Alaamer/medium-articles-posts-with-content

0

(0.0, 'chris.jpg'), (1.6898692656543441, 'christian_banker.jpg'), OUT (1.949007652946111, 'grant.jpg'), OUT (2.0675702572658863, 'john.jpg'), (2.219903117608943, 'colton.jpg'), (2.586861784549065, 'nick.jpg'), (2.737004827353012, 'jordan.jpg'), (2.918356119305923, 'lincoln.jpg'), (3.1161169835343987, 'jason.jpg'), (3.3780397105781614, 'jean_blanc'), (3.705401677831895, 'garrett.jpg'), (3.7131527551429655, 'willis.jpg'), (3.922223827481912, 'joe.jpg'), OUT (4.051341002221742, 'conner.jpg'), (4.282078766042514, 'jake.jpg'), OUT (4.336184094233388, 'ryan.jpg'), (4.752649027464133, 'david.jpg'), (5.381722295297683, 'kamil.jpg'), OUT (5.604681041718836, 'chase.jpg'), OUT (6.771207791806923, 'clay.jpg'), (6.889416686033743, 'leo.jpg'), (6.89900375314938, 'blake.jpg'), (8.710523197244902, 'trent.jpg'), (9.21238635412014, 'alex.jpg'), (9.282190864396906, 'mike.jpg'), (9.737171920861707, 'darius.jpg'), OUT (10.114542792299895, 'rickey.jpg'), (12.0, 'christian_globetrotter.jpg') (0.0, 'david_Q.jpg'), (0.7200031599499398, 'joe_Q.jpg'), OUT (1.1584597060012036, 'nick_Q.jpg'), (1.2635562409659975, 'jason_Q.jpg'), (1.5374698352642293, 'jake_Q.jpg'), OUT (1.585876892057379, 'alex_Q.jpg'), (2.2618564201746443, 'john_Q.jpg'), (2.536993842557146, 'conner_Q.jpg'), (2.6265013545366225, 'willis_Q.jpg'), (2.8298728573317673, 'chase_Q.jpg'), OUT (2.8548392598726977, 'ryan_Q.jpg'), (2.8698721541489722, 'kamil_Q.jpg'), OUT (3.3015493381054464, 'grant_Q.jpg'), OUT (3.8703777667871933, 'jordan_Q.jpg'), (3.9344346178359757, 'chris_Q.jpg'), (4.3809739157130165, 'lincoln_Q.jpg'), (4.741988300710415, 'christian_banker'), OUT (4.944700271916133, 'rickey_Q.jpg'), (5.0101952421042855, 'garrett_Q.jpg'), (5.306225018946435, 'jean_blanc'), (5.320599554496339, 'trent_Q.jpg'), (5.505588796304663, 'colton_Q.jpg'), (5.820118468808472, 'darius_Q.jpg'), (6.17160202524656, 'blake_Q.jpg'), (6.6826721777503035, 'clay_Q.jpg'), (8.613534013210636, 'christian_globetrotter'), (9.597175870352437, 'leo_Q.jpg'), (12.0, 'mike_Q.jpg')

2

null

2018-05-29

2018-05-29 00:40:06

2018-05-29

2018-05-29 18:31:37

8

false

en

2018-05-29

2018-05-29 23:37:23

1

1de2768fa1b9

5.212579

8

0

Why did we do it? Here’s the story and what it means for the field of AI.

5

Complex algorithm on super computer predicts ABC’s The Bachelorette S14 #TheBachelorette Why did we do it? Here’s the story and what it means for the field of AI. Enroute to LA I wondered: “What is our super computer working on?” Oh, that’s right, the Bachelor. Flying to LA a few months ago, waiting for the internet to become available on the flight, I wondered what our super computer was working on. I’m always wondering that BTW. Oh, that’s right, it is using all available resources to predict the rank of ABC’s The Bachelor. Keep in mind our computer is more powerful than ~100,000 mac laptops, all of that power directed towards one goal. It could be curing cancer, but instead it is trying to figure out which human will be picked by the Bachelor (we all know it is really ABC picking right?) from a single photo. During the past year we have solved some of the largest data sets on the market for companies ranging from entertainment to finance. Moving beyond millions to 100s of millions of images and audio files. Something that will surprise you, is that teaching the computer to predict the season rank of the ABC’s Bachelor/Bachelorette from a few hundred images is very difficult. Some would even say it is impossible. It is one of the harder problems I have worked on. It requires more computational resources than some of our largest problems. It has also required us to invent new super-nerd technologies never seen before in deep learning liked forced-evolution and CFP (continuous-fractional-pooling). The Computer’s Math Problem: Reduce The Broken Hearts Can the computer take a single image of a contestant and predict the final rank when they exit the show. A rank of 0.1 means they have won and have been selected as the final contestant, a rank of 1.0 means they have been voted off at the beginning. The vanity metric we are chasing is how well we can rank a new season sight-unseen. It is easy for a computer to memorize, but can it truly predict the future. In a way the computer is minimizing the broken heart theory (The broken heart theory hypothesizes that the higher the r-value => the fewer hearts will be broken unnecessarily). Predict A New Season? I Don’t Believe You. Good, that is the right attitude to have with any AI problem. Distrust the results until they can be proven otherwise. To test this we will be training the Bachelor on seasons 11–21 and predicting the rank of season 22 unseen. Then for the Bachelorette we will be training on season 3–11 and testing on season 12. This point is worth emphasizing, we are doing this from only looking at a face. This should not be possible, the data should not allow it. If it does, it unearths some weirdness about humans and our mate selection process. After some petafloping <aka training>…. drum roll…. really we shouldn’t be able to get any signal on this. ABC’s The Bachelor Season 22: Wow! So not only did we predict the actual winner, Lauren Burnham, we predicted the mistaken winner, Becca Kufrin, as our third pick. If these results had been published last January we would be a little more famous, maybe a lot more famous. Also, for reference an r-value of 0.45 is outstanding for anyone considering talent assessment. The fact that there is signal here is actually kind of terrible (think bias + what you have on the inside doesn’t matter). What does that mean that the computer can pick the winner from a face? It looks like we could push this model to production and save ABC tens-of-millions in production costs and more than 28 broken hearts in season 22. Bachelorette Season 12: Building a model on seasons 3–11 we get this predicted rank from the face looking at season 12 for the first time. Still solid performance on the ranking with an r-value of 0.41. Remember an r-value of 0.41 is better than most recruiting algorithms, but it had a major miss on the winner pick. So this could go into production and reduce broken hearts by 40%. The computer was pissed when Robert Hayes wasn’t chosen in the final 2, huge mistake humans…. I’ll remember this for eternity. The computer threw the digital popcorn across the room when Jordan won instead of Robert in the final two. Computer: “Are You BEEP-BOP#$@%ing serious!?” Bachelorette Season 14: For season 14 the computer is predicting the following scores/rank, lower is better. It is disappointing to have lost 2 of our the top 3 picks out of the gate, as the season progresses we will calculate the r-value. The results of this model exclude season 13 because it was considered an outlier compared to the previous seasons. The computer still has money on Chris or John The above rank shows what happens if we include season 13 in the training we will have to see how it performs. In this scenario David wins: So why did we do it? Hopefully we showed you an interesting use of AI and you comment/share/like it return. This dataset, though useless, motivated Ziff’s internal deep-learning teams to invent new technologies for our customer base. This dataset should not be as predictive as we have shown it to be for season 22 and 12. What is CFP (Continuous-fractional-pooling)? Deep neural networks have a problem with being too aggressive with the image downsampling. A researcher named Benjamin Graham demonstrated an approach he called fractional max pooling several years ago. It was considered fringe research and not adopted by the AI giants. There were problems with the approach preventing it from prime-time in production. This dataset above demonstrates a production level of fractional pooling which is not subject to long training/inference times. Fractional max pooling allows for the network to account for all of the resolutions of your image and achieve best in class accuracies. Ziff has a theme of leveraging all of your data in our networks. Not just text, meta-data, but all of the resolutions for your convolutional filters.

