減少模型的overfitting 比較有趣的一點是Learning Rate Annealing，搭配Pytorch 所提供的動態圖運算，可以在模型訓練的過程中，用不同的Learning Rate進行計算；應用這項特別的技巧，可以有利於模型找到更細膩的特徵與更好的參數，提升模型的成效而Test Time Augmentation 則是針對驗證集的資料，隨機進行資料擴增，並增加驗證的次數，最後的準確性會根據這幾次的驗證取平均，可以比較準確的估計模型的準確性這些方法都能有效的降低模型在訓練集上overfitting 狀況同時Fast AI 也提供了許多常見的模型來讓大家做Transfer Learning，例如Vgg Net、ResNet、DenseNet… 調用起來也非常方便，只需要指定模型的名稱就好總言之，Fast AI 大大減少了在一開始使用ＡＩ技巧的門檻，利用這個Library ，可以非常簡單的去調用ＡＩ的模型，如果大家有興趣，可以進一步參考官方網站上所提供的公開課以及他們的GitHub 這裏是我自己練習這項工具訓練用的代碼，是訓練辨識15種場景的圖形資料： AdamYuCheng/Machine-Learning Machine-Learning - Personal notes in machine learninggithub.com 讓我們一起用這些工具，進入ＡＩ的世界吧！

應用Fast AI Library 操作CNN Transfer Learning

0

應用fast-ai-library-操作cnn-transfer-learning-1ca73d047e10

2018-07-09

2018-07-09 03:01:09

https://medium.com/s/story/應用fast-ai-library-操作cnn-transfer-learning-1ca73d047e10

false

119

紀錄所想、所見、所聞、所感動

null

yucheng.adam

null

無邊拼圖框

aabbzztw@gmail.com

無邊拼圖框

DATA SCIENCE,ARTIFICIAL INTELLIGENCE,MACHINE LEARNING

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Chang Yu-Cheng

致力於探索信息中的價值；應用數據科學的技術，在充滿不確定性中的未來不斷地尋找可以被掌握的事物

26bd1d9cd66b

AdamCheng

1

2

20,181,104

null

0

null

0

null

2017-09-22

2017-09-22 04:30:25

2017-09-22

2017-09-22 04:32:33

1

false

en

2017-09-22

2017-09-22 04:32:33

2

1ca97b36938e

2.939623

0

1. Understand the business before you start solving problems

3

Tips about Data science 1. Understand the business before you start solving problems I know you’re an analyst and all you want is numbers. But what distinguishes a brilliant business analyst from the average data analyst? It is your potential to understand business. You should try to understand the business before you even resume your first project. Here are some things to explore: Customer-level information: total number of active customers, customer friction month by month, segments defined by the companies in the portfolio. Business strategies: how to acquire new customers, what are the channels. How do we keep valuable customers? Product information: How does your customer interact with your products? How do you make money from your product? Is your product a direct revenue provider, or is it just an engagement tool? If you can answer all of these questions, you are in good shape to start your first project. 2. Do you think if you solve an underlying problem or simply a result? I’ve noticed that analysts are targeting targets that are not even the main concern of the problem. For example, imagine that we find that the more customers serve customer service, the more willing they are to leave services. Now, if we start solving the method to minimize customer service calls, we probably will not reduce the rate of friction. Instead, I see greater dissatisfaction with your customers if you do not have a human being to justify your faults. 3. Spend more time looking for the correct measurement and quantity needed for implementation This is probably the easiest puzzle to solve an analyst, but a simple trap to fall into. Let me explain this using some examples. Suppose you are trying to build a targeting model for a marketing campaign. Which of the metrics you choose to measure your model: • Statistics of KS • Raise to the tenth • AUC-ROC • Log-Likelihood I will always choose KS in this case, since Lift will only give you estimates in a certain decile. Log-Likelihood is probably the most inappropriate in this case because all that matters to us is the sort order and not the actual probability. 4. Follow the divergent-convergent reflection process to avoid pre-mature convergence I’ve seen this as the biggest problem in many functions / industries. Today, business leaders are seeking innovation in everything they do. To truly innovate, you must follow a systematic way of diverging and converging. The extent to which you need to diverge will come when you have more experience in this approach. What we are trying to say is thinking of all possible ways of fearing that the problem is fair in feasibility, development time, and traditional approaches 5. Participate with business counterparts throughout the process From the first day of your review, you should interact with business partners. One thing I’ve seen wrong overall is for the analyst and business partner to get in touch with the solution infrequently. Business partners want to stay away from 6. Think about the simplest implementation levers to support your idea I know you are a statistician and like to confuse business people using complex formulations. This complexity in discussing with business people can help you get out of the instant conversation, but lessens the chances of a successful implementation. Here’s what you need to do: Once you have the output variables, try to find a simple lever that makes it easier to understand the business. 7. Learn to speak business language while introducing you to business leaders I recently started learning Chinese for one of my projects. The whole project was extremely easy, but I found that, even with a robust model, I was doing a poor job selling the business. For having understood your inner discussions. 8. Actively monitor the implementation plan With regard to the last but not the least, this happens since everyone is convinced with the effectiveness of their model. Your work is not yet complete. Set up monthly follow-ups with companies to understand how the project was implemented, is used in the correct shipment. If you want to build a career in Data Science, start right away. The field is growing quickly, and the sooner you understand the scope of Data Science Training, the sooner you’ll be able to provide solutions to complex work problems.

Tips about Data science

0

tips-about-data-science-1ca97b36938e

2018-06-19

2018-06-19 09:31:19

https://medium.com/s/story/tips-about-data-science-1ca97b36938e

false

726

null

Data Science

data-science

Data Science

33,617

Nandhini G

null

e768a51157e7

nandhinign22

2

7

20,181,104

null

0

null

0

null

2018-06-03

2018-06-03 16:38:06

2018-06-03

2018-06-03 20:02:48

6

false

en

2018-06-04

2018-06-04 00:07:00

0

1caa177fc277

2.757547

8

0

Suppose that we have a sample of height of Pakistani men. Lets say we have a sample of 10,000 Pakistani men heights and the mean height is…

4

Sampling Variability and Central Limit Theorem Suppose that we have a sample of height of Pakistani men. Lets say we have a sample of 10,000 Pakistani men heights and the mean height is 1.74 m with standard deviation=0.09 m then sample mean=1.74 m sample standard deviation= 0.09 m but we are interested in the population mean and population standard deviation. Here is the Central Limit theorem comes in. Central Limit Theorem The distribution of sample statistics is nearly normal, centered at the population mean and standard deviation equal to the population standard deviation divided by the square root of the sample size. x̄ ~N(mean= μ ,SE= σ/ √n) Sampling Statistics Suppose we collect n random samples from the population without replacement and calculate the mean for each random sample. These means are called sample statistics. If we plot the distribution of these sample statistics the distribution will be nearly normal and known as sampling distribution. So as the sample increases the standard error decreases i.e to have less variability around the mean in our sampling distribution we can increase our sample size. Note that the more skewed our population distribution is the larger sample size is required to get the nearly normal sampling distribution. Conditions For Central Limit Theorem Sampled observations must be independent. if sampling without replacement then n<10% of the population. The sample should not be too small. If the sample size is too small then we will not get a a nearly normal sampling distribution. n>30 is a general rule of thumb. Demonstration of Central Limit Theorem in Python Importing the necessary packages import pandas as pd import numpy as np import matplotlib.pyplot as plt import seaborn as sns import numpy as np Loading our dataset age=pd.read_csv(“heights.csv”) age.head() first five rows of our dataset print(‘Population mean age: ‘+str(age[‘age’].mean()),’Population age standard deviation: ‘+str(age[‘age’].std())) plt.title(‘Population Age Distribution’) sns.distplot(age[‘age’]) The population age is right skewed Choosing 1000 random observation from our population. samples=age[‘age’].sample(n=1000,replace=False).tolist() n_samples = [samples[x:x+30] for x in range(0, len(samples), 30)] n_samples_means=[sum(i)/len(i) for i in chunks] n_samples_means=np.array(n_samples_means).mean() standard_error=np.array(n_samples_means).std() print(‘Sampling distribution mean: ‘+str(n_samples_means),’Standard error: ‘+standard_error)) sns.distplot(n_samples_means) The sampling distribution mean (41.23) is approximately equal to the population mean (41.37) but the standard error (3.06) is very less than population standard deviation (15.86) since sampling distribution is nearly normal. Standard Error =σ/ √n according to Central Limit Theorem Standard Error=σ/√n=15.86/ √30=2.89 which is approximately equal to 3.06 We got standard error=3.06 We have just seen demonstration of Central Limit Theorem. Our sampling distribution is nearly normal with mean approximately equal to the population mean and standard error approximately equal to population standard deviation divided by square root of sample size. Why Central Limit Theorem? Central Limit Theorem helps in determining the unknown population parameter. Since the sampling distribution is nearly normal we can apply z scores to it. End Note Its just first version. i will write in more detail about Central Limit Theorem in the next version. Thanks for reading :)

Sampling Variability and Central Limit Theorem

9

sampling-variability-and-central-limit-theorem-1caa177fc277

2018-06-15

2018-06-15 15:23:14

https://medium.com/s/story/sampling-variability-and-central-limit-theorem-1caa177fc277

false

479

null

Data Science

data-science

Data Science

33,617

Usman Abbas

https://www.linkedin.com/in/usman-abbas-404385108/

cee8cfedb3c8

usmanabbas

14

31

20,181,104

null

0

null

0

59039de05b80

2018-03-01

2018-03-01 13:43:03

2018-03-05

2018-03-05 04:35:04

3

false

en

2018-03-05

2018-03-05 04:35:04

1

1caad29668ef

3.342453

0

null

2

What will Marketing Dashboards look like in 2020? But first, what is a marketing dashboard?! A marketing dashboard is a reporting tool that summarizes marketing KPI’s and metrics using data visualizations. They are specifically designed to simplify analytics and provide instant and continuous updates on your marketing performance. They are customizable and can be updated frequently. It facilitates tracking of marketing initiatives and aligns marketing expenditures with anticipated results. Metrics that are usually tracked usually belong to the following categories: customers (acquisition, retention, value), product (adoption, innovation, price, and margin), competitive positioning (market share, brand preference), and financial (budget, payback). Today’s marketing dashboards, provide easy to comprehend visualizations, a much-needed escape from data-heavy spreadsheets and emails. They are a focal point of all information that is needed to make decisions, spot trends and determine insights. A marketing dashboard generally has three types of metrics: Online Marketing Metrics like click-through rates, new sessions, cost-per-lead, bounce rate etc. Hard Metrics related to sales and revenue measurement like internal rate of return, net present value of presently running campaigns Soft Metrics like customer satisfaction, brand awareness, engagements and interactions Marketing Dashboards: What the future has in store? Predictive Analytics: Marketing dashboards give insights into what has happened in the recent past. The next logical step will be to identify market trends before they emerge using Predictive Analytics. Marketing professionals will be able to predict their consumer’s behavior patterns and use this knowledge to plan their next move. Going forward, in 2020, many dashboards will have recommender systems to leverage machine learning algorithms that will dig deep into consumer data based on actual interactions with the brand (through the e-commerce website, social media or in-store). Using these insights, it will predict actions of consumer profile(s) and help determine strategies to capture their attention in real-time. Such rich insights will help marketers channelize their resources in the right direction and plan their campaigns accordingly. Integrated Management System: Marketers today often make use of different dashboards for multi-channel campaigns. An ideal dashboard would be one which can pull data from different channels and campaigns for optimizing marketing plans. AI Driven Dashboards: Artificial Intelligence along with machine learning algorithms can help marketers automate insights for all data and get recommendations from within the dashboard itself. Though it may seem too futuristic, with a plethora of data already available in dashboards, algorithms can be made to learn from past data so that they can draw better conclusions. Instead of just giving numbers on how a particular campaign is performing, the dashboard may suggest making strategical changes like focusing more on action X to convert more leads. Dashboards, in future, may help marketers understand what mix of campaigns, channels, and audience will drive ROI. Dashboards may have newsfeeds where insights will appear and be updated at regular intervals. AI may provide real-time behavioral indicators and notifications of exceptional events through intelligent alerts. A dashboard may be programmed to provide contextual feedback on the performance of a campaign. Natural Language Processing and Search Console in the dashboard: A dashboard with a search console will make it easier to find precise information a marketer needs at any given moment. Natural language processing (NLP) will be like the icing on the cake. “NLP is the combination of machine learning, AI, and linguistics that allows us to talk to machines as if they were human”. Introducing NLP will make data more user-friendly and accessible. Voice-Driven Dashboards: With the evolution and accelerating growth in voice technology, the days are not far, where you could interact with your dashboards verbally. Probably integrating Alexa, or Siri or Google Assistant with your dashboard to ask questions like How many people added Product X to your “Buy Later List”, and voila your dashboard assistant can respond with a number and a suggestion that it’s now time to offer these customers a discount to convert them! If Artificial Intelligence, NLP, and Voice are integrated with marketing dashboards, they will begin to understand semantics. Instead of simply giving you data they will be able to answer and make suggestions by inferring from queries, performing filtering activities. In 2020, you may not need to view your mobile or desktop to see your dashboard, you may simply need to ask a question and it will answer with relevant insights and intelligent suggestions. What will Marketing Dashboards look like in 2020?

What will Marketing Dashboards look like in 2020?

0

what-will-marketing-dashboards-look-like-in-2020-1caad29668ef

2018-03-05

2018-03-05 04:35:06

https://medium.com/s/story/what-will-marketing-dashboards-look-like-in-2020-1caad29668ef

false

740

The World’s Leading Source for Marketing Technology News, Research, Product Comparisons & Expert Views

null

martechadvisor

null

MarTech Advisor

shabana.arora@martechadvisor.com

martech-advisor

MARKETING,BUSINESS,TECHNOLOGY,TECH,NEWS

MarTechAdvisor

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

MarTech Advisor

Helping Marketers Succeed

4a1305dc5a36

martechadvisor

172

448

20,181,104

null

0

null

0

null

2017-09-14

2017-09-14 03:59:57

2017-09-14

2017-09-14 04:04:41

1

false

en

2017-09-14

2017-09-14 15:56:22

4

1cab05ba3f49

4.739623

0

Varun is the Head of Early-Stage Fund of Qualcomm Ventures. In this episode, he makes us jump into the future to imagine how cities will…

5

JUMP INTO THE FUTURE Varun is the Head of Early-Stage Fund of Qualcomm Ventures. In this episode, he makes us jump into the future to imagine how cities will look with autonomous vehicles and how it will have a dramatic impact on our lives. I met Varun Jain at Startupfest in Montreal where he was a keynote speaker, one of the best among those I had the opportunity to listen to. Varun is the Head of Early-Stage Fund of Qualcomm Ventures. In this episode, he makes us jump into in the future and imagine how cities will look with autonomous vehicles and how it will have a dramatic impact on our lives. If you haven’t listened to the first episode with Varun, What was disruptive for the past three years, please do so. You will discover how Varun made Qualcomm Ventures become the third investor in Cruise Automation, the startup sold to GM for $1 billion $ seventeen months ago. A fascinating story! In this episode, Varun reflects on what he shared in his talk, The future of Automation and Robotics. “We are at the threshold of a dramatic shift. One in fifty Americans drives a truck for a living. Cities make over US$6 billion a year in speeding tickets. And there are billions of dollars in taxicab medallions. Add up the impact of self-driving cars, and it’s going to transform everything about society — from where we put service stations, to the value of houses with garages, to the very idea of car ownership. Qualcomm Ventures’ Varun Jain has been at the center of this disruption. Qualcomm invests in all stages of startups, the fields that they like most are machine learning and robots. Music: Smash feat. Ridley — Future People (AFP Anthem) (VOLAC Remix) The automation has the reputation that will cause labor loss but on the other hand, it will create new opportunities for those involved in this shift of technology. For the past three years, there were some huge acquisitions, strategic partnerships which increased the valuation of those companies. On the other hand, it attracted a lot of talent to work to develop those new technologies. There is a lot of excitement and paranoia around this subject and the regulation that will be required. What does safety mean? What new regulations will be needed? Within the next three years, Varun says that we will have autonomous cars in controlled environments. This has already been tested in some cities. Despite of the investment of big companies, there are a lot of startups searching for new solutions. This doesn’t mean that the biggest ones will disrupt the technology. The trucking industry is the one that might be the first one to use the autonomous trucks. On the roads, there are fewer obstacles. And the labor cost is a significant part of the total cost of the industry. Varun Jain explained about how other industries could be affected. If his scenario is correct, more people will probably live more outside cities. Hotels on the road might suffer because you’ll be able to sleep in your car. When autonomous cars become a reality, which Qualcomm Ventures expect to happen within 5 to 10 years, the impact won’t be limited to just the car industry, Varun thinks it will be a systemic shift which will change the way we live our lives, particularly in urban environments. If we were to consider how our lives will change in an autonomous vehicle world, Varun and his team believe that there are five or six major changes that we will see and these are all dramatic shifts which will not just change our everyday life but also the future of certain pre-existing multibillion dollar industries that will be phenomenally affected once this technology becomes a day to day reality. Varun gave the examples: When your car becomes your office, your living-room or your bedroom, the time you spend to commute is not wasted but rather used for relaxing, or working on your laptop, or doing something else. People will then worry less about the time they spend going from point A to point B. Qualcomm Ventures thinks that it will cause them to think: Rather than living in a small two-bedroom apartment in San Francisco, which is close to my work, why don’t I live in a four-bedrooms mansion in a suburban area 60 minutes away from work because the 60 minutes I commute everyday is actually my work time and I can be productive.” “We fundamentally believe that the first autonomous vehicle will be in the form of sharing services. Moving toward, airplanes, business models, fleet services. We believe the overall number of cars in cities will come down over a period of time as more and more people start thinking of transportation as a service, as opposed to something that needs to be owned. That will not just have an impact on the number of cars in the city but this will also have an impact on adjacent businesses like, for example, parking garages. In major urban cities, 30% of overall real-estate is used for parking. When will move to transportation as a service world all that space will be freed up. It will have an impact not only on the real-estate in the cities but also on the quality of life. At that point, it will become much easier to move around in the cities for people and families. It will be much more pleasant. In the next two or three years, commercial autonomous cars will be services offered in cities where autonomous vehicles are tested now. Over the next five years, it will become mainstream. The future generation won’t learn how to drive a car. It’s difficult to evaluate but experts think that accidents will decrease by 60% to 80%. In conclusion, Varun said that It is a fantastic time if you understand the systematic shift that is happening. He invited us to think what is next not by building those vehicles but imagining new services and products. NOW YOU PLAY! I recommend that you read the book by Kevin Kelly, The Inevitable: Understanding the 12 Technological Forces That Will Shape Our Future to fully understand at what point we are living in an amazing time. Download the Checklist-1 for Future People to reflect on the first three forces explained by Kevin Kelly and start imagining new products and services for this new world. Do not miss the next episode with the futurist Robert Scoble, it will be another big jump into the future. Let us know what you think? Do you have ideas how you can contribute to the future? What new products or services should we create? Leave me comments, I would like to know what are your thoughts on the subject. The Creative Mind Trick of the day! Sylvie is one of the co-founders of The New School of Creativity, a school without walls that developed innovative learning experiences based on emotional intelligence and neuroscience. She also teaches creativity and innovation to the PhD Candidates at the Polytechnique School of Montreal. Subscribe to her podcast They Make Me Smarter.

JUMP INTO THE FUTURE

0

jump-into-the-future-1cab05ba3f49

2017-12-25

2017-12-25 04:49:20

https://medium.com/s/story/jump-into-the-future-1cab05ba3f49

false

1,203

null

Self Driving Cars

self-driving-cars

Self Driving Cars

13,349

Sylvie Gendreau

CAHIERS DE L'IMAGINAIRE, nouveau média sur l'art et la création. http://www.cahiersdelimaginaire.com #nouvelle école de créativité #lelaboratoire créatif

ca5272cd6c93

SylvieGendreau

206

205

20,181,104

null

0

null

0

null

2018-04-21

2018-04-21 21:28:56

2018-04-21

2018-04-21 23:02:11

2

false

en

2018-04-23

2018-04-23 06:22:08

4

1cab3604e875

2.534277

2

0

Beginning today, I will start a series of article demonstrating what I am currently struggling with my coding projects

5

My Coding Diary with R Beginning today, I will start a series of article demonstrating what I am currently struggling with my coding projects Currently, most of these projects are volunteer works motivated by merely an interest to understand more about numerical methods. These are just some of the examples: Simulation of Dispatchment Schemes in a Power System Using R Generation of a Triangular Network Using R Numerical Simulation of a Non-Divergent Geostrophic Wind Using R Transient Convention Diffusion Modeling Using R So how do you analyze the energy efficiency performance of a building, and what to do next? Experience in Promoting Energy Efficiency Initiatives on CampusData Visualization Using R However, it is very likely that some of my current projects will later be used in my works. For example, I probably will have to improve my spatial interpolation algorithm from my previous term project: Why Do I Use R? I know that some of my readers will now be thinking, “how unprofessional! Why doesn’t he use Python?” These is a very legitimate question, so I must make it clear from the very beginning that I am not a professional coding guy. The only reason I have been using R is because it was the only program I was taught during my undergraduate period (other than C). Since then I have been using it for almost everything from simple calculations to developing taboo algorithms. So let’s just say that there exists an inertia too large for me to just change my default coding program to Python or Matlab or whatever. If you are interested in learning the syntax of these other programs (which is a very good thing and I encourage you to do so), then my diary probably wouldn’t contribute much to your resource base. But if you are more interested in how to develop numerical methods without the need to comprehend too much robust math, than this series might help you to some extent. I am neither a mathematician nor a computer scientist. Rather, I am just a mere mortal who needs some algorithm to do his term projects or master thesis. But I think that is why most of us want to learn coding in the first place, isn’t it? Why do I Use R Studio? It is not necessary to run R on R Studio. However, the platform provides a easier way to code with R. Is This a Tutorial? Actually it isn’t. It is a learning diary. I don’t know anything about the subjects I am about to cover more than most of you do. So if you have a better solution to my problems or simply a different idea, you are very welcome to leave a response in any of my upcoming articles. It is better to treat these articles as a track of how I think and how I tackle the problem. They might not be the canonical way to tackle the problem, and they might contain flaws. Be very picky and cautious to my reasoning. It is anything but the collary-proof-theorem stuff you see on a math textbook. What Will be Featured? I am planning to develop a spatial interpolation scheme next. To do so, I have built a triangular network. Later I wish I can do a functional principal component analysis with my network. Let’s see what will happen then and decide what shall be featured beyond that. How Can I Generate a Triangular Network with R? Part of My Coding Diary with Rmedium.com How Can I Do a Principal Component Analysis in R? Part of My Coding Diary with Rmedium.com

My Coding Diary with R

11

my-coding-diary-with-r-1cab3604e875

2018-04-23

2018-04-23 13:33:15

https://medium.com/s/story/my-coding-diary-with-r-1cab3604e875

false

570

null

R

r

R

1,558

Tony Yen

A Taiwanese student who studies in the master program of Renewable Energy Engineering and Management in University of Freiburg.

32333e6125a6

yenishrepublic

255

277

20,181,104

null

0

null

0

null

2018-08-25

2018-08-25 17:18:49

2018-08-25

2018-08-25 17:18:59

0

false

en

2018-08-25

2018-08-25 17:18:59

1

1cab6db2a423

1.177358

0

[PDF] Download Superintelligence: Paths, Dangers, Strategies By Nick Bostrom Best Ebook download Link…

1

Pdf Download Free eBook Superintelligence: Paths, Dangers, Strategies By Nick Bostrom PDF eBook #EPUB [PDF] Download Superintelligence: Paths, Dangers, Strategies By Nick Bostrom Best Ebook download Link https://reviewskindlenew.icu/?q=Superintelligence%3A+Paths%2C+Dangers%2C+Strategies . . . . . . . . . . . . . . . . . . . Read Online PDF Superintelligence: Paths, Dangers, Strategies, Download PDF Superintelligence: Paths, Dangers, Strategies, Download Full PDF Superintelligence: Paths, Dangers, Strategies, Download PDF and EPUB Superintelligence: Paths, Dangers, Strategies, Read PDF ePub Mobi Superintelligence: Paths, Dangers, Strategies, Reading PDF Superintelligence: Paths, Dangers, Strategies, Read Book PDF Superintelligence: Paths, Dangers, Strategies, Read online Superintelligence: Paths, Dangers, Strategies, Download Superintelligence: Paths, Dangers, Strategies Nick Bostrom pdf, Download Nick Bostrom epub Superintelligence: Paths, Dangers, Strategies, Read pdf Nick Bostrom Superintelligence: Paths, Dangers, Strategies, Download Nick Bostrom ebook Superintelligence: Paths, Dangers, Strategies, Read pdf Superintelligence: Paths, Dangers, Strategies, Superintelligence: Paths, Dangers, Strategies Online Download Best Book Online Superintelligence: Paths, Dangers, Strategies, Read Online Superintelligence: Paths, Dangers, Strategies Book, Read Online Superintelligence: Paths, Dangers, Strategies E-Books, Read Superintelligence: Paths, Dangers, Strategies Online, Read Best Book Superintelligence: Paths, Dangers, Strategies Online, Read Superintelligence: Paths, Dangers, Strategies Books Online Download Superintelligence: Paths, Dangers, Strategies Full Collection, Download Superintelligence: Paths, Dangers, Strategies Book, Read Superintelligence: Paths, Dangers, Strategies Ebook Superintelligence: Paths, Dangers, Strategies PDF Read online, Superintelligence: Paths, Dangers, Strategies pdf Download online, Superintelligence: Paths, Dangers, Strategies Read, Download Superintelligence: Paths, Dangers, Strategies Full PDF, Read Superintelligence: Paths, Dangers, Strategies PDF Online, Read Superintelligence: Paths, Dangers, Strategies Books Online, Read Superintelligence: Paths, Dangers, Strategies Full Popular PDF, PDF Superintelligence: Paths, Dangers, Strategies Read Book PDF Superintelligence: Paths, Dangers, Strategies, Read online PDF Superintelligence: Paths, Dangers, Strategies, Download Best Book Superintelligence: Paths, Dangers, Strategies, Read PDF Superintelligence: Paths, Dangers, Strategies Collection, Read PDF Superintelligence: Paths, Dangers, Strategies Full Online, Read Best Book Online Superintelligence: Paths, Dangers, Strategies, Download Superintelligence: Paths, Dangers, Strategies PDF files

Pdf Download Free eBook Superintelligence: Paths, Dangers, Strategies By Nick Bostrom PDF eBook…

0

pdf-download-free-ebook-superintelligence-paths-dangers-strategies-by-nick-bostrom-pdf-ebook-1cab6db2a423

2018-08-25

2018-08-25 17:18:59

https://medium.com/s/story/pdf-download-free-ebook-superintelligence-paths-dangers-strategies-by-nick-bostrom-pdf-ebook-1cab6db2a423

false

312

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

tfjsdhrsrwett664

null

5a290f11b8e7

tfjsdhrsrwett

0

1

20,181,104

null

0

null

0

null

2018-02-14

2018-02-14 11:14:01

2018-02-14

2018-02-14 11:38:33

1

false

en

2018-02-15

2018-02-15 17:15:01

1

1caced707e1d

1.007547

3

0

It is obvious, most companies in Medical Technology or their Investors haven’t read, Zero to One or The Art of War.