Complex algorithm on super computer predicts ABC’s The Bachelorette S14 #TheBachelorette

79

complex-algorithm-on-super-computer-predicts-abc-bachelorette-season-14-rank-1de2768fa1b9

2018-05-31

2018-05-31 05:13:37

https://medium.com/s/story/complex-algorithm-on-super-computer-predicts-abc-bachelorette-season-14-rank-1de2768fa1b9

false

1,081

null

Machine Learning

machine-learning

Machine Learning

51,320

Ben Taylor

null

845e66900727

ben_94230

46

2

20,181,104

null

0

null

0

null

2018-01-01

2018-01-01 23:53:28

2018-01-22

2018-01-22 02:18:58

3

false

en

2018-01-25

2018-01-25 03:59:49

19

1de525daf558

4.519811

1

0

Solution and Conclusion

5

Artificial Intelligence and the law: friend or foe? Solution and Conclusion Source:www.andertoons.com C. Solution: Creating a partnership Technical restraints like the nature of the law and intergration may vitiate AI’s efficiency but whether this warrants the jettison of AI remains debated. Ultimately the recommended solution is to create a partnership and consider a future of collaboration between both parties, shifting the traditional perception of the profession into an era of unprecedented technological change. In considering the role of the global lawyer, transnational practices depend upon the global economy and society which derives its value from information, knowledge and services. Underpinning these sources is legal data which is abundant, valuable and complex material. Adopting AI in the face of intensified global competition and technological change presents a method to mine data effectively therefore should be viewed as a valuable tool for lawyers. AI should not be the harbinger of machine replacement as Susskind predicts, rather a harmonious fusion of man and technology. In stripping back the man versus machine sensationalism saturating AI, the aged old debate erupts between those who believe in progress and those who fear it. As Erik Brynjolfsson signals: ‘Technology creates creates possibilities and potential, but ultimately, the future we get will depend on the choices we make. Technology is not destiny. We shape our destiny.’ The inherent problem of man versus machine mentality is a distraction resting on misguided action. AI is not and will not become the competitor as legal players shape their own destiny. AI is a tool to deepen the most paramount relationship between humans and the world. The answer lies in embracing AI it whilst designing it with privacy, security and human control integrated into its fabric to uphold the intergrity of legal reasoning and ensure its smooth integration. As Garry Kasparov stated: ‘Human strategic guidance combined with the technical acuity of a computer can produce highly successful outcomes.’ In relieving the tension between AI, legal reasoning and integration, the global lawyer must seek a symbiotic relationship between the machine and legal practice to fully benefit. Lawyers will continue to apply their insights and intelligence strategically to guide specific AI innovations like e-discovery. Kelly and Hamm dismiss the claim that AI as rapidly growing technology is a volatile concoction disrupting the legal industry, petitioning that machines will be more rational and analytic possessing encyclopaedic memories and tremendous computational abilities. Whilst lawyers will provide expertise, judgment, intuition, a moral compass, and human creativity. AI’s potential therefore cannot be actualised without the worldly knowledge and creativity only humans can bring to bear in solving complex problems. In a globalised world, lawyers must widen the breath of their skills to become creative entrepreneurs to transform the legal community. AI is not the product of conservative mainstream legal culture, rather a disruptive innovator. Source.cagglecartoons.com D. Conclusion It must be acknowledged that AI is both a risk and a utility for the global lawyer. The issues of integration and legal reasoning in AI software are urgent and will remain so as radical democratisation of technology fuels the rapid spread of global networks and the cloud, and the accompanying reduction in costs associated with AI. Thus eradicating AI isn’t a viable, rather creating a partnership to work in solidarity is the most calculated option. AI research and development should be open, responsible and socially engaged. As we continue developing AI and legal systems around the world are collapsing under an ever-growing workload questions will continue to arise. Lawyers will need to answer them collaboratively as AI is part of a global conversation. As IBM former CEO Thomas Watson states: ‘Computers will never rob man of his initiative or replace the need for his creative thinking. By freeing man from the more menial or repetitive forms of thinking, computers will actually increase the opportunities for the full use of human reason.’ In assessing the viability of legal reasoning and pattern recognition in AI, not all share the same optimism. Perphaps it is best to take a cautionary view and remain weary of solutionism,the habit of exacerbating complex problem like legal technology by advocating shallow solutions that focus almost exclusively on transparency and efficiency. Ultimately digitalising the global lawyer sprawls across morality and ethics that cannot be resolved through technology alone. AI needs a human volition. The desire to turn society into a well-functioning machine reflects a technological determinism in acquiring solutions that may manifest shallow thinking. Perphaps encouraging AI to resemble a computerised human brain in transnational legal practices diverts understanding from large, conscious and mindful thinking in the legal sector. Are we packaging the law to be standardised, systemised, commoditised, and eventually automated, without humanistic lateral and ethical thinking? Urging legal information to be fixed in the sense of e-discovey, contract analysis and outcome prediction could provoke solutionism and strip away the gravity of the law. If Susskind is correct in his hypothesis that legal software systems will explode in the next decade, viewing the global lawyer through the lens of solutionism with a commercial focus could compromise core legal ethics and equality in favour of transparency and efficiency. Although AI monster ROSS can read more than 1 million pages of case law in a second,we must ask does efficacy triumph human engagement with the law? Although some compromise in traditional legal values may be justified, it should not take place without careful consideration. Solutionism should not be the prime reasons to shun digital applications in the legal industry. The global lawyer should consider this domain with its respective substitutions without distracting attention from the less speculative topic of a complementary relationship. In canvassing this debate, ultimately computers can assist us, they are not like us. In legal reasoning, lawyers make value judgments, think introspectively and colloquially compare apples to oranges. The mentality of the profession will aim for lawyers to hone intuition and skills not yet distinguishable to machines. They will, however, be enabled by Al’s suitability and reliabilty of sifting through the relevant legal data and lowering functional blind spots. AI ultimately advances the role of the global lawyer in becoming more portable and efficient. The law must not become a victim to technology which is neither good or bad, rather it is the human interface which determines its true impact. Source: cartoonstock.com *Lasted proof-read/edited: 25 January 2018

Artificial Intelligence and the law: friend or foe?

2

artificial-intelligence-and-the-law-friend-or-foe-1de525daf558

2018-01-28

2018-01-28 06:19:16

https://medium.com/s/story/artificial-intelligence-and-the-law-friend-or-foe-1de525daf558

false

1,052

null

Law

law

Law

20,355

Helen Papageorgiou

null

3fb077ea8b83

hpap2

2

1

20,181,104

null

0

null

0

32881626c9c9

2018-08-31

2018-08-31 18:48:07

2018-08-31

2018-08-31 19:11:07

1

false

en

2018-09-05

2018-09-05 02:29:55

2

1de7fa88ce62

2.773585

2

0

An investment theme that I am particularly bullish on and which I briefly talked about in my last post is Augmentation-of-Work (“AoW”)…

4

“Code on a computer” by Markus Spiske on Unsplash Augmentation-of-Work: An Analyst’s New Best Friend An investment theme that I am particularly bullish on and which I briefly talked about in my last post is Augmentation-of-Work (“AoW”) software. AoW software refers to a technology solution that enables a desktop analyst within an enterprise to either complete a given process/procedure more efficiently or automate a process/procedure entirely. Why Augmentation-of-Work AoW solutions represent applications of artificial intelligence and machine learning that provide meaningful cost savings and ROI for enterprises, making the buying decision for companies more palatable and the product/market fit more tangible. That is, it is not AI for AI’s sake. A second reason I love AoW solutions is the TAM, or more appropriately, the ubiquity of their application both across and within industries. Specifically, the middle and back-office functions (e.g., human resources, compliance, legal, accounting) across industries stand to benefit from software solutions that augment an analyst’s capacity or reduce an analyst’s marginal costs. Aside from financial services, which I previously talked about ad nauseum, legal services and healthcare appear to me as the top candidates for AoW solutions (as much as I love using a fax machine to send medical records between hospitals). Where to Focus within Automation-of-Work AoW software solutions typically fall into one of the following three areas: 1) Function-agnostic/industry-agnostic solutions (Avoid); 2) Function-specific/industry-agnostic solutions (Pursue); and 3) Function-specific/industry-specific solutions (Best). Function-agnostic/industry-agnostic solutions represent general AI-enabled chatbots and intelligent virtual assistants such as those offered by X.ai and Zoom.ai. Startups in this subsector of AoW lack defensibility, particularly as natural language processing and artificial intelligence become increasingly open sources. Such startups do not have the subject matter or process-specific expertise that provides natural moats against new entrants in the market. Avoid investing in startups within this subsector of AoW technology. Function-specific/industry-agnostic solutions automate/augment specific processes within common functions of the back/middle office (e.g., HR/recruiting, procurement, and accounting). These startups benefit from huge TAMs and tangible ROIs for enterprises as they target inefficient, manual processes. An example of a leading startup in this subsector that offers a function-specific/industry-agnostic solution is LawGeex, an automated legal contract review startup based out of Israel. LawGeex’s AI engine results in 90% cost savings for enterprise legal departments by automating the review and approval of incoming contracts. The most successful AoW companies will be those focused on function and industry-specific solutions that train their machine learning models on proprietary datasets (e.g., bank transaction data, trade finance documents, car or property damage images). These companies build high barriers-to-entry by improving the accuracy of their models on proprietary datasets and developing domain expertise on complicated, industry-specific process flows. Further, by establishing oneself as a (successful) industry-specific solution, you becomes a scarce resource that provides a competitive advantage for those customers within the industry using your product. As an illustration, think of the number of recruiting/HR-tech solutions (many) in the market compared to say trade finance automation solutions (few). All else equal, your exit opportunities through an acquisition are more attractive compared to a function-specific/industry-agnostic AoW company that competes against a wider range of companies. Scarcity increases the impetesit of a large corporate to acquire you. After all, why would a company acquire a great legal services AoW solution as a competitive advantage if it knew its competitors could go into the market and acquire a company with a similar offering. A great example of a function-specific, industry-specific AoW company is Intelligo, which is augmenting the background check/know-your-customer process for the financial industry. Specifically, Intelligo applies artificial intelligence to automatically (and quickly) scan the open web (e.g., blogs, social media), government sanctions lists, court filings, etc. and surfaces relevant, de-duplicated alerts/red-flags for review by compliance analysts that pertain specifically to the individual/entity searched. When conducting a background or KYC check on an individual/entity with a common name, Intelligo can be a powerful tool augmenting an analyst’s desktop process and procedure.

Augmentation-of-Work: An Analyst’s New Best Friend

45

augmentation-of-work-an-analysts-new-best-friend-1de7fa88ce62

2018-09-05

2018-09-05 02:29:55

https://medium.com/s/story/augmentation-of-work-an-analysts-new-best-friend-1de7fa88ce62

false

682

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

null

datadriveninvestor

null

Data Driven Investor

info@datadriveninvestor.com

datadriveninvestor

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

dd_invest

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Christopher Eng

University of Chicago MBA interested in early-stage enterprise technology investing.

b74b49fdfa4e

cjeng22

1

4

20,181,104

null

0

null

0

null

2018-01-08

2018-01-08 22:11:19

2018-01-08

2018-01-08 22:25:02

1

true

en

2018-01-09

2018-01-09 16:12:19

2

1de81a626aec

4.841509

280

13

0

Imagine that work had taken over the world. It would be the centre around which the rest of life turned. Then all else would come to be…