5

“Fight the Competition where they are not.” Sun Tzu It is obvious, most companies in Medical Technology or their Investors haven’t read, Zero to One or The Art of War. The quote in the title, in terms of business strategy, suggests we look at areas where our competition is not concentrating. Look for something they are neither focusing on, nor spending resources developing. In the case of Ophthalmology a disproportionate amount of resources, talent and capital, has been allocated to fight the scourge of Diabetic Retinopathy (DR). While DR is no doubt an awful condition, it barely accounts for 1% of the causes for Global Blindness (data courtesy of the International Association for the Prevention of Blindness). Total Market Size for Opticals Worldwide = $150B Total Market Size for Diabetic Retinopathy = $7B Diabetic Retinopathy Screening ML Models = 30+ (and more coming) Refractive Error Predictive ML Model = 1 (At LVPEI Center for Innovation) If I can not appeal to the Medtech community’s sense of fairness, I hope I can appeal to their desire for market dominance. There are innumerable other significant conditions in Ophthalmology that would benefit from their know how and investment. And along the way of giving many the gift of sight, they will make more than a few bucks.

“Fight the Competition where they are not.” Sun Tzu

4

fight-the-enemy-where-they-are-not-sun-tzu-1caced707e1d

2018-03-05

2018-03-05 18:11:01

https://medium.com/s/story/fight-the-enemy-where-they-are-not-sun-tzu-1caced707e1d

false

214

null

Machine Learning

machine-learning

Machine Learning

51,320

Raghu Gullapalli

Leveraging Data & Computation to bring Last Mile HealthCare to the First Mile. Executive Director of Emerging Technology, LV Prasad Eye Institute, LVPEI.org

5276cd2f8e26

tradergu

20

139

20,181,104

null

0

null

0

84bb99513fba

2017-10-20

2017-10-20 14:21:57

2017-10-20

2017-10-20 14:23:33

2

false

en

2017-10-20

2017-10-20 14:23:33

1

1cadd344021f

2.288994

2

0

by Gemma Wisdom

5

Are you ready to meet your AI self? by Gemma Wisdom We recently came across a new chatbot AI called Replika which was designed to learn your unique patterns of communication and how you would respond to certain situations so that it can copy them. While this may all sound like the set up for a technology-themed Halloween movie so far, the social applications for an AI that can replicate your behavior are exponential. How helpful would it be to have a duplicate of you that could handle all the nuisance interactions that you have to deal with on a daily basis? A personal assistant that goes beyond the quick Q&A sessions that Siri and Cortana provide and can answer questions on your behalf, in a decent facsimile of your voice. How comforting will it be to your loved ones that some version of you will persist in the world after you have expired? Even before death, the ability to revisit an earlier iteration of a partner and to remember them as they were when you first fell in love might do wonders for relationships. Decision making will be so much easier when you can form a quorum of your past-selves to chime in on an issue; combining the idealism of youth with the wisdom gained through experience. Illustration by Simon Wood It is telling that the best-known attempts at AI at the moment are those personal assistant applications. Simple machine intelligences that we have imbued with human personalities to make them more palatable. Everyone in the technology sector is intently focused on the practical applications of AI; the ways that the technology will simplify data analysis, the ways that AI will make life easier. The societal impact of AI is being left on the table. Humans, as a species, use emotion as the basis for most decision making, so the way that we are going to interact with AI on an emotional level is at least as important as the practicalities when it comes to adoption of the technology. Even if AI are the most useful pieces of technology since the invention of the smart phone; if they fail to connect with their users then they are never going to catch on. Illustration by Simon Wood This is why it is vital to integrate design research into these projects from the very beginning. By centring the user and the ways that they are going to rely on AI, the emotional component of that interaction can be given the same value in the design process as the technological solutions. Whether that is by creating a chatbot that mimics your patterns of speech, following the psychological principle of “mirroring” or by giving Siri a dry sense of humor. We may not know all the ways that AI are going to affect society, but by focusing on putting the user at the heart of development we have a better chance to predict all of the myriad ways that life is going to be changed.

Are you ready to meet your AI self?

6

are-you-ready-to-meet-your-ai-self-1cadd344021f

2018-04-09

2018-04-09 17:26:17

https://medium.com/s/story/are-you-ready-to-meet-your-ai-self-1cadd344021f

false

505

Stories driven by curiosity to improve everyday experiences

null

sutherlandlabs

null

Sutherland Labs

info@sutherlandlabs.com

sutherland-labs

TECH,USER EXPERIENCE,DESIGN,DESIGN THINKING

sutherlandlabs

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Sutherland Labs

We design and run bespoke user experience strategy, research, and design for major brands to help them understand the end-to-end customer experience.

b3970c03afa

SutherlandLabs

135

146

20,181,104

null

0

null

0

null

2018-07-25

2018-07-25 22:12:49

2018-07-25

2018-07-25 22:17:48

1

false

en

2018-07-25

2018-07-25 22:17:48

0

1cae9fec57f6

2.464151

0

Predicting Progression of Alzheimer’s

5

Artificial intelligence system projects future symptoms Predicting Progression of Alzheimer’s Alzheimer’s disease, the sixth leading cause of death in the US, kills more senior citizens than breast cancer and prostate cancer combined. The disease affects millions of people annually, says the Alzheimer’s Association. If diagnosed early, as much as $7.9 trillion in medical and care costs could be saved, according to an article by Kyle Wiggers in Venture Beat. Researchers at Unlearn.AI think they have a solution, the article says. The startup company designs software tools for clinical research and anticipates a major role for artificial intelligence in personalizing diagnosis and treatment of Alzheimer’s. Their paper (“Using deep learning for comprehensive, personalized forecasting of Alzheimer’s Disease progression”) on the preprint server Arvix.org describes a system for predicting disease progression. In other words, the system projects the symptoms that individual patients will have at any time in the future. The researchers published a training video on the website Unlearn.AI. As they explained, “Two patients with the same disease may present with different symptoms, progress at different rates, and respond differently to the same therapy. Understanding how to predict and manage differences between patients is the primary goal of precision medicine. Computational models of disease progression developed using machine learning approaches provide an attractive tool to combat such patient heterogeneity.” While presenting an interesting possibility for precision medicine in the management of Alzheimer’s disease, AI-enabled systems for tracking cognitive decline have been suggested before. Neurologists at Montreal’s McGill University created a “PET scan-ingesting algorithm that identified which patients ended up with dementia with 84 percent accuracy,” and scientists at North Carolina’s Duke University and Croatia’s Rudjer Boskovic Institute “used machine learning to pick up changes in brain tissue loss over time,” Wiggers related. The innovative element in Unlearn.AI’s system is its unsupervised learning approach. This system utilizes data that is unclassified or unlabeled. Additionally, it can compute predictions and confidence intervals for numerous patient characteristics at the same time. The researchers developed the system in two stages. Initially, they modeled clinical data using a Boltzmann Encoded Adversarial Machine (BEAM), a form of neural network highly effective for “classification and feature modeling tasks.” They trained and tested BEAM on the Coalition Against Major Diseases (CAMD) Online Data Repository for Alzheimer’s Disease, that includes 1,908 patients evaluated during a timeframe of 18 months including 42 variables, such as the individual components of ADAS-Cog (a frequently used cognitive subscale) and Mini-Mental State Examination (a questionnaire that measures cognitive impairment in clinical and research venues). In the second phase, the researchers utilized the trained model to develop “virtual patients” and their related cognitive exam scores, laboratory tests, and clinical data. They ran simulations for individual patients to figure out their disease progression in such areas as orientation, word recall, and naming, which were then utilized to figure out the overall ADAS-Cog score. The Unlearn.AI researchers claim that the unsupervised model could make accurate ADAS-Cog predictions to at least 18 months in the future. They suggest that the system can be adapted to preject the outcomes for patients who have other degenerative diseases. The team concluded, “The approach to simulating disease progression that we describe here can be easily extended to other diseases. Widespread application of deep generative models to clinical data could produce synthetic datasets with lower privacy concerns than real medical data, or could be used to run simulated clinical trials to optimize study design. In certain disease areas, tools that use simulations to forecast risks for specific individuals could help doctors choose the right treatments for their patients.”

Artificial intelligence system projects future symptoms

0

artificial-intelligence-system-projects-future-symptoms-1cae9fec57f6

2018-07-25

2018-07-25 22:17:48

https://medium.com/s/story/artificial-intelligence-system-projects-future-symptoms-1cae9fec57f6

false

600

null

Health

health

Health

212,280

dan sfera

Future owner of the NBA's Clippers. Entrepreneur. Clinical Trials. 👋🏻. Arizona Wildcat for life. http://www.TheClinicalTrialsGuru.com

4b939f468944

TheRealDanSfera

579

259

20,181,104

null

0

null

0

null

2017-12-21

2017-12-21 01:49:44

2017-12-21

2017-12-21 01:52:53

1

false

en

2017-12-21

2017-12-21 01:52:53

3

1caec8f81188

2.237736

0

Artificial intelligence (AI) and robots have had the mainstream fired up since they made their silver screen debut in the 1920’s film…

4

AI is taking off in showbiz Artificial intelligence (AI) and robots have had the mainstream fired up since they made their silver screen debut in the 1920’s film, Metropolis. In the last decade alone, at least 30 movies with a spotlight on AI have been released. From Star Wars to Iron Man, AI technology has kept the masses entertained and chomping at the bit for more. In recent years, however, AI has since transcended its role as a plot focus to add more value to show business as a practical solution. Cognitive movie trailers The thriller Morgan adopted AI to create the movie trailer in 24 hours, a task that usually takes humans 10 to 30 days to complete. The machine was put into film school to learn from more than 100 horror movie trailers with experimental APIs before processing Morgan in about 90 minutes, extracting its scary scenes for the film’s trailers. However, while AI is the tool that picks and arranges the visuals, a human element is still necessary to supervise the creative aspect of production. Hence, an editor stepped in to create the final trailer from those selected clips. Automatic screenwriters Benjamin, an algorithm — LSTM recurrent neural network — which is commonly used in text recognition, wrote the movie script and created a soundtrack for the 9-minute long sci-fi movie Sunspring. Created by filmmaker Oscar Sharp and AI researcher at NYU Ross Goodwin, the movie won a top ten spot at Sci-Fi London’s 48-Hour Film Challenge. Benjamin was fed the scripts of different science fiction movies including such classics as Highlander Endgame and The Fifth Element before embarking on script development. Some of the lines don’t make sense, yet it is still an interesting watch. Optimization for customers To increase views and offer personalized choices, Netflix developed Meson, an AI framework to operate a large volume of machine learning workflows on a daily basis. Meson manages personalization algorithms that drive video recommendations based on users’ previous viewing activities and behavior, subscription history, interaction with content, viewing times and devices. It also refines the way the home page is presented to users, where the “Continue Watching” section is optimized to maximize views. Judging the performances of talents Last year, an AI developed by Alibaba Cloud was able to predict with 100% accuracy the finalists and grand winner of hit Chinese reality singing show, I’m a Singer. By evaluating variables including the ability of the singer, the popularity of their song, and crowd responses, the AI predicted the winners the show’s judges would select. This example demonstrates the ability of AI to judge the talent of performers in show business, in a way which emulates how humans evaluate talent. While AI is incapable of true creativity, its technology speeds up labor-intensive work, giving those in the industry more time to focus on the creative process. While showbiz is still at the initial phases of adopting AI, it’s still exciting to see the ways it augments the production and distribution processes. It’s safe to say that this duo of machine intelligence and human creativity has made a lasting impact on showbiz, and there’s much more to come. Reference: https://www.alibabacloud.com/blog/AI-is-taking-off-in-showbiz_p96495