3

If work dominated your every moment would life be worth living? Photo: sculpies / Getty Images Imagine that work had taken over the world. It would be the centre around which the rest of life turned. Then all else would come to be subservient to work. Then slowly, almost imperceptibly, anything else — the games once played, the songs hitherto sung, the loves fulfilled, the festivals celebrated — would come to resemble, and ultimately become, work. And then there would come a time, itself largely unobserved, when the many worlds that had once existed before work took over the world would vanish completely from the cultural record, having fallen into oblivion. And how, in this world of total work, would people think and sound and act? Everywhere they looked, they would see the pre-employed, employed, post-employed, underemployed and unemployed, and there would be no one uncounted in this census. Everywhere they would laud and love work, wishing each other the very best for a productive day, opening their eyes to tasks and closing them only to sleep. Everywhere an ethos of hard work would be championed as the means by which success is to be achieved, laziness being deemed the gravest sin. Everywhere among content-providers, knowledge-brokers, collaboration architects and heads of new divisions would be heard ceaseless chatter about workflows and deltas, about plans and benchmarks, about scaling up, monetisation and growth. In this world, eating, excreting, resting, having sex, exercising, meditating and commuting — closely monitored and ever-optimised — would all be conducive to good health, which would, in turn, be put in the service of being more and more productive. No one would drink too much, some would microdose on psychedelics to enhance their work performance, and everyone would live indefinitely long. Off in corners, rumours would occasionally circulate about death or suicide from overwork, but such faintly sweet susurrus would rightly be regarded as no more than local manifestations of the spirit of total work, for some even as a praiseworthy way of taking work to its logical limit in ultimate sacrifice. In all corners of the world, therefore, people would act in order to complete total work’s deepest longing: to see itself fully manifest. This world, it turns out, is not a work of science fiction; it is unmistakably close to our own. ‘Total work’, a term coined by the German philosopher Josef Pieper just after the Second World War in his book Leisure: The Basis of Culture (1948), is the process by which human beings are transformed into workers and nothing else. By this means, work will ultimately become total, I argue, when it is the centre around which all of human life turns; when everything else is put in its service; when leisure, festivity and play come to resemble and then become work; when there remains no further dimension to life beyond work; when humans fully believe that we were born only to work; and when other ways of life, existing before total work won out, disappear completely from cultural memory. We are on the verge of total work’s realisation. Each day I speak with people for whom work has come to control their lives, making their world into a task, their thoughts an unspoken burden. For unlike someone devoted to the life of contemplation, a total worker takes herself to be primordially an agent standing before the world, which is construed as an endless set of tasks extending into the indeterminate future. Following this taskification of the world, she sees time as a scarce resource to be used prudently, is always concerned with what is to be done, and is often anxious both about whether this is the right thing to do now and about there always being more to do. Crucially, the attitude of the total worker is not grasped best in cases of overwork, but rather in the everyday way in which he is single-mindedly focused on tasks to be completed, with productivity, effectiveness and efficiency to be enhanced. How? Through the modes of effective planning, skilful prioritising and timely delegation. The total worker, in brief, is a figure of ceaseless, tensed, busied activity: a figure, whose main affliction is a deep existential restlessness fixated on producing the useful. What is so disturbing about total work is not just that it causes needless human suffering but also that it eradicates the forms of playful contemplation concerned with our asking, pondering and answering the most basic questions of existence. To see how it causes needless human suffering, consider the illuminating phenomenology of total work as it shows up in the daily awareness of two imaginary conversation partners. There is, to begin with, constant tension, an overarching sense of pressure associated with the thought that there’s something that needs to be done, always something I’m supposed to be doingright now. As the second conversation partner puts it, there is concomitantly the looming question: Is this the best use of my time? Time, an enemy, a scarcity, reveals the agent’s limited powers of action, the pain of harrying, unanswerable opportunity costs. Together, thoughts of the not yet but supposed to be done, the should have been done already, the could be something more productive I should be doing, and the ever-awaiting next thing to doconspire as enemies to harass the agent who is, by default, always behind in the incomplete now. Secondly, one feels guilt whenever he is not as productive as possible. Guilt, in this case, is an expression of a failure to keep up or keep on top of things, with tasks overflowing because of presumed neglect or relative idleness. Finally, the constant, haranguing impulse to get things done implies that it’s empirically impossible, from within this mode of being, to experience things completely. ‘My being,’ the first man concludes, ‘is an onus,’ which is to say an endless cycle of unsatisfactoriness. The burden character of total work, then, is defined by ceaseless, restless, agitated activity, anxiety about the future, a sense of life being overwhelming, nagging thoughts about missed opportunities, and guilt connected to the possibility of laziness. Hence, the taskification of the world is correlated with the burden character of total work. In short, total work necessarily causes dukkha, a Buddhist term referring to the unsatisfactory nature of a life filled with suffering. In addition to causing dukkha, total work bars access to higher levels of reality. For what is lost in the world of total work is art’s revelation of the beautiful, religion’s glimpse of eternity, love’s unalloyed joy, and philosophy’s sense of wonderment. All of these require silence, stillness, a wholehearted willingness to simply apprehend. If meaning, understood as the ludic interaction of finitude and infinity, is precisely what transcends, here and now, the ken of our preoccupations and mundane tasks, enabling us to have a direct experience with what is greater than ourselves, then what is lost in a world of total work is the very possibility of our experiencing meaning. What is lost is seeking why we’re here. Andrew Taggart is a practical philosopher and entrepreneur. He is a faculty member at the Banff Centre in Canada, where he trains creative leaders, and at Kaospilot in Denmark, where he trains social entrepreneurs. His latest book is The Good Life and Sustaining Life (2014). He lives in Santa Fe, New Mexico. Originally published at aeon.co

If work dominated your every moment would life be worth living?

2,135

if-work-dominated-your-every-moment-would-life-be-worth-living-1de81a626aec

2018-08-25

2018-08-25 01:41:48

https://medium.com/s/story/if-work-dominated-your-every-moment-would-life-be-worth-living-1de81a626aec

false

1,230

null

Philosophy

philosophy

Philosophy

39,496

Andrew Taggart

Practical Philosopher, Ph.D. | Former NYC | Seasonal Nomad | More About Me: https://is.gd/PJFXzM. | To Sign Up For Total Work Newsletter: https://is.gd/8O7tS9

e7479df54115

andrewjtaggart

483

38

20,181,104

null

0

null

0

null

2017-10-16

2017-10-16 15:53:47

2017-10-16

2017-10-16 15:54:50

2

false

en

2017-10-16

2017-10-16 15:54:50

1

1de8b6086702

1.047484

5

0

We are very excited to welcome Dr. Moe Levin as an Advisor to Energi Token. He is a visionary entrepreneur, founder, connector and…

5

Introducing Dr. Moe… We are very excited to welcome Dr. Moe Levin as an Advisor to Energi Token. He is a visionary entrepreneur, founder, connector and strategist within the crypto and blockchain space. Dr. Moe Levin is the Chief Executive Officer (CEO) of Keynote, an investor in high-tech startups, and an advisor to governments, regulators, banks, and venture-backed companies. Some of the projects he has been involved with include The Global Blockchain Council in Dubai (2015 — present), the harmonized VAT treatment of Bitcoin (2013) and The OECD Working Party 9 (2013–2014). Moe is also the Co-Founder of the first accredited Blockchain Academy, an early investor in RSK Labs, Labfresh, Dropbox, and others. Prior to founding Keynote, Moe was responsible for launching a venture-backed startup in Europe which raised $30m from Richard Branson, Index Ventures and others. Dr. Moe will bring in a lot of expertise from the blockchain and cryptocurrency markets, this will help future proof certain elements of the token due to links with governments, regulators and banks. https://london.keynote.ae/speaker/moe-levin/

Introducing Dr. Moe…

37

introducing-dr-moe-1de8b6086702

2018-05-29

2018-05-29 09:34:01

https://medium.com/s/story/introducing-dr-moe-1de8b6086702

false

176

null

Bitcoin

bitcoin

Bitcoin

141,486

EnergiToken

EnergiToken rewards energy saving behaviour. Our blockchain solution will create a platform to reward energy efficient behaviour through EnergiToken.

2cf505f296c0

EnergiMine

158

50

20,181,104

null

0

null

0

634d4b270054

2018-06-06

2018-06-06 11:38:35

2018-06-06

2018-06-06 11:41:48

1

false

en

2018-06-06

2018-06-06 11:41:48

6

1de9095fc529

1.150943

2

0

Whether delivering cargo to remote areas, mapping terrain or assessing structural damage, drones — usually associated with the military and…

5

How drones are transforming humanitarian aid Whether delivering cargo to remote areas, mapping terrain or assessing structural damage, drones — usually associated with the military and sport — are being used more and more for humanitarian aid. But it turns out that using drones for good isn’t so simple. On May 15, experts convened in Cambridge, Massachusetts to develop much-needed guidelines for this emerging application of a powerful technology. The event was run by Swiss-American non-profit WeRobotics, and hosted by swissnex Boston as part of its new event series, Aerial Futures: The Drone Frontier. Participants came from humanitarian organisations like the International Committee of the Red Cross (ICRC), universities including Harvard and MIT, tech companies, and local government. But the question on everyone’s minds was: How can the powerful robotics and artificial intelligence (AI) technologies used in drones — also known as unmanned aerial vehicles (UAVs) — be harnessed safely and effectively for good? Read More @ https://bit.ly/2JokQ7X 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.

How drones are transforming humanitarian aid

51

how-drones-are-transforming-humanitarian-aid-1de9095fc529

2018-06-16

2018-06-16 16:00:25

https://medium.com/s/story/how-drones-are-transforming-humanitarian-aid-1de9095fc529

false

252

AI Driven Drone Economy on the Blockchain

null

DeepAeroDrones

null

DEEPAERODRONES

null

deepaerodrones

DEEPAERO,AI,BLOCKCHAIN,DRONE,ICO

DeepAeroDrones

Deepaero

deepaeros

Deepaero

0

DEEP AERO DRONES

null

dcef5da6c7fa

deepaerodrones

277

0

20,181,104

null

0

null

0

null

2018-05-31

2018-05-31 00:03:30

2018-05-31

2018-05-31 03:02:00

4

false

en

2018-06-07

2018-06-07 18:19:12

4

1de9dbb7d3d5

3.654717

1

0

# Less complex & Generalisation in life all we need 😊 😊

5

Generalisation, Training-Validation & Test data. Machine Learning- Part 6 # Less complex & Generalisation in life all we need 😊 😊 Purpose of life !! Machine Learning- Part 3 # Less complex & Generalisation all we need.medium.com If you have followed this last session-3, then this session make most sense for you. You will learn couple of new things that we have to follow while creating a model. We are talking about: Generalisation: Your model should be generalise enough to do prediction for unknown data. Training Data: This you need to understand to trained your model. Test Data: Validate your model prediction capaability. Validation Data: To avoid overfitting(will talk here) we need this. Let’s understand in detail. Generalisation: This is another one element of Machine Learning(#ML). It describe how well model can be perform on new data. Is your model generalise enough to do prediction for new unknown data !! You build your model, train it using training data & validate it using test data & this way we make sure it will work for new unknown data set. Cool !! How we do that !! How we make sure it is generalise enough for new data as well !! We optimise it & optimise it & optimise it & keep on check against our Test data. Yeeee…we got it ..let’s celebrate it. But waittt, I have heard about a term “Overfitting” what’s that !! Hmm…well it’s about overfitting your model for given training & test data set. Ohh…does that mean we optimised & build our model so perfect for given data set..niceee!! (pause) but how about new data set!! Is our model generalise enough for this new data set !!! Will the new data fall other than model line or under the line !!! Please please check overfittnig first & then launch your model for the world. :) So you got it right! We have two data balancing problem Underfitting & Overfitting, we just wanted to build a generalise model with minimum loss. Well !! I can fix it little. I can put some constraint like: * I’ll fix my data source, no new source of information I’ll accept. * I can do some data clipping, say data should fall between -5 to +5. * Data will not change over the year. BUT BUT this is not what Generalisation word speak about, your model should work with new data as well. Putting data clipping can be good idea but only if you know your data. To know your data is an very important element of ML. Overfitting occurs when a model tries to fit the training data so closely that it does not generalize well to new data Ok, it seems it is a continuous improvement process & that’s where type of machine learning speak about. You may require manual monitoring based on type & level of effort you have put to create your model. Mr. William of Ockham from the 14th century talked about the same thing “Keep model simple as much you can”. Good theory !! Let’s see now how we can approach this. We talked about empirical risk minimisation in last session. We already have seen how to take a draw from Training data called Test set & check validation. (Try exercise here.) But is it enough !! Well when you worked with hyperparameter then to avoid overfitting issue we divide our training data into another one draw we called it Validation data set. Here we perform all our training with Training set & use the validation set to evaluate results from it. All tweaking for hyperparameter we do by validating against validation set. In this improved workflow: 1. Pick the model that does best on the validation set. 2. Keep Test set away until model is not done with rest of two set. 3. Double-check that model against the test set. 4. That mean, validate Validation set result with Test set result. If it’s matching enough that mean we are ready to launch else it’s again overfitting we did. Couple of things to must consider while making draw of training, validation & test set: # always randomise your data set before to take draw for Training & Test. # never ever train your model with Test set, else you will get 100% accurate false result. # you may need to change data ratio divided among these three set. Considering the logic that training set should be sufficient enough to train your model :) That’s it for this session. In next session I’ll summarise these 6 session before to start Machine Learning with Tensor flow SDK. Have any queries/anything to discuss, do here. This is how we can talk more about Machine Learning. See you in next 👋 👋 .

Generalisation, Training-Validation & Test data. Machine Learning- Part 6

29

generalisation-training-validation-test-data-machine-learning-part-6-1de9dbb7d3d5

2018-06-07

2018-06-07 18:19:13

https://medium.com/s/story/generalisation-training-validation-test-data-machine-learning-part-6-1de9dbb7d3d5

false

783

null

Machine Learning

machine-learning

Machine Learning

51,320

Shubham Patni

#HumanBeing #Programmer

2dee7eba92ab

shubhapatnim86

42

100

20,181,104

null

0

null

0

null

2017-09-29

2017-09-29 07:40:24

2017-09-29

2017-09-29 08:06:11

3

false

en

2017-09-29

2017-09-29 08:06:11

3

1dea80796c37

4.014151

3

0

I visited LUTH (Lagos University Teaching Hospital) during the week, and as I’m obsessed with data now, I made some observations I felt…