AI is taking off in showbiz

0

ai-is-taking-off-in-showbiz-1caec8f81188

2018-04-06

2018-04-06 21:30:14

https://medium.com/s/story/ai-is-taking-off-in-showbiz-1caec8f81188

false

540

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Alibaba Cloud

A global Cloud Computing and AI company, a subsidiary of Alibaba Group. Website: https://www.alibabacloud.com/

8cdb52f9cdca

Alibaba_Cloud

504

14

20,181,104

null

0

null

0

7f48c10a07d

2018-03-01

2018-03-01 16:32:28

2017-10-27

2017-10-27 11:12:20

1

false

en

2018-03-01

2018-03-01 17:04:30

8

1cb062d93ff0

1.641509

0

null

5

Digital Marketing News: Pinterest PPC, Facebook Sets, Online Beats Offline Shopping Author: Lee Odden / Source: Online Marketing Blog — TopRank® Need help understanding Machine Learning? We now live in an age where machines can teach themselves without human intervention. Sound scary? It should. Scary amazing that is. Applications for machine learning extend from marketing to medicine to interstellar space travel. Find out what Machine Learning is, how it works, and how it will change the world. Infographic. Google’s New AI Is Better at Creating AI Than the Company’s Engineers. Google CEO Sundar Pichai says his team has achieved “AI inception” with AutoML. AutoML is an artificial intelligence that can assist in the creation of other AIs. By automating some of the complicated process, AutoML could make machine learning more accessible to non-experts. Futurism Survey: 37% of online retailers started holiday preparations earlier this year. How early you ask? 1 to 4 months earlier than 2016 according to a survey by BigCommerce. Along with early, retailers are optimistic. 88% expect an increase in holiday revenue. Marketing Land Oh joy (sarcasm) Facebook is bringing paywalls to Instant Articles in your mobile feed. Since more people than ever before are getting their news from social media, it makes sense that Facebook wants to help publishers by introducing subscriptions for content on its platform. And it’s starting on mobile. The Next Web Digital Video Marketing Is A $135 Billion Industry In The U.S. Alone, Study Finds. Video capturing, creation, hosting, distribution, analytics and staffing is big business! In contrast, advertisers are expected to spend $83 billion on digital ads and $71 billion on TV commercials (a total of $154 billion) in the U.S. this year. Forbes Businesses can now sign up to add booking buttons to their Google local results. Google has finally added a feature to let you easily add a ‘book online’ button to your local business on Google Maps or Google Search. Soon, some businesses might not even need a website. Search Engine Land Snap is turning to programmatic ads for Snapchat shows. Advertisers can make programmatic buys on Snap Ads — 10-second vertical video units — across the app’s public user stories, Snapchat-curated live stories and Discover publisher channels and Snapchat shows. Digiday Click here to read more

Digital Marketing News: Pinterest PPC, Facebook Sets, Online Beats Offline Shopping

0

digital-marketing-news-pinterest-ppc-facebook-sets-online-beats-offline-shopping-1cb062d93ff0

2018-03-01

2018-03-01 17:04:30

https://medium.com/s/story/digital-marketing-news-pinterest-ppc-facebook-sets-online-beats-offline-shopping-1cb062d93ff0

false

382

oneQube’s AI powered audience automation software stack enables brands to build powerful relevant digital audience for their product, content and brand. oneQubes software and team of audience architects develop, and manage highly engaged passionate audiences.

null

OneQube

null

oneQube

Connect@oneqube.com

oneqube

BLOCKCHAIN,MARKETING TECHNOLOGY,SOCIAL MEDIA MARKETING,AUDIENCE DEVELOPMENT,DIGITAL MARKETING

oneQube

Digital Marketing

digital-marketing

Digital Marketing

59,312

oneQube

Intelligent audience automation software to develop highly engaged digital audiences, grow brand awareness, drive traffic and transactions.

e7162a4f66e5

oneQube

19

180

20,181,104

null

0

null

0

null

2018-07-31

2018-07-31 18:22:41

2018-07-31

2018-07-31 18:39:11

2

false

en

2018-07-31

2018-07-31 18:39:11

4

1cb25d5a391d

13.836164

0

I just finished reading a few more op-eds on how Google, Walmart, Facebook, Comcast and the yet-to-be-named Amazon-Berkshire…