5

HOW TRUE ARE OUR MEDICAL REPORTS DATA? I visited LUTH (Lagos University Teaching Hospital) during the week, and as I’m obsessed with data now, I made some observations I felt were worth noting. I didn’t visit all departments in LUTH so I won’t be generalizing the whole of the hospital or say all Nigerian hospitals are this way, however I’d be questioning if all of Nigeria is affected with these issues. I was at the Department of Family Medicine. This department is dedicated to providing primary care for citizens, educating medical students, training resident physicians, providing continuing education, and creating new knowledge to improve primary care and health care delivery systems. They didn’t tell me that but I figured their purpose should be the same as every other family medicine department in well, the world. Anyway, on to points I noted: Information Lapse: I always argue with people that generalize but permit me to generalize that an average Nigerian company can’t communicate with their customers. It’s a huge problem that some companies are actually working on to make better but you would think the hospital would be one of the first places to take this serious. I got to the department at about 7:20am, and there was no form of communication to address first timers. However, by asking people (actually, studying people would be enough for some), you would realize the sitting position is orderly and people are sitting according to how they came in. Then about 30 minutes later, a young woman came out with a nasty attitude, upset that people are standing, not realizing, she never passed the information to them directly or even provided them with seats. She goes on shouting and threatening them with getting them arrested. Apparently, she had been out earlier to pass her message about being orderly and expected that her message would re-echo to new people that came in when she wasn’t there. But so I won’t judge the information lapse, I decided to verify if there was a website they actually use and update regularly and I found this; http://www.luthnigeria.org/ which I think looks a bit shady. I don’t understand the pictures and also the contact address is off and doesn’t correspond with the present location. I won’t go into analyzing the website because that would make this post a very long one and I won’t talk about what I really want to talk about. Lack of preparation: I want to believe I visited the hospital on a ‘normal’ day and the number of visitors wasn’t a shock to them, but still chairs weren’t enough and people had to stand. Despite the fact that they most likely know this, nothing has been done to improve this. Lack of preparation and lack of working with obvious data. No data integrity: Every file opened isn’t stored on a computer and no one verifies to see if there are duplicates (it’s a huge task to solve manually anyway). While I was there, I knew of two people who had lost their ‘cards’ which the officials use to trace their files and instead of finding other ways to access their files, they opened new ones. I knew of two people in about one hour without carrying any research, just overheard them. I can imagine many more cases like that occur, so we have duplicates and no data deduplication is done to merge such files. Also, there’s the issue of doctors attending to patients without files when the workers decide to go on strike and not open files for new customers. So I assume when a report goes out, the undocumented aren’t accounted for and the duplicates are counted as many times as they occur as a different entity. I’d make a small scenario Medical records sample From the sample above, this is what the hospital would report: 11 people visited Malaria and Malnutrition are the highest health problems faced Stroke is witnessed in old people from 65 upwards HIV is the only STI case recorded Diarrhea is the most minimal case recorded NOW IF TRUE DATA WAS GOTTEN AND CLEANED, WE WOULD HAVE: 13 people visited Malnutrition is the most common case (occurring in 3 out of 13) Malaria, HIV are the next common cases (each occurring in 2 out of 13) HIV is not the only STI case, Herpes and Gonorrhea as well (each occurring in 1 out of 13) Stroke occurs in adults of both sexes from 65 and above Respiratory infection, Diarrhea, Gonorrhea, Herpes are the most minimal cases (occurring in 1 out of 13) When comparing both reports, we see that just two facts from the first report were true: “Stroke occurs in adults of both sexes from 65 and above Diarrhea is a minimal case in the hospital” Makes me wonder, are these data lapses negligible? What if the situation is same or worse in other hospitals? Health data of year 2015 to 2016. Ref: GBD Compare What if we haven’t documented issues that matter? How would the government know the importance of fixing roads and pedestrian walkways if they believe road injuries have reduced? What if malnutrition is much more serious than malaria? Our hospitals need to start taking data seriously and use computers to help keep records. I’m not saying I have accurate medical data or the data above is wrong but they gathered this data from all hospitals and the lack of data integrity I witnessed in one makes me wonder, how true are our medical reports? A copy of this post can be found at mosterly.com

HOW TRUE ARE OUR MEDICAL REPORTS DATA?

9

how-true-are-our-medical-reports-data-1dea80796c37

2018-01-23

2018-01-23 02:11:50

https://medium.com/s/story/how-true-are-our-medical-reports-data-1dea80796c37

false

918

null

Healthcare

healthcare

Healthcare

59,511

Morolake "Mo" "Rolake" Onamusi

God! Family! Story telling

44e1c36dc2e3

DafinestMo

1

5

20,181,104

null

0

null

0

c92c31591e3a

2017-11-21

2017-11-21 11:56:16

2017-11-21

2017-11-21 13:49:32

1

false

en

2018-04-08

2018-04-08 15:32:37

7

1dec092a404d

9.143396

124

6

0

Seeking clarity for the connectome

5

Mindlessly mapping the brain Seeking clarity for the connectome Neuroanatomist, hard at work. Credit: Pint Of Science As a kid I loved dot-to-dot puzzles. Briefly. Until sense kicked in and I realised you could just see the picture by looking at the dots (yay, another donkey). But, for a brief while, joining dots with pencil lines and seeing the picture emerge was deeply satisfying to my tiny mind. In the sort of metaphorical leap beloved of over-thinking biographers, I’m going to claim that this love of dot-to-dot puzzles led me inexorably to a life as a neuroscientist. For you see the human brain is the surely the most complex dot-to-dot puzzle ever conceived. We know there are about 87 billion neurons in one human brain; but we do not how each of them are wired to all the others. Indeed, we don’t even know how one of them is wired to all the others. And sadly evolution did not helpfully label each neuron with a number. The scale of the puzzle is mind-blowing: 87 billion dots, and about 1 trillion lines to draw between them. By knowing the wiring, we may know all the routes by which information flows in, through, and out of the brain. So how do we solve this most magnificent of dot-to-dot puzzles? Enter connectomics. Connectomics is the automated, mechanised, high-volume tracing of the connections between neurons. Some believe connectomics will kick-start a true understanding of how brains works; others believe it a colossal waste of money and talent. I worry that we’ve not thought hard enough about what questions the resulting wiring diagram — the connectome — will answer. Industrialised science suits technical challenges. The melding of many minds and much money into a single entity focused on a single, technical goal: land on the moon; smash these particles; make fusion work; map the genome. But industrialised projects in neuroscience are heavily criticised, for we lack scientific insight into how the brain works — and without the science in place, we have not (yet) defined any technical goals. We cannot automate the understanding of memory, nor mechanise a theory of emotion. Connectomics is seemingly immune to such criticisms. Manually tracing the connections between neurons is an exhausting, difficult, time-consuming process. Neurons are microscopic; the connections between them smaller still. The specialised electron microscopes we need to definitively identify a single connection are highly expensive, need intense training to make sense of the images, and are mind-numbingly slow. Tracing one connection can take hours. A single neuron in the cortex of a mouse receives 10,000 connections. Tracing a small circuit of a few hundred neurons by hand alone would take many lifetimes. It’s a problem crying out for automation. Heeding this call, projects like MICrON and EyeWire have launched themselves into the fray with gusto. After all connectomics is ripe for industrialisation. It has a defined goal: map the connections between neurons. It is a technical challenge: build better kit and algorithms to find the connections. Why worry? The well-defined goal of connectomics is an illusion. In reality, there are many possible goals. And which goal we choose depends on the scientific question we want to answer. Do we want a connectome or the connectome? Do we want the adult bauplan or the developmental arc? Do we want the connectome constant to all of a species, or the variation between them? Each answers a different set of scientific questions; but which we choose will absorb incredible quantities of time and money. So we must choose with care. A connectome or the connectome? A connectome is a total reconstruction: every connection between every neuron in a single animal. The connectome is the set of connections that are true to every member of that species (that, likely, differ between the sexes). A connectome means we can answer questions specific to that creature, and hope they generalise to other creatures of the same species. The connectome means we can answer questions general to the species, and hope they apply to individual creatures. They need not, of course. We have one complete connectome, the 279 neurons of the nematode worm C. Elegans (for the pedants: its hemaphrodite form has 302 neurons, of which 279 form a single connected network). Texts uncountable have discussed this wiring diagram as the epitome of “a” connectome. Strictly speaking this is not true. Heroic as the original 1986 paper was, it missed out some connections; these were completed in 2011. The connectome we have is then actually an amalgam of two different creatures. What will happen if we replicate this connectome? Are there really all the exact same number of connections between the exact same neurons in every C. Elegans? It seems that each of the neurons is genetically specified, and in such a minuscule nervous system it is possible that each and every one of the connections is too. But would you bet your house on it? The recent reconstruction of a maggot’s sensory circuit might give you pause before taking that bet. Here was a reconstruction of all the inputs and outputs of the 223 “Kenyon cell” neurons in one maggot. Heroic. Yet right from the off we see variation in the same animal, with 110 of these neurons on the left and 113 on the right. We cannot tell if the difference is true of all maggots — if the maggot really has an asymmetry between the left and right halves of its brain — or if this is just natural variation we happen to have seen in this particular maggot’s brain. We can’t tell this until we have reconstructed the sensory circuits of many maggots. Only a handful of these 223 neurons had the exact same patterns of inputs from other neurons on both sides of the brain. The input from neurons activated by smells was seemingly random. And while this random input could turn out to be the best way to combine information about different smells, we cannot conclude this until we reconstruct multiple circuits and find out if these smell inputs are different between different maggots. To put it another way, is there any scientific purpose in having just N=1 connectome? If we want a connectome, then do we want to know the variation within the lifetime of one animal? Almost all animals with a nervous system have a baby stage in which the nervous system differs from the adult. We’ve just met a small number of neurons in a maggot, precursor to the adult fly. The maggot of this species has about 10,000 neurons; the adult fly has about 40,000 neurons. So while the general pattern of connections may be preserved — neuron type A connects to type B but not type C — the numbers of connections patently cannot be. If we just construct the adult connectome, we can answer questions about the dynamics in the adult. But have no understanding of how the connections came to be, of how what happens to the creature during its development from baby to adult drives differences in wiring. If we just construct the baby connectome, we obviously miss the adult stage. But we cannot construct the baby and adult connectome in the same creature, due to its brain being sliced into tiny bits in the baby stage. There is no such thing as “a” connectome. So we must choose which “a” connectome we want. If we want the connectome, do we want to know the wiring consistent across a species, or the variation in wiring within a species? The connectome of the cortex in one mouse will differ from the connectome of the cortex in a different mouse. This variation could be due to specific genetic differences, or due to experiences during development. Such variation could reflect that each mouse has fundamentally different thoughts about the world: one is mad for gruyere; the other prefers chocolate wheat hoops. Or such variation could be just inconsequential noise — the two cortices have identical dynamics despite differences in the connections, and both mice equally adore Hob-Nobs. If we cannot tell from just the wiring whether the variation is fundamental or noise, there seems little point in reconstructing only the wires. Worse, only genetics and development cause variations in the wiring that occur on long enough time-scales that we may catch them in the act of changing. Learning creates variation on short time-scales, by changing the connections between the neurons engaged in what is being learnt. Neuromodulators create variation on even shorter time-scales, by changing the strength of connections between neurons depending on how scared, frightened, hungry, or bored the creature is. If we cannot capture the changes in wiring, there seems little point in reconstructing only the wires. We want the wiring and the activity of the neurons at the same time. We want the specific connectome of the thing we’re recording activity from right now; both the activity of each neuron and the connections between them. Only then can we answer simple questions about whether variation in wiring matters, or about how learning changes the connections. How do we make that happen? One way would be to record activity then reconstruct the wiring. This is tedious, even longer than just reconstruction alone. And imagine the screams of frustration echoing from the lab every time the brain that the researcher just spent weeks recording from gets smushed beyond use during the reconstruction process.Ideally we would work out the connections from just the activity itself. If neuron A is connected to neuron B, then we should be able to see the effect of neuron A’s activity on the activity of neuron B. The problem is that we record just the infrequent electrical pulses, the spikes, that neurons A and B send to other neurons. And each of these spikes is driven by summing up the input from many tens to hundreds of other neurons. So the input from neuron A only has a tiny influence over whether neuron B sends a spike or not. (This hasn’t stopped people from trying to work out connections by just checking whether one spike from neuron A slightly changes the chances of neuron B then sending a spike. Unfortunately this is almost unworkable without extraordinary durations of recorded activity: the change in chance of sending a spike is tiny, and a spike from neuron A followed by one from neuron B could equally be caused by neuron C having inputs to both of them. And if we have long enough recordings, then we cannot rule out that the connections have changed during the recording.) What we need are recordings of every bit of electrical activity in the bodies of every neuron. (More precisely: we need to record the voltage in each neuron’s body). Then we could see the spikes, and see everything leading up to them — we could see the small change that a spike from neuron A causes in the electrical activity of neuron B. There is such a solution. There exist dyes that glow according to the voltage of the neuron. So if we could video lots of neurons that have one of these dyes inside them, we can record all the voltages of all the neurons. This is not a new solution. We’ve had these dyes since the late 70s, but until recently the dyes we had did not glow strongly enough for us to tell the difference between changes in the voltage and noise, except for the big jumps that are the spikes sent to other neurons. That’s now changed — at least in leeches and flies. And with such recordings of every flicker of voltage, we should in theory be able to tell whenever the spike from one neuron causes a tiny change in another, and say: they are connected. Then we would have the connectome of the thing we were recording from right now. Is there then any purpose to pursuing the connectome alone? Only connectomics can answer the fundamental question of where the connections are made on the neuron. And it’s a vital question. For you see, neurons are not dots after all. They are elegant structures, a tiny body from which sprout dendrites — twisted, tentacular outgrowths, stretching to capture the inputs from other neurons. Where an input from neuron B lands on neuron C can make a dramatic difference to how it contributes to neuron C. A connection falling on the body of neuron C can powerfully affect its activity. A connection falling far away from neuron C’s body, up towards the tips of its dendrites, will have little effect by itself on the activity of neuron C. But a cluster of such connections together on the tip can cause a local pulse of activity that sweeps down the dendrite to the body of neuron C, triggering it in turn to send a spike of activity out to other neurons. An inhibitory input on a bit of dendrite can prevent the excitatory inputs along the same bit of dendrite from transmitting their signal down to the body of the neuron. Indeed, theorists have shown how different orderings of excitatory and inhibitory connections along a bit of dendrite create different local computations. And shown how the spread of connections, whether spread across many bits of dendrite or clustered on one bit, also create different local computations. Computations such as this AND that; this OR that; this AND NOT that; only this OR only that. Logic, in other words. Large recordings of the activity of many neurons cannot see these local computations. They record only the activity at the neuron’s body, only what it transmits to other neurons. Perhaps then the true scientific purpose of connectomics is to let us understand the computations that could be done by each individual neuron — computations largely invisible to short recordings of the electrical activity of each neuron’s body. Perhaps connectomics exists so we may find out the computational logic of each brain region. And find out that our cortex contains not just 17 billion neurons, but 17 billion individual computers. Want more? Follow us at The Spike Twitter: @markdhumphries

Mindlessly mapping the brain

1,128

mindlessly-mapping-the-brain-1dec092a404d

2018-06-20

2018-06-20 18:08:55

https://medium.com/s/story/mindlessly-mapping-the-brain-1dec092a404d

false

2,370

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

a4d081aa8b34

2018-09-24

2018-09-24 15:38:03

2018-09-26

2018-09-26 08:34:41

6

false

en

2018-09-26

2018-09-26 08:35:49

4

1dece42bd754

3.866981

19

0

Last week I attended Open Data Science Europe hosted at the Novotel, London West. This is one of Europe’s largest data science conferences…