5

Introducing the Real Amazon Prime Health. Ready? I just finished reading a few more op-eds on how Google, Walmart, Facebook, Comcast and the yet-to-be-named Amazon-Berkshire Hathaway-JPMorgan Chase (ABC) not-for-profit are going to fix healthcare. Unfortunately, most new entrants will likely give way to an industry perfectly designed to yield the same current results — questionable outcomes, gross inefficiencies and perverse, out-of-control consumer pricing. The C-suites of all these companies openly admit how difficult and complex the task will be, and few are envisioning board meetings ending with everyone high-fiving and chanting “we got this.” Whether new entrants succeed or not is irrelevant. Either way, they’re positively pushing the incumbents, challenging the status quo, placing the industry under a microscope and, more importantly, pressure testing a system to see if it has any chance of becoming a true consumer market. From what I’ve read, those predicting how a company like Amazon will disrupt healthcare all seem to be pointing to the obvious like leveraging their tremendous buying power to negotiate better pricing and having Alexa remind us to take our medications or order our next PillPack. Simple tactics and technology alone will not fix the complexities and fundamental issues of our current healthcare system, for which “sick care” is a more accurate description. At the risk of sounding like I’ve loaded the silver bullet, I think most predictions are missing the real power that Amazon could exert to extend our individual health spans, improve population health, reduce pricing, deliver consistent quality and, as Warren Buffet has famously stated, ultimately kill “the tapeworm of the U.S. Economy.” Low-Hanging Fruit Could these business giants reduce current inefficiencies by 30‑50% and over-diagnosis and over-treatment within their employee populations? Absolutely. It’s low-hanging fruit. Many large employers have already paved the way toward reducing their annual healthcare spend by taking an active part in implementing “novel” or alternative delivery models like direct primary care (DPC), on-site clinics, telehealth, specialty carve-outs, direct contracting, incentivizing physician and employee behavior, and by demanding financial transparency. All these, by the way, aren’t really novel — they’ve been around for over 20 years and are once again being recycled as innovative with hope that maybe they’ll work this time around. Regardless, these courageous large employers are outliers. To most employers, sponsored coverage is a nuisance. Some people, like Dan Munro, author of “Casino Healthcare” and Forbes contributor, might argue that employer-sponsored insurance is the reason for the majority of what’s wrong with the U.S. health system. And as much as employers complain about the high cost of employee healthcare, they do get a tax write-off at the end of the year. Maybe it’s enough to make them feel a little better about 5+ percent annual cost increases. Unfortunately, most are not equipped with the knowledge or resources to take greater control of their healthcare spend even when healthcare costs are one of the highest expenses on their balance sheet. The new CEO of the ABC healthcare company, Dr. Atul Gawande, will likely face many of the same challenges as he first focuses on creating greater value for ABC employees. Outside of their own employees, will ABC and other new entrants inspire external innovation, disrupt the status quo, drive down healthcare prices and improve people’s health spans? No, unless they all create or greatly influence a near-complete system redesign by turning the existing healthcare food chain on its head and, more importantly, help change the collective human behavior that feeds it. The reason being that the current healthcare system has been designed to get the exact results it gets, namely: - Hospitals make their own pricing, like an unregulated utility. - Physicians make more money by over treating and increasing utilization. - Insurance companies benefit from high cost, as it results in greater profit. - Brokers benefit from higher cost, as it increases their income. - Pharmaceutical companies enjoy no limitations on price or profit. - Employees have no interest or influence in purchasing decisions. Healthcare consultants are well aware of the system’s shortfalls, as are most healthcare CEOs and administrators, but often conflicted about making fundamental changes. To get different results, the fundamentals of the system must change. And changing the fundamentals isn’t popular because it will result in many stakeholders losing money or, if they’re smart, cannibalizing their own business before someone else does. If any company can force change, reinvent a business model, and overhaul an inefficient and ineffective market, it’s a new entrant like ABC. Much of what happens moving forward and the speed at which change occurs will depend largely on courage, passion, trust, creativity, empathy, persistence, patience and the permission to fail. The Healthcare Food Chain Let’s review how the healthcare food chain is designed today (reference Figure 1). Currently insurers are at the top, then large healthcare systems, physicians and, at the VERY bottom, employers and individual consumers. The closer to the top, the larger the influence in decision making relative to what services get delivered and paid and take the lion’s share of profits. Stakeholders on the two lower tiers are stuck paying for poor quality and questionable value. How many other industries or markets relegate the individual consumer to the least influential part of a purchase? Think about how upside down it is and who SHOULD have market power and influence. There are approximately 300 million U.S. consumers, one million doctors, 5,000 hospitals and 25 commercial health insurers, with only a handful controlling the majority market share (sans the largest payer, Centers for Medicare and Medicaid Services (CMS)). There are more consumers than doctors, more doctors than healthcare systems and more healthcare systems than insurers. In a true market-based system, consumers would be influencing the products and services delivered and the price they’re willing to pay. For instance in India, open-heart surgery performed at institutions equal in clinical outcomes (and in some instances better) to their U.S. peers costs about $2,000. In the U.S. the same procedure averages $150,000. Conceptually, employers, individual consumers and physicians could work together and turn the current food chain upside down, but it’s a Herculean effort that most are too afraid or unwilling to try. They feel the boat is just too damn big to turn and have become accustomed to the “healthcare normal” of it being an expensive, disjointed, poor experience. I find it fascinating that many physicians are happy to get 110% of what Medicare pays. Some might say that’s analogous to being happy with making 10% over minimum wage. The Blame Game It’s convenient to point a finger and demonize CEOs of insurance companies, hospital systems, large physician groups and big pharma for out-of-control price increases. Let’s not forget, they’re just doing what they’re hired to do — create maximum profit and shareholder value. The CEO of United Healthcare has received compensation of over $270 million over the last six years. However, the root cause of the systemic problems in medical care is not the high compensation of United’s CEO, it’s actually closer to you and me. We’ve individually and collectively decided that we’re powerless to change things and just put up with current reality. Psychologist Martin Seligman calls this phenomenon “learned helplessness” where we’re conditioned to accept pain and discomfort to the point of believing it’s so inescapable and out of our control that we stop trying even when given an option to avoid the pain. Ask any hospital CEO or CFO what their biggest priority is and they’ll tell you (if they’re totally honest) “filling beds” or the new-top-of-mind term, “control patient leakage” (losing an opportunity to treat a consumer). Again, don’t lose sight of the fact that hospitals, physicians and everyone in the “medical industrial complex” only make money when people are ill. To this end, our healthcare system has been brilliantly designed and adjusted to extract about $4 million out of every millennial’s life span. Treating illness pays well. Preventing illness, historically, hasn’t paid anything. What we have is a system that capitalizes on individuals not willing or able to make better lifestyle choices, change behavior, understand the basis of what they’re paying for and agree on how healthcare value is defined. Are insurers, hospitals, physicians and big pharma “working within the system” to the financial detriment of employers, employees, individual consumers and the economy? But how much is our individual behavior to blame? The Root Cause In Ray Dalio’s video, “How the Economic Machine Works,” he emphasizes that economics are driven by one key factor — human behavior. Not unlike the economy, our health and wellness is greatly influenced by behavior. It’s the single most important factor in our lives — more so than our genetic makeup. The root cause of our healthcare (sick care) problem is not only our acceptance of the system’s dysfunctional design and pricing, but equally the greater impact of our individual lifestyle and behavior choices. Aside from an occasional random cell mutation, our overall health is largely dependent upon lifestyle and what goes into our mouths and lungs. Social determinants play a significant part and represent a huge opportunity that’s finally getting the attention it deserves by CMS and startups alike. Are non-transparent pricing, 30–50% waste, over-treatment, upside-down incentives and lack of trust feeding this economic tapeworm? Absolutely. Is individual behavior cited as the root cause of most illness and reduction of life spans? Not so much. Is Population Health Improving Financial Health? With the cost of healthcare reaching an all-time high of over $28,000 per year for a family of four, maybe we’re getting closer to a proverbial tipping point. If your employer is picking up 60 percent of the bill, it’s still expensive. If you’re an independent or member of the gig economy, it’s a choice between paying a mortgage on a $500,000 house in Austin, Texas, or paying healthcare premiums. On the delivery side of healthcare, other onerous trends and statistics are accelerating the battle for market share, mergers and acquisitions and collaborative partnerships between providers, payers and technology companies. And although the party line is that consolidation will lead to decreased costs, greater efficiencies and better outcomes, it remains to be seen if it comes to fruition and if decreased costs and increased efficiencies will result in lower healthcare premiums, improved outcomes and better experiences for consumers, or if providers and payers will just continue to maintain growing profit margins. Managing and improving the health of large patient cohorts or populations of people is a really difficult task to do effectively. Many providers believe they can do it internally without external expertise, advice or technology. Managing the health of populations is a collaborative undertaking, and going it alone imparts more risk and probability of failure. With a few exceptions, value-based care programs that have been totally provider sponsored have not produced great results. The work of more than 70 population health and AI companies all on a mission to improve care and reduce costs are moving the needle in the right direction. Cerner’s recent $266 million investment in Essence Health and Lumeris is a perfect example of complementary partnerships that could make a big impact, strengthen providers’ control in the food chain, make healthcare more frictionless, improve efficiency, and become a hallmark for population health and value. However, healthcare and prevention is a behavior-change business. You can install the latest technology, evidence-based protocols, processes and methodologies, but if you can’t change the behaviors of consumers and providers, and incentivize those involved in the healthcare ecosystem, nothing will change. What’s more, if the fundamental design of the system stays the same, one cannot and should not expect different results. Innovative startups, mid-market companies and new entrants like Google and Apple have their work cut out for them even though they understand human behavior better than most. But if any new entrant can help change our health-related behavior and redesign a system to produce better outcomes with better economics, I’m leaning toward ABC. However, the metrics by which we measure population health success must be radically different than they are today. Behavior and Lifestyle Change It’s really hard. It’s why we try the latest diet, re-join the gym every January and stop going in March and regress to consume over 66 pounds of sugar each year, smoke, overindulge in alcohol, sleep poorly, avoid vacation, and work in stressful jobs we hate. The bottom line — we need help to change our behavior patterns. Although Amazon is often demonized for its unending appetite for cannibalizing just about every industry, Facebook for selling detailed consumer data with the result of manipulating political views, and Google for knowing more about us than we do, they all have a fantastic opportunity to help us change the course of chronic illness and the behaviors that cause it — IF, and it’s a big IF, they commit to protecting our personal data, securing our trust and leveraging the data they currently have to help us lead a healthier lifestyle with a vision of extending our health span. One of the reasons why the CEOs of major insurance companies, pharmaceutical companies and large healthcare systems are concerned with new entrants like ABC is because they can provide something that insurers and providers struggle with — the trust of consumers. According to a recent Harris Poll, only 7% of Americans trust their health insurance company. The winners of Healthcare 3.0 (or 4.0) will be those delivering the best experience, lowest price and highest trust (great outcomes should be a given). Fearful Incumbents Incumbents should realize, if they haven’t already, that with a few additional simple pieces of personal data, trust (our opting in), security, frictionless experiences and a different profit motive, endeavors like ABC have an opportunity to redefine the meaning of healthcare value and change consumer behavior. And if they’re successful changing consumer behavior, they can turn the current healthcare food chain upside down so it more closely resembles a true market-based system. Now before you say, “I’m not providing any more personal information to any company — look what happened with Facebook and Cambridge Analytica,” have you closed your Facebook account? Have you noticed smartphone and social media addictions waning? Social media and the smartphone we carry are the new addictions, and they’re going to be very difficult, like most behaviors, to change. At the same time, both could be the keys to improving population health in a very personalized way. The word addiction carries a negative connotation, but it doesn’t have to. The phrase “addicted to good health” may sound oxymoronic but it could become part of a business model with a unique value proposition. Silicon Valley insider Jaron Lanier believes that social networks have become “behavior modification empires” dividing people through manipulation and surreptitious intent. Instead of social networks and data brokers vacuuming up consumer information to divide, there’s an opportunity to create “healthy behavior modification engines” for the common good with complete transparency. I can think of plenty of ways that Amazon and others can monetize addiction to good health. At this point, it’s remains unclear and too soon to predict what Dr. Gawande’s initial business strategy and approach might be as he first leads his new company to provide improved healthcare value to the employees of ABC. And although it may be cloudier as to what external value propositions ABC might offer the external health consumer, one thing is clear — the data that ABC already has is far more valuable than their ability to negotiate better pricing or implement novel alternative delivery models. Introducing ABC Prime Health Regardless of recent data scandals, if Cambridge Analytica could be as successful with micro-targeting individuals in the same household to sway their behavior to vote or behave in a desired way, imagine what Amazon could do with what it knows about us for the good of improving individual health. The Amazon “know me” and “trust me” factors are pretty amazing when you think about it. If you opted in to letting an Amazon delivery driver open the front door of your home or leave a package in the trunk of you car, then it’s pretty safe to say you trust them. If you didn’t, that fact that Amazon felt confident enough to offer the option speaks volumes about their internal perception of customer trust. If you shop on Amazon, it already knows your name, gender, address, location, credit card number, creditworthiness, reading preferences, approximate weight (if you buy clothes), education level, purchase and search history, and personality type (if you leave reviews). From these data points Amazon could infer a tremendous amount about your lifestyle, income level, personality and health. If you shop at Whole Foods, Amazon could curate data on what you eat and how much you spend on various types of food. With the recent acquisition of PillPack, Amazon will be able to determine your chronic health issues by the medications you take, and even apply an overall health-risk rating. If you’re a Medicare or Medicaid beneficiary that’s having trouble getting to and from doctor appointments, Amazon could leverage their Flex driver program and offer a non-emergent medical transport (NEMT) service. Google, Apple, Microsoft and a host of healthcare providers and payers have tried to get people to enter health data into a personal health record (PHR) only to eventually shutter efforts. Apple is trying again and has entered into an agreement with 40+ hospital systems. It’s been a challenge to get consumer adoption and continual use of PHRs. It remains to be seen whether Apple’s Open API and integration with provider electronic medical records (EMRs) will be successful this time around. The problem is people just aren’t interested in taking the time to enter their health data. The ROI just isn’t strong enough, especially if you’re healthy. They want someone else to do it for them. With the data Amazon already has, all that’s missing, clinically, would require a consumer to write a short history of surgical procedures, ailments currently bothering them, and family health history. Amazon’s use of artificial intelligence (AI), machine learning and natural language processing will take all that is written and codify it automatically to create a nearly complete personalized health profile. As an Amazon (or ABC) subscriber, all you have to do is opt in to the Prime Health Platform and get suggestions on how you could improve your health based on the data aggregated as part of your personalized health profile and continued purchasing habits. Not much different than how Amazon currently suggests books or products you might like. Tie in Alexa, a smartwatch and phone, and other IoT devices like Bluetooth blood pressure cuffs, glucometers, a weight scale and pill box, and Alexa could provide reminders, alert and coach you when your goals are missed, and congratulate you when you’re on target. Based on your health challenges or other interests, Prime Health could make referrals to physicians, health organizations or insurers with the best experiences, outcomes and prices. And there’s no better time to reinforce a behavior or nudge someone to act than right at the moment of a purchase decision when you have their undivided attention. Further reinforcement could occur through daily “nudges” to help you change your habits and behaviors. I could go on and on ideating with much more advanced capabilities, but you get the idea…. What’s more, ABC’s revenue could come from subscriptions to the Prime Health Platform and associated purchases made from personalized recommendations. The big unknown for the likes of ABC is whether their healthcare vision and mission includes redesigning our healthcare system with new fundamentals and leveraging their knowledge, know-how, technology and expertise in consumer behavior to produce better health outcomes, redefining healthcare value and turning the healthcare food chain upside down. I anticipate ABC stacking the deck by forming partnerships and acquiring population health companies, status-quo-busting providers, human-centered design firms, digital health innovators, AI and semantic inference technologists and novel social media health platforms. The collective knowledge, behavior change know-how, methods, processes, data and new success metrics could lead to the disruption that everyone so often talks about. Disruption, regardless of its pace, must include drastic reductions in consumer pricing and lateral changes in market expectations, value propositions, consumer and provider behavior, self-care and personal responsibility. It’s safe to say ABC has the resources, time and expertise to disrupt or reshape a market. And although it’s getting difficult to discern what Amazon is as a company, what is clear is that it’s a disruption machine. It remains to be seen whether Dr. Gawande will drive disruption in an aggressive manner or in a more tempered way that seeds a vision of improving the health of the U.S. population, changes the fundamental economics and redefines how value is defined. It takes a great deal of courage to change the world in a meaningful way. Personally, I can think of no better mission for Dr. Gawande and everyone involved in healthcare transformation than to increase our individual and collective health spans and, simultaneously, kill the “tapeworm of the U.S. economy.” © 2018 Vince Salvo Trademarks and copyrights are the property of their respective companies

Introducing the Real Amazon Prime Health. Ready?

0

introducing-the-real-amazon-prime-health-ready-1cb25d5a391d

2018-07-31

2018-07-31 18:39:11

https://medium.com/s/story/introducing-the-real-amazon-prime-health-ready-1cb25d5a391d

false

3,565

null

Healthcare

healthcare

Healthcare

59,511

Vincent Salvo

null

858673dc9c01

vinnie_salvo

7

16

20,181,104

null

0

null

0

null

2018-06-08

2018-06-08 09:40:48

2018-06-08

2018-06-08 09:44:22

1

false

en

2018-06-08

2018-06-08 09:44:22

1

1cb2f805096c

3.267925

0

New approach can help organizations scale their data science efforts with artificial data and crowdsourcing.

2

Artificial data give the same results as real data — without compromising privacy. New approach can help organizations scale their data science efforts with artificial data and crowdsourcing. Although data scientists can gain great insights from large data sets — and can ultimately use these insights to tackle major challenges — accomplishing this is much easier said than done. Many such efforts are stymied from the outset, as privacy concerns make it difficult for scientists to access the data they would like to work with. In a paper presented at the IEEE International Conference on Data Science and Advanced Analytics, members of the Data to AI Lab at the MIT Laboratory for Information and Decision Systems (LIDS) Kalyan Veeramachaneni, principal research scientist in LIDS and the Institute for Data, Systems, and Society (IDSS) and co-authors Neha Patki and Roy Wedge describe a machine learning system that automatically creates synthetic data — with the goal of enabling data science efforts that, due to a lack of access to real data, may have otherwise not left the ground. While the use of authentic data can cause significant privacy concerns, this synthetic data is completely different from that produced by real users — but can still be used to develop and test data science algorithms and models. “Once we model an entire database, we can sample and recreate a synthetic version of the data that very much looks like the original database, statistically speaking,” says Veeramachaneni. “If the original database has some missing values and some noise in it, we also embed that noise in the synthetic version… In a way, we are using machine learning to enable machine learning.” The paper describes the Synthetic Data Vault (SDV), a system that builds machine learning models out of real databases in order to create artificial, or synthetic, data. The algorithm, called “recursive conditional parameter aggregation,” exploits the hierarchical organization of data common to all databases. For example, it can take a customer-transactions table and form a multivariate model for each customer based on his or her transactions. This model captures correlations between multiple fields within those transactions — for example, the purchase amount and type, the time at which the transaction took place, and so on. After the algorithm has modeled and assembled parameters for each customer, it can then form a multivariate model of the these parameters themselves, and recursively model the entire database. Once a model is learned, it can synthesize an entire database, filled with artificial data. Outcome and impact After building the SDV, the team used it to generate synthetic data for five different publicly available datasets. They then hired 39 freelance data scientists, working in four groups, to develop predictive models as part of a crowd-sourced experiment. The question they wanted to answer was: “Is there any difference between the work of data scientists given synthesized data, and those with access to real data?” To test this, one group was given the original data sets, while the other three were given the synthetic versions. Each group used their data to solve a predictive modeling problem, eventually conducting 15 tests across 5 datasets. In the end, when their solutions were compared, those generated by the group using real data and those generated by the groups using synthetic data displayed no significant performance difference in 11 out of the 15 tests (70 percent of the time). These results suggest that synthetic data can successfully replace real data in software writing and testing — meaning that data scientists can use it to overcome a massive barrier to entry. “Using synthetic data gets rid of the ‘privacy bottleneck’ — so work can get started,” says Veeramachaneni. This has implications for data science across a spectrum of industries. Besides enabling work to begin, synthetic data will allow data scientists to continue ongoing work without involving real, potentially sensitive data. “Companies can now take their data warehouses or databases and create synthetic versions of them,” says Veeramachaneni. “So they can circumvent the problems currently faced by companies like Uber, and enable their data scientists to continue to design and test approaches without breaching the privacy of the real people — including their friends and family — who are using their services.” In addition, the machine-learning model from Veeramachaneni and his team can be easily scaled to create very small or very large synthetic data sets, facilitating rapid development cycles or stress tests for big data systems. Artificial data is also a valuable tool for educating students — although real data is often too sensitive for them to work with, synthetic data can be effectively used in its place. This innovation can allow the next generation of data scientists to enjoy all the benefits of big data, without any of the liabilities. The project was funded, in part, by Accenture and the National Science Foundation.

Artificial data give the same results as real data — without compromising privacy.

0

artificial-data-give-the-same-results-as-real-data-without-compromising-privacy-1cb2f805096c

2018-06-08

2018-06-08 09:44:23

https://medium.com/s/story/artificial-data-give-the-same-results-as-real-data-without-compromising-privacy-1cb2f805096c

false

813

null

Data Science

data-science

Data Science

33,617

OSA Hybrid Platform

Creating value for the retailer, supplier, and shopper through introduction of technological innovations

a85c7db5016d

osahp

0

1

20,181,104

null

0

null

0

null

2018-09-08

2018-09-08 10:30:25

2018-09-08

2018-09-08 12:05:34

3

false

th

2018-09-08

2018-09-08 12:05:34

11

1cb441722bf4

2.64434

1

0

ระบบการเงินของ Bank กลายเป็นแอปในมือถือ ซึ่งมีข้อมูลเป็น Digital ที่สามารถนำมาใช้วิเคราะห์และประมวลผลได้อย่างที่ไม่เคยทำได้มาก่อนด้วย AI

5

Upgrade ระบบการเงินด้วย Big Data และ AI Big Data เป็นคำที่ทุกคนชอบพูดถึง แต่หลายคนอาจจะไม่ได้เข้าใจมันจริงๆ ตอนนี้ถ้าใครอยากระดมทุนอย่างรวดเร็วก็มักจะใส่คำว่า Big Data, Artificial Intelligence (AI), Machine Learning, หรือ Deep Learning เข้าไปให้ดูเท่ๆ ในความเป็นจริงแล้ว ศาสตร์ทางด้านนี้มีความหลากหลาย และมีความซับซ้อนหลายระดับ ซึ่งหลายคนไม่เข้าใจ ไม่ใช่ว่าทุกอย่างที่มีคำว่า Big Data จะประสบความสำเร็จ แต่ความสำคัญของ Data ในชีวิตประจำวันของเรานั้นมีมากขึ้นเรื่อยๆ อย่างหลีกเลี่ยงไม่ได้ Trend ของ Big Data และ AI เติบโตมากับบริษัทที่ทุกวันนี้กลายเป็น Tech Giant อย่าง Amazon Google Facebook บริษัทเหล่านี้สามารถเก็บข้อมูลได้ในแบบที่โลกไม่เคยเห็นมาก่อน ด้วยพัฒนาการทางสังคม 2 ด้าน 1) การเติบโตของมือถือ Smart Phone (Mobile Technology) ที่อยู่ในกระเป๋าของทุกๆคน และ 2) การเข้าถึงข้อมูลและ Software ผ่าน Internet (Cloud Technology) Photo by rawpixel on Unsplash บริษัทยักษ์ใหญ่เหล่านี้จึงสามารถดึงข้อมูลผู้ใช้จำนวนมหาศาลมาเก็บไว้บนระบบ Digital โดยข้อมูลที่เป็น Digital เหล่านี้สามารถนำมาวิเคราะห์ได้ว่าลูกค้าต้องการอะไร และสร้าง Platform ที่ดีขึ้นเรื่อยๆ อย่างต่อเนื่อง เพื่อตอบสนองความต้องการของลูกค้า การพัฒนาอย่างต่อเนื่องแบบนี้เองสามารถทำให้ Amazon เปลี่ยนรูปแบบการค้าปลีกไปตลอดกาล ทำให้ปัจจุบัน Amazon เติบโตจนมีมูลค่าถึง 1 ล้านล้านเหรียญสหรัฐ หรือ กว่า 2 เท่าของ GDP ไทย บรรดา Startup ทั้งหลายจึงมองหาโอกาสเปลี่ยนแปลงหรือ Disruption ในอุตสาหกรรมอื่นที่เก่าแก่ อย่างธุรกิจ Bank โดยเป้าหมายก็คือการสร้างบริษัทที่จะแข่งกับ Bank เหมือนกับสิ่งที่ Amazon ได้ทำกับธุรกิจค้าปลีก Data รูปแบบใหม่ในโลกการเงิน ด้วย Mobile และ Cloud Technology ระบบการเงินของ Bank จึงกลายเป็นแอปในมือถือ ซึ่งมีข้อมูลเป็น Digital ที่สามารถนำมาใช้วิเคราะห์และประมวลผลได้อย่างที่ไม่เคยทำได้มาก่อน และเปิดช่องทางให้ Fintech Startup ในสหรัฐ สามารถนำ Technology ใหม่ๆ มาโจมตีช่องโหว่ของธนาคารได้เกือบทุกจุด เช่น Credit Karma บริษัทที่ให้คนเข้าไปตรวจสอบ Credit Score ของตัวเองได้ฟรี ได้ผันตัวมาให้บริการเป็นผู้แนะนำสินค้าทางการเงินต่างๆ โดยใช้ AI และ Algorithm มาวิเคราะห์ข้อมูลผู้ใช้และให้คำแนะนำเช่น คุณควรจะย้ายบัตร Credit Card จากธนาคาร Chase ไปที่ธนาคาร Capital One นะ เพราะธนาคารนี้ดอกเบี้ยถูกกว่า Credit Karma ก็ได้รายได้จากข้อตกลงกับ Partner หรือผู้ที่มาโฆษณาขายสินค้าทางการเงินต่างๆ บน Platform ส่วนบริษัท Fintech ที่ปล่อยกู้ Consumer finance บนมือถืออย่าง Affirm และ LendUp ที่ปล่อยกู้ Personal loan ในสหรัฐ พบว่าลูกค้าที่ใช้เวลานานในหน้าแอป ใบสมัครขอกู้บนมือถือ มักจะมีโอกาสผิดนัดชำระหนี้ตำ่กว่า ลูกค้าที่ใช้เวลาสั้น เพราะคนที่ใช้เวลาสั้นมักจะไม่ได้คิดคำนึงถึงความสามารถในการผ่อนจ่ายของตัวเองมากนัก การวิเคราะห์ข้อมูลแบบนี้นำไปสู่การคาดการณ์โอกาสผิดนัดชำระซึ่งสำคัญต่อการอยู่รอดของบริษัทเงินกู้อย่างมากโดยเฉพาะลูกค้าที่ไม่มีประวัติการกู้ยืมมาก่อน ข้อมูลประเภทนี้มาจากการวิเคราะห์ Big Data ซึ่งมีได้ในระบบที่เป็น Digital เท่านั้น เนื่องจากเมื่อ 20 ปีที่แล้วนายธนาคารคงไม่สามารถใช้การวัดแรงกดปากกาของคนที่มาเซ็นขอเงินกู้ เพื่อวิเคราะห์โอกาสการผิดนัดชำระหนี้ได้เหมือนข้อมูลที่เป็น Digital อีกตัวอย่างหนึ่งของการใช้ข้อมูลในแบบที่ธนาคารคิดไม่ถึงเช่น Startup อย่าง SoFi ที่ปล่อยเงินกู้ Personal loan ในสหรัฐ ใช้มาตราฐานการปล่อยกู้ที่ต่างไปจากธนาคารทั่วไป โดย SoFi จับกลุ่มลูกค้าที่บริษัทเรียกว่า HENRY หรือ High Earners Not Rich Yet โดยเลือกปล่อยกู้นักเรียนวิทยาศาสตร์ วิศวะ แพทย์ และกฎหมาย ซึ่งเป็นกลุ่มนักเรียนที่น่าจะมีเงินเดือนที่สูงในอนาคต แต่ปัจจุบันยังไม่รวยและไม่มีประวัติการกู้ยืมเงิน เพราะฉะนั้นธนาคารทั่วไปจึงไม่กล้าปล่อยเงินกู้คนกลุ่มนี้ เปิดช่องให้ SoFi สามารถเติบโตได้อย่างรวดเร็วในช่วง 3–4 ปีที่ผ่านมา ปล่อยเงินกู้ได้กว่า 3 พันล้านเหรียญสหรัฐ ธนาคารเองก็ต้องปรับตัวนำข้อมูลและฐานลูกค้าเดิมที่ตัวเองมีอยู่แล้วมาใช้ผสมผสานกับ Big Data แอปของธนาคารในประเทศไทยเองก็เริ่มมีการนำข้อมูลมาใช้ทดลองเริ่มปล่อยกู้ Personal loan โดยส่งเป็นข้อเสนอวงเงินกู้ ไปที่แอปเป็นรายบุคคล ส่วน Fintech ที่เป็น e-wallet ในประเทศไทย ณ ปัจจุบันนี้ เช่น Rabbit LinePay TrueMoney หรือรายที่กำลังมาในอนาคตอย่าง Grab Financial และ JD Finance ก็ย่อมต้องใช้ Big Data มาสร้างรายได้ขายสินค้าทางการเงินต่างๆ (Cross-selling) บน Platform ตัวเองอย่างแน่นอน Photo by Markus Spiske on Unsplash AI ยกระดับการวิเคราะห์ข้อมูล เมื่อยุค Digital นำมาซึ่งข้อมูลที่มากมาย เทคนิคการวิเคราะห์และประมวลผลจึงสำคัญมากไปด้วย เพื่อนำข้อมูลไปใช้ให้เกิดประโยชน์ทางธุรกิจจริงๆ ศาสตร์ทางด้านนี้เกิดการพัฒนาแบบก้าวกระโดด เมื่อเทคนิคหนึ่งในศาสตร์การพัฒนา AI แบบ Machine Learning ซึ่งเรียกว่า Deep Learning สามารถวิเคราะห์และประมวลผลนำข้อมูลจำนวนมากได้อย่างมีประสิทธิภาพแบบที่ไม่เคยทำได้มาก่อน เทคนิคนี้จึงเปิดประตูให้ Fintech และ Bank สามารถนำ AI มาประมวลผลข้อมูลที่มีจำนวนมหาศาลเพื่อแก้ปัญหาหลากหลายด้านไม่ว่าจะช่วยจับการฉ้อฉล ปล่อยกู้ลูกค้า หรือ chat กับลูกค้า โลกการลงทุนก็นำ AI มาใช้สร้างความได้เปรียบในการลงทุน จะเห็นได้ว่ากองทุนต่างๆ นำ Machine Learning มาวิเคราะห์งบการเงินบริษัท หรือจับความนิยมของสินค้าด้วยจำนวนการ Search ชื่อสินค้าใน Google เพื่อสร้างความได้เปรียบในการลงทุน ด้าน Hedge fund ในสหรัฐ ก็ร่วมมือกับ Startup อย่าง Orbital Insight สามารถวิเคราะห์ราคานำ้มันด้วยการใช้ภาพถ่ายดาวเทียมติดตามนับจำนวนเรือขนส่งนำ้มัน หรือคาดการณ์ราคาข้าวโพดจากการวิเคราะห์สีความเข้มของภาพถ่ายไร่ข้าวโพดจากดาวเทียม แม้การทั่งนับจำนวนรถยนต์ในที่จอดรถของ Walmart เพื่อทำนายยอดขายในไตรมาสนี้ Photo by Rod Long on Unsplash ประกันภัยก็เริ่มใช้ AI เช่นกัน Ant Financial ซึ่งเป็นบริษัทลูกของ Alibaba ในประเทศจีน ร่วมมือกับบริษัทประกันภัยต่างๆ พัฒนาแอปที่ให้คนขับรถส่งวีดิโอของรถยนต์ที่เสียหายจากอุบัติเหตุมาวิเคราะห์ด้วย AI เพื่อบอกผู้ใช้ว่าจะเรียกร้องความเสียหายจากบริษัทประกันได้เท่าไร และนำรถยนต์ไปซ่อมได้ที่ไหน ถ้า Alibaba ทำสำเร็จ การใช้ AI จะช่วยลดเวลาและลดค่าใช้จ่ายได้อย่างมากต่อทั้งบริษัทประกันภัย และผู้ซื้อประกันเอง AI คือการยกระดับการวิเคราะห์ข้อมูลที่สร้างความได้เปรียบอย่างมหาศาลต่อบริษัทที่มี Data อยู่ในมีจำนวนมากไม่ว่าจะเป็นธนาคารเองที่มีข้อมูลลูกค้าจำนวนมากอยู่แล้ว และ Startup ใหม่ๆ ใช้กลยุทธ์ให้บริการฟรีๆ หรือค่าธรรมเนียมศูนย์เพื่อดึงลูกค้ามาที่ Platform ตัวเอง ดังนั้น AI จึงเป็นเครื่องมือที่ทุกธุรกิจสามารถนำมาใช้สร้างประโยชน์ได้ เพราะจริงๆแล้วเราก็ใช้ AI กันทุกวันไม่ว่าจะเป็นการแปลภาษาใน Google Translate หรือการ Tag รูปที่ post ใน Facebook รวมถึงการหาสินค้าใน Amazon การเติบโตของ Big Data และ AI ที่กล่าวมาจึงผลักดันให้ความจำเป็นในการเก็บข้อมูลเพิ่มขึ้นมากไปด้วย บริษัท Cisco ประเมินว่าจำนวน Data Center จะเติบโต 13% ต่อปีจาก ปี 2016 ถึง 2021 ในขณะที่การสื่อสารผ่าน Data Center จะเติบโตถึง 25% ต่อปีในช่วงเวลาเดียวกัน ผู้เขียนเองในฐานะที่ปัจจุบันบริหารกองทุน K-PROP ของ บลจ.กสิกรไทย ซึ่งมีการลงทุนใน REIT ที่มีทรัพย์สินเป็นอสังหาริมทรัพย์ประเภท Data Center (จากรายงานประจำปี ณ วันที่ 31 พฤษภาคม 2018) จึงจับตาการเติบโตของอุตสาหกรรม Data Center เป็นพิเศษ และอยากให้ผู้อ่านได้ตระหนักถึงความสำคัญของ Big Data และ AI ด้วยเพราะ Trend นี้จะส่งผลต่อทุกธุรกิจ ในอนาคต Big Data และ AI คงไม่ใช่คำพูดที่ดูเท่ๆ สำหรับให้บริษัท Startup ระดมทุนได้ง่ายๆ อีกต่อไป แต่คงเป็นคำพูดที่น่าเบื่อเพราะทุกบริษัทต้องเป็นบริษัทที่มี Big Data และ AI อยู่ในมือถึงจะสามารถแข่งขันทางธุรกิจได้ ธีรวัฒน์ บรรเจิดสุทธิกุล, CFA, FRM, CAIA บทความนี้เป็นความเห็นส่วนตัวของผู้เขียน ไม่เกี่ยวข้องใด ๆ กับหน่วยงานที่ผู้เขียนสังกัดอยู่ การลงทุนมีความเสี่ยง ผู้ลงทุนควรศึกษาข้อมูลก่อนการตัดสินใจลงทุน The comments, opinions and analyses are the personal views expressed by Theerawat Bunjerdsutikul and are intended to be for informational purposes and general interest only and should not be construed as individual investment advice or a recommendation or solicitation to buy, sell or hold any security or to adopt any investment strategy. Investing involves risk, including possible loss of principal. It does not constitute legal or tax advice. The information provided in this material is rendered as at publication date and may change without notice and it is not intended as a complete analysis of every material fact regarding any country, region, market or investment. Reliance upon information in any posting on this site is at the sole discretion of the viewer.