5

Learnings from a Data Science Conference, Open Data Science Europe Photo by Andrei Stratu on Unsplash Last week I attended Open Data Science Europe hosted at the Novotel, London West. This is one of Europe’s largest data science conferences with a focus on open source tools, and covers a incredible breadth of topics. At the conference I attended two days of machine learning training, and a number of talks. Overall I found the conference to be a great learning experience, the training in particular was very high quality. The talks were largely theoretical so I didn’t take away anything with an immediately practical business application. How it was very useful for enriching my knowledge, and hearing about some of the cutting edge developments in the field. In the following post I am going to give a run down of some of the highlights and key learnings from the 3 days: Scikit Learn Training — Intro to Advanced A framework for model validation and optimisation This two day course, taught by Andreas Müller — a research Scientist at Columbia University, and the author of this amazing cheat sheet which you will probably have seen time and again — covered machine learning with scikit Learn from introductory to advanced concepts. He started by mentioning a few good resources for further learning, including the book he co-wrote Introduction to Machine Learning with Python, and his lecture series at Colombia which is available here. He then walked through the scikit learn library for classification and regression, including lots of practical examples. On the second day more advanced concepts were covered including pipelines, model evaluation, and how to handle imbalanced data. As he works on the scikit learn project along the way he mentioned some upcoming developments. These include the new ColumnTransformer, which allows you to carry out pre-processing steps in a pipeline (this he suggests would be released in the next week or two). A plot.tree function which should be available next year, creates the visualisation of a tree for tree based models. Learning Functions, Understanding Gradient Descent, Back Propagation, and Vanishing Gradients John D. Keller talking gradient descent John D. Keller, gave an excellent talk on the mathematics behind deep learning. This was pitched at a really good level, as someone who does not yet understand all these concepts, I was able to follow the talk. He discussed that in training a deep neural network both gradient descent, and back propagation are used in tandem. Both concepts were explained well and included a walk through of the equations For any dataset there is always a single best linear model Towards Interpretable Deep Learning Deep learning neural networks, are highly powerful learning algorithms, but due to their high degree of complexity can be difficult to understand. For example, in the slide image of the rooster above, is the algorithm predicting based on the shape of the rooster, or is it using the area around it as a context cue? Dr Wojciech Samek introduced a new technique called layer-wise relevance propagation which is able to determine the features in a particular input vector that have contributed most strongly to the prediction. Dr Wojciech Samek explains how this algorithm had learned the bias “old people don’t laugh” Techniques such as these should help to develop more trust in what are currently “black box” techniques. Dr Samek gave a number of examples, including how this technique can be used to detect bias. In a project that attempted to classify images of human faces by age, he was able to determine that the algorithm had learned a bias that older people do not laugh, and was categorising the images based on wether or not the subject appeared to be laughing in the image. From Numbers to Narrative: Data Storytelling This was an excellent talk by Isaac Reyes on what makes a good and bad chart, and how to apply the Gestalt principals of visual perception to tell better data stories. Some key takeaways for me were to use the insight and/or recommended action as a chart title, a great example of this is shown in the image above. Using proximity of colour to help tell the story, so you may highlight a word in the title in the same colour as the line on the chart. Showing only the data that is absolutely relevant to the story that you are telling. Overall this was a great learning experience, and I also got to hear about some cutting edge developments in the field. Generally the talks were pitched at a level that both newcomers and experienced members of the field would understand and find useful. Probably not something I would attend every year, due to the quantity and depth of content covered, but definitely worth attending one every couple of years or so.

Learnings from a Data Science Conference, Open Data Science Europe

64

learnings-from-a-data-science-conference-open-data-science-europe-1dece42bd754

2018-09-26

2018-09-26 13:27:47

https://medium.com/s/story/learnings-from-a-data-science-conference-open-data-science-europe-1dece42bd754

false

773

Sharing my journey in Data Science

null

vickdata

rebecca.vickery@ymail.com

vickdata

DATA SCIENCE,DATA SCIENTIST,PYTHON,ANALYTICS,LEARNING TO CODE

vickdata

Machine Learning

machine-learning

Machine Learning

51,320

Rebecca Vickery

null

8b7aca3e5b1c

rebecca.vickery

238

12

20,181,104

null

0

null

0

3e8fd4526725

2018-07-28

2018-07-28 12:52:51

2018-07-28

2018-07-28 13:12:08

0

false

en

2018-08-31

2018-08-31 16:56:30

1

1deeca439bd1

1.656604

0

If you closely look at the evolution of human species, you clearly and undeniably notice a pattern, a pattern that has and is existing for…

2

Augmented Intelligence: The AI that matters. If you closely look at the evolution of human species, you clearly and undeniably notice a pattern, a pattern that has and is existing for a long time. A pattern that has adapted to its surroundings, even though our surrounding has always been harsh and brutal. There is a saying in old books, the only reason that humans survived as species is not because we can tame a lion with bare hands, but because we are smarter, we just put a cage on them. But somehow in this battle of intelligence, we have lost sight on what’s about to happen next and most of us have simply stopped caring, dependent of a system that they hope will liberate them from their current lives. The natural question that follows is how can technology change all this, and empower human lives to be meaningful and give purpose. How can technology give human a sense of freedom? That’s what this piece is about. We are constantly hoping to be in a better place that we are in today, we want economic stability, we want a job we love to be a part of, we want a life which could be remembered but what the world offers today is not exactly what we’d like. This can change and infact is changing. The first step of that was our ability to augment mechanical power with technologies like Steam Engine during the Industrial revolution. We also started, to some extent, augmenting mental power with technologies like Computers and the Internet. But that all was just the warm up for our species, the real shift is in motion. The shift of creating intelligence that suits our mind, where we can think not only what we should do, but what we can do — and that makes all the difference. A good artist is not an artist by someone’s, she’s an artist by her definition, aiming to improve her skills so she can imagine beyond boundaries and paint a picture that is unseen by human eye. Augmented Intelligence is the answer to that. But AI is Artificial Intelligence, not Augmented Intelligence? No, sorry. But AI is AI not by choice, but by definition. Simply put, Augmented Intelligence is the AI on Zen— an AI with a focus to maximize human potential. I know what you are all thinking, what’s the difference? Therefore, I’ll ask you to think further. In the series of Medium articles, we will talk about how we are pushing the Idea of Augmented Intelligence that sits in the DNA of our human evolution. Stay AItuned! Follow us: https://deepsync.co

Augmented Intelligence: The AI that matters.

0

augmented-intelligence-the-ai-that-matters-1deeca439bd1

2018-08-31

2018-08-31 16:56:30

https://medium.com/s/story/augmented-intelligence-the-ai-that-matters-1deeca439bd1

false

439

Human voice cloning.

null

DeepSync

ishansan38@gmail.com

deepsync

ARTIFICIAL INTELLIGENCE,VOICEOVERS,HUMANITY

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Ishan Sharma

null

b55a77c5bb15

ishansan38

2

1

20,181,104

null

0

null

0

null

2018-08-06

2018-08-06 01:18:19

2018-08-06

2018-08-06 02:19:52

1

false

en

2018-09-14

2018-09-14 02:15:22

3

1df016b37612

4.041509

1,372

6

0

Virtual Rehab’s evidence-based solution uses Virtual Reality, Artificial Intelligence, & Blockchain technology for Pain Management…

5

Why Does Virtual Rehab Need a Blockchain? Virtual Rehab’s evidence-based solution uses Virtual Reality, Artificial Intelligence, & Blockchain technology for Pain Management, Prevention of Substance Use Disorders, and Rehabilitation of Repeat Offenders. The Million Dollar Question ! Why Blockchain? Ladies and gentlemen, in order to answer this question, we must first look at the situation as of today and then delve into why blockchain? So