Upgrade ระบบการเงินด้วย Big Data และ AI

1

upgrade-ระบบการเงินด้วย-big-data-และ-ai-1cb441722bf4

2018-09-08

2018-09-08 12:05:34

https://medium.com/s/story/upgrade-ระบบการเงินด้วย-big-data-และ-ai-1cb441722bf4

false

555

null

ลงทุน

18

Champ Theerawat Bunjerdsutikul

Fund Manager

cb6cd4314f17

ChampTheerawat

3

9

20,181,104

null

0

null

0

null

2018-07-16

2018-07-16 09:38:31

2018-07-16

2018-07-16 09:38:28

3

false

en

2018-07-17

2018-07-17 13:52:22

4

1cb4a274db7a

3.780189

2

0

null

4

Which jobs will survive when the robots take over? I’ve started reading an excellent book by “Talk Like Ted” author Carmine Gallo titled “5 stars: From Good to Great”. The opening chapters affirm my thinking about the value of the skill of public speaking. Gallo provides a number of examples — some I have presented below. Gallo suggests that “Persuaders are irreplaceable”. He argues that “the days of being average in business are over. If a computer can recognise average, it can replicate average. Average simply isn’t good enough to stand out in the digital age.” “If you can persuade, inspire and ignite the imagination of others, you will be unstoppable, irresistible and irreplaceable.” he says. He goes on to say that “Emotional connection is, indeed the winning ticket in a world where technologies such as automation, big data, artificial intelligence and machine learning are eliminating millions of jobs and disrupting entire industries, businesses and careers.” His key point in the opening chapter is that in every economic shift — and especially in the digital revolution, communication skills become more valuable and not less. My favourite quote from the opening chapter is “People who know how to talk — and talk well — will be be rewarded, and those with the ability to inspire people — to ignite another person’s imagination — will be especially well positioned”. He also quotes data expert Anthony Goldbloom who says “If you have the ability to communicate ideas that grab people’s attention, then you won’t be replaced anytime soon”. “The unusual person will jump out” — says Warren Buffett. “You will jump out much more than you can anticipate if you get really comfortable with public speaking. It’s an asset that will last you 50 or 60 years and it is a liability if you don’t like doing it.” He goes on to say that mastering the art of public speaking is the single greatest skill a person can acquire today to boost their career in the future. This was music to my ears, as I have spent the last 19 years trying to perfect my craft as a public speaker, with the intent to inspire and influence others in each and every talk. An event last week reminded me that you’re never too old to realise the need to “rise above the noise”, and in this increasingly automated world, especially in the digital revolution — communication skills become more valuable and not less. For nearly 3 years now, I have been involved with the charity set up by ITV journalist Robert Peston called Speakers for Schools. This organisation pairs secondary schools in the UK with speakers to motivate and inspire the next wave of leaders. I spoke to a group of the mixed 6th form at Wilmington Grammar School for Girls. I always start my talk by saying that I can predict which of the students in front of me will be successful, and this event was no exception. After my talk of around 30 minutes where I employed the students to “rise above the noise and I have the students a mini-masterclass on personal branding. I delivered a very similar talk to Ruislip high in December 2017 which you can watch below. After I had finished the talk, there were a number of the students that asked very interesting questions. As I was packing to leave, one student, Billy Ellis approached me and said “I’m doing all of the things that you mentioned in your talk”. He then proceeded to give me his professionally produced business card (one of my personal branding suggestions), and also a copy of his self-published book. I was so impressed with Billy’s achievements I asked if we could take a photo, which I then shared on Twitter. I’ve been wanting to set aside the time for a while now to write a book, and I have already started writing one, but I was spurred on to accelerate my plans given the fact that someone a generation younger than me has already published one. It shows you that you’re never too old to be inspired into rising above the noise yourself, and that no matter what stage you are in your career, communications skills have incremental vale to both you and the organisation you work for. Carmine Gallo’s book has also emphasised how important being able to stand out — by using the art of public speaking or having a well respected personal brand can help future-proof you against the rise of the robots. So answering the question I posed in this post’s title “Which jobs will survive when the robots take over?”, the answer is that it is not always about the position, the type of work, but about your own uniqueness – your own ability to stand out and be different and irreplaceable. In the words of New York Times columnist and bestselling globalization expert Thomas Friedman: “Everyone needs to find their extra – their unique value contribution that makes them stand out in whatever is their field of employment.” What are you doing to future-proof your career?

Which jobs will survive when the robots take over?

2

which-jobs-will-survive-when-the-robots-take-over-1cb4a274db7a

2018-07-19

2018-07-19 13:30:40

https://medium.com/s/story/which-jobs-will-survive-when-the-robots-take-over-1cb4a274db7a

false

856

null

Carmine Gallo

carmine-gallo

Carmine Gallo

0

Andrew Grill

The Practical Futurist. Former Global Managing Partner @ IBM. Husband & Dad. TEDx & International Keynote Speaker. Aussie in London.

ff6b886950a6

andrewgrill

3,389

1,315

20,181,104

null

0

null

0

null

2018-06-26

2018-06-26 06:54:25

2018-06-26

2018-06-26 08:17:26

2

false

en

2018-06-26

2018-06-26 08:26:05

5

1cb4fe4c1824

1.334277

1

0

There are as many dimensions as there are variables, so in a data set with 8 variables the sample space is 8 dimensional. That makes your…

5

Dimensionality Reduction — Principal Component Analysis (The Easy Way) There are as many dimensions as there are variables, so in a data set with 8 variables the sample space is 8 dimensional. That makes your head hurt? It’s tough to visualise. Dimensionality Reduction attempts to distill higher- dimensional data down to a smaller number of dimensions while preserving as much of the variance in the data as possible. Remember K Means Clustering? It is an example of dimensionality reduction algorithm. It reduces data down to K dimensions. Principal Component Analysis: The main linear technique for dimensionality reduction is principal component analysis, performs a linear mapping of the data to a lower-dimensional space in such a way that the variance of the data in the low-dimensional representation is maximised. So the covariance (and sometimes the correlation) matrix of the data is constructed and the eigenvectors on this matrix are computed. The data gets projected onto the hyperplanes, which represents the lower dimensions you want to represent. The greatest variance of the data set comes to lie on the first axis and the second greatest variance on the second axis, and so on.. This process allows us to reduce the number of variables used in an analysis. PCA IN ACTION PCA is really useful for things like image compression and facial recognition, however the most challenging part of the PCA is interpreting the components Fortunately, Scikit Learn makes it very easy to use PCA and you can do it with just 3 lines of code.

Dimensionality Reduction — Principal Component Analysis (The Easy Way)

1

dimensionality-reduction-principal-component-analysis-the-easy-way-1cb4fe4c1824

2018-06-26

2018-06-26 08:26:05

https://medium.com/s/story/dimensionality-reduction-principal-component-analysis-the-easy-way-1cb4fe4c1824

false

252

null

Data Science

data-science

Data Science

33,617

Sammy

null

af012ff9befa

shivam.somani09

7

42

20,181,104

null

0

null

0

null

2018-06-06

2018-06-06 16:10:41

2018-06-06

2018-06-06 16:22:17

0

false

en

2018-06-06

2018-06-06 16:22:17

3

1cb59b7cfc1a

4.166038

3

0

Code for America is a nationwide voluntary organization, founded in 2010 to bring top digital expertise to bear on government projects…

5

Code for America, civic tech and DOD Code for America is a nationwide voluntary organization, founded in 2010 to bring top digital expertise to bear on government projects. (The name and business model were inspired by Teach for America, which recruits new college grads, many from top-ranked universities, to work for two years teaching in disadvantaged communities.) CfA is at the forefront of a movement often called “civic tech” to encourage people with tech skills and interests to work, at least for a while and/or part-time, for projects for government. This is a very important role, given the problems governments have competing for tech talent given yawning gaps between government and the tech industry in salaries, job responsibilities, and the coolness factor. There is a space for civic tech in the tech ecosystem for two reasons: One is the idealism of some young people in tech, who want a chance to work on something more meaningful than the latest photo sharing app. The other is the cultural shift away from one job for life, which means that more are open to a temporary stint helping government. In some sense, CfA is a voluntary-sector counterpart to the U.S. Digital Service and the General Services Administration’s 18F. CfA founder Jen Pahlka actually pitched and worked standing up USDS while on a one-year stint as deputy federal CTO in 2013 under Todd Parks. However, CfA currently works only for state and local government. Sometimes that tech work — which can touch on Medicaid, SNAP (formerly food stamps), criminal justice, and workforce development — includes work with federal agencies that oversee projects in these areas, but the federal government is not the client. Since 2011 CfA has hosted an annual summit as a signature event. The most recent one was at the end of May, with around 1,200 attendees, of whom, Pahlka estimates, 20 percent were federal employees, including USDS and 18F. One of the sessions at this summit included two speakers from the Defense Department, Raj Shah (the first managing director of Defense Innovation Unit Experimental, who left DOD in February but spoke about his DIUx experience) and LTC Enrique Oti, who is the Air Force liaison for DIUx. This was the first time DOD speakers had ever appeared at the summit, which normally centers on the social services and anti-poverty space. The invitation to appear came from Pahlka herself, who serves on DOD’s Defense Innovation Board, and had heard the pair give a presentation on a project DIUx was sponsoring to deal with scheduling and fuel economy in a combined military operations center in Qatar. Pahlka was ultra-impressed by the project, which achieved some of CfA’s general goals, such as producing working software at a fraction of the time and cost that such projects normally require, and using non-traditional, non-defense contractor partners. (The specifics of the actual project are not really the subject of this post, but they are sufficiently interesting that I intend to discuss them in my post next week.) The day after the summit ended, a piece appeared on the website Medium, which publishes lots of tech-related content, called Code for America Summit: Is it still ‘civic tech’ when we help the US Department of Defense? The article is by Aaron Wytze, a master of global affairs student at the University of Toronto who also writes for a Taiwan-based news outlet. It led the next day’s email edition of highlights from Medium for people interested in government, which is why the piece landed in my inbox. Including the DOD Qatar effort in in the summit, Wytze wrote, “broaden(s) the meaning of civic tech into eerie new territory” of war and killing rather than the traditional one of support and succor. The majority of U.S. fighter aircrafts involved in bombing and missile campaigns against ISIS, Wytze wrote, go out of the Qatar base. In a 2017 report, he continued, “DIUx’s digital planning tools were discussed in relation to ‘Operation Inherent Resolve,’ the US bombing campaign to eliminate ISIS forces in Iraq and Syria. The operation in question has had a very high civilian casualty rate, with Human Rights Watch stating that the US-led forces have ‘failed to take necessary precautions to avoid and minimize civilian casualties, a requirement under international humanitarian law.” The post quoted two CfA employees who criticized the invitation on Twitter. “Not all software should be built better and faster,” wrote one. “Machines of war fall into that bucket.” Another tweeted that “civic tech engineers should not take any jobs that end up with bombs dropped on people. Why are we listening to this?” Wytze quoted no supporters of inviting DOD; Pahlka said there were internal folks who supported the invitation and commented on this through private CfA channels, though none on Twitter as far as she knows. This dispute recalls the recent controversy within Google about accepting a DOD contract to use artificial intelligence used to analyze drone video. The Washington Post has reported that Google has decided not to renew the contract when it runs out. I would like to strike a (rare?) blow for civility here. On the one hand, those who don’t want to be involved in defense work should refrain from doing so. And Google, in making a decision for the organization, is perfectly within its rights to withdraw from a DOD AI contract. But I draw the line at suggestions CfA should not invite speakers from DOD, whose messages listeners are then free to embrace or reject. “Certainly the defense of our country’s freedom and independence qualifies as a central civic mission, including the use of tech in that defense.” Withholding invitations to speak is something that should be limited to extreme cases of people with clearly evil ideas, such as Nazis, Klansmen or advocates of terrorism. Even if you disagree with our (democratically determined) military policies, I don’t think that DOD activities come close to the line that merits exclusion. Instead, they are policies on which reasonable people can disagree. Certainly the defense of our country’s freedom and independence qualifies as a central civic mission, including the use of tech in that defense. So my answer to the question of whether DOD activities can qualify as civic tech is, “Sure!” We can and should debate whether U.S. policies in the world are right or wrong. But let’s have our community demonstrate in such debates the civility that is so valuable for society but is such a rare commodity today. Published with permission from Fcw.com. This originally appeared in fcw.com: https://fcw.com/blogs/lectern/2018/06/comment-kelman-cfa-dod.aspx

Code for America, civic tech and DOD

3

code-for-america-civic-tech-and-dod-1cb59b7cfc1a

2018-06-10

2018-06-10 06:12:31

https://medium.com/s/story/code-for-america-civic-tech-and-dod-1cb59b7cfc1a

false

1,104

null

Government

government

Government

28,711

Steve Kelman

Harvard Kennedy School professor, does research on improving government performance. also strong amateur interest in China and learning Chinese

48796854583f

KelmanSteve

592

367

20,181,104

null

0

null

0

null

2018-07-02

2018-07-02 12:25:29

2018-07-02

2018-07-02 12:25:30

0

false

en

2018-07-02

2018-07-02 12:25:30

1

1cb5c45b6bdf

0.335849

0

null

5

Data is often messy or incomplete, requiring human intervention to make sense of it before being usable as input to machine learning projects. This is problematic when the volume scales beyond a handful of records. In this episode Dr. Cheryl Martin, Chief Data Scientist for Alegion, discusses the importance of properly labeled information for machine learning and artificial intelligence projects, the systems that they have built to scale the process of incorporating human intelligence in the data preparation process, and the challenges inherent to such an endeavor. Listen Now!

Integrating Crowd Scale Human Intelligence In AI Projects (Interview)

0

integrating-crowd-scale-human-intelligence-in-ai-projects-interview-1cb5c45b6bdf

2018-07-02

2018-07-02 12:25:30

https://medium.com/s/story/integrating-crowd-scale-human-intelligence-in-ai-projects-interview-1cb5c45b6bdf

false

89

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Boundless Notions

Content and Consulting to make sense of your data

4db925846be1

TobiasMacey

7

2

20,181,104

null

0

from pyspark.sql import SparkSession from pyspark.sql import functions as F from pyspark.sql.types import * import pandas as pd from pyspark.sql import DataFrame from pandas.io import gbq import calendar import matplotlib.pyplot as plt %matplotlib inline import seaborn as sns spark = SparkSession.builder.appName(‘APP_COHORT’).getOrCreate() project_id=xxxx (aca es su proyecto ID de bigquery) def get_month(x): name = calendar.month_name[x] return name get_month_udf = F.udf(get_month) QUERY1 = “”” select cohort,_month,months_from,active_users, total_users, percent_active from ( SELECT c.cohort cohort, c._month _month, INTEGER(c._month) + INTEGER(RIGHT(c.cohort, 2)) — 2 AS months_from, c.actives AS active_users, d.total_users AS total_users, ROUND((c.actives/d.total_users)*100, 2) AS percent_active FROM ( SELECT cohort, _month, EXACT_COUNT_DISTINCT(passenger) actives FROM ( SELECT ROUND((DATEDIFF(DATE(b.requested_at), DATE(a.sign_up))/30),0) AS _month, a.cohort cohort, a.passengerID passenger FROM ( SELECT passengerID, sign_up, SUBSTR(STRING(sign_up), 1,7) cohort, FROM [passengers] WHERE DATE(TIMESTAMP(sign_up))>=’2017–02–01' and DATE(TIMESTAMP(sign_up))<=’2018–02–01' ) a INNER JOIN ( SELECT passengerID, requested_at FROM TABLE_DATE_RANGE(events, TIMESTAMP(‘2017–02–01’), CURRENT_TIMESTAMP()) WHERE AND DATE(TIMESTAMP(requested_at))>= ‘2017–02–01’ and DATE(TIMESTAMP(requested_at))<= ‘2018–02–01’ )b ON a.passengerID=b.passengerID ) where _month>0 GROUP BY 1, 2 )c INNER JOIN ( SELECT SUBSTR(STRING(sign_up), 1,7) cohort, EXACT_COUNT_DISTINCT(passengerID) total_users FROM [passengers] WHERE DATE(TIMESTAMP(sign_up))>=’2017–02–01' and DATE(TIMESTAMP(sign_up))<=’2018–02–01' GROUP BY 1 )d ON c.cohort=d.cohort ORDER BY 1, 2 ASC ) where months_from >0 “”” raw_data = gbq.read_gbq(QUERY1, project_id=project_id) df = spark.createDataFrame(raw_data) df2 = df.withColumn(‘name_month’, get_month_udf(F.col(“months_from”))) df3 = df2.toPandas() fig = plt.figure(figsize=(14, 6)) a = sns.barplot(x=’name_month’, y=’percent_active’, hue=’cohort’, data=df3) a.set_title(‘Cohort Analysis of User Engagement over 12 month period’) a.set_xlabel(‘Month’) a.set_ylabel(‘% Users Active in Cohort’)

12

1979b6cc00ad

2018-06-06

2018-06-06 22:48:26

2018-06-08

2018-06-08 14:54:46

7

false

es

2018-07-17

2018-07-17 04:46:55

4

1cb69b4b5397

5.472642

5

0

En este artículo hablaré de análisis de cohorte para retención de sus clientes con observación en aumento o descenso usando solo Bigquery y…

3

Cohort usando Bigquery y Pyspark En este artículo hablaré de análisis de cohorte para retención de sus clientes con observación en aumento o descenso usando solo Bigquery y para mostrar Pyspark ¿Qué es el análisis de cohortes? El análisis de cohorte es un subconjunto de análisis de comportamiento que toma los datos de una determinada plataforma de comercio, y, en lugar de considerar a todos los usuarios como una unidad, los divide en grupos relacionados para su análisis. Estos grupos relacionados, o cohortes, generalmente comparten características o experiencias comunes dentro de un lapso de tiempo definido. El análisis de cohortes es una herramienta para medir el compromiso del usuario a lo largo del tiempo. Ayuda a saber si la participación del usuario en realidad está mejorando con el tiempo o si solo parece mejorar debido al crecimiento. El análisis de cohorte es valioso porque ayuda a separar las métricas de crecimiento de las métricas de participación, ya que el crecimiento puede enmascarar fácilmente los problemas de participación. En realidad, la falta de actividad de los usuarios antiguos está siendo ocultada por el impresionante número de nuevos usuarios, lo que permite ocultar la falta de compromiso de un pequeño número de personas. El análisis de cohortes tienes varias utilidades pero en este caso usaremos para saber como esta reteniendo la empresa a sus usuarios Uso de cohort Ingreso periódico recurrente Los nuevos costos de adquisición de usuarios se suman. Siempre es mejor retener a más clientes con un valor de vida más alto que tratar de buscar clientes más pequeños con vidas más cortas. Una vez que entienda cuánto gana por mes y cuánto quema, puede hacer planes y predicciones para sus productos, servicios y negocios. Tasa de conversión de visitante a prueba. Es el elemento Act clásico del marco RACE. Cuanto mayor sea el número de visitantes (Alcance) y cuanto mayor sea su Porcentaje de conversiones en este paso, más usuarios se habrá registrado para su prueba gratuita. Por lo general, la siguiente fase del embudo está transformando a los usuarios de prueba en clientes de ingresos recurrentes. Tasa de conversión de prueba a pago. Una vez que finalice la prueba, querrá saber cuántos de esos usuarios están lo suficientemente satisfechos con su producto / servicio como para querer pagarle mensualmente. Aquí es clave comenzar a hablar y escuchar a sus clientes lo antes posible, para que puedan hacerse una idea de cuáles son sus puntos débiles y ayudarlos en consecuencia. Visitantes únicos mensuales. No solo la cantidad es importante aquí, sino también la calidad de ese tráfico; una mejor ventaja vale más que dos malas. Los usuarios por semana o mes cuando se inscribieron Esto lo ayudará a comprender qué canal es el mejor para la adquisición del usuario, en términos de calidad y costo Usuarios por fuente de tráfico Cuando se relacionaron por primera vez con su sitio web de manera significativa, se registraron para el boletín / webinar / demo, comenzaron una prueba gratuita La hora se activó por primera vez En base a esto, pensará en formas de recompensar a los usuarios premium o reducir los límites de prueba gratuitos Frecuencia / frecuencia de uso Los usuarios fueron referidos al sistema o se inscribieron por su cuenta ¿Qué es el bigquery? Bigquery es una base de datos de google tiene funcionamientos que son familiares a las bases de datos relacionales y a las no relacionales. Actúa sobre conjuntos de datos (datasets) de un peso considerable para el humano medio a través de lenguaje SQL (un poco cocinado, ya que tiene funciones propias) con relativa rapidez. Lo que hay detrás de esto, para los más curiosos, es el motor de búsquedas Dremel. para mas información puede ver aquí Este sera un ejemplo sencillo, un análisis practico para observar como es la atracción de los clientes durante el periodo de 1 año (MENSUAL), también te puede ayudar a observar en una siguiente análisis el CLV de su cliente, dentro un grupo de cliente potenciales Ejemplo bigquery usando sql Imagen de bigquery 8GB en 36 segundos Registro lo puedes obtener haciendo aquí Usando Pyspark existen 2 formas Podríamos descargar el csv y leerlo Podríamos simplemente ejecutar el query directo con Spark creería la la opción 1 es la mas fácil y la mas común, pero en esta ocasión usaremos la opción 2 , usando una consulta directa el detalle de esto es que si deseamos un análisis vaya ejecutando diario en notebook, seria una buena opción para que las áreas como Marketing o Comercial u otra pueda observar como va el día a día y que podría hacer si no logro lo cometido durante el día, semana , mes, etc. Importar librerías Definir la session en spark Definir el proyecto de bigquery Funcion UDF Crear una variable llamada QUERY1 y pegamos la instrucción sql Crear un dataframe usando spark Cambiamos en nombre del cambo en meses usando la funcion UDF convertimos a pandas para la visualización Usado maplotlib para la visualización Observaciones: Se observa que durante los meses la retenciones han comenzado el subida , llego unos meses que bajaron y luego tiene una creciente inmensa a la mitad del año y para los últimos meses del año esta comenzando a descender . posiblemente hubo un percance en las áreas o lo mas seguro no tuvieron un análisis de cohort para poder medir su desempeño como área encargada de retener usuario =). Libro para estudio fuente(https://amzn.to/2xQdGaV) Espero le hayas gustado esta publicación sobre Análisis de cohort usando bigquery con Pyspark, ya comenzare con unas aplicaciones usando Pytorch linkedin: Jonathan Quiza | LinkedIn View Jonathan Quiza’s profile on LinkedIn, the world’s largest professional community. Jonathan has 10 jobs listed on…www.linkedin.com Gracias

Cohort usando Bigquery y Pyspark

81

cohort-usando-bigquery-y-pyspark-1cb69b4b5397

2018-07-17

2018-07-17 04:46:55

https://medium.com/s/story/cohort-usando-bigquery-y-pyspark-1cb69b4b5397

false

1,172

Publicación de Ciencia de Datos, Machine Learning, Deep Learning, Inteligencia Artificial y mucho más en Español. Compartiendo conocimiento para hacer de este mundo un lugar mejor :)

null

Ciencia y Datos

favio@cienciaydatos.org

datos-y-ciencia

CIENCIA Y DATOS,CIENCIA DE DATOS,DATA SCIENCE,MACHINE LEARNING

null

Pyspark

pyspark

Pyspark

132

Jonathan Quiza

Data Scientist student forever| Developer | Machine Learning | Deep Learning Enthusiast

6373afa3bbcf

jonap

69

24

20,181,104

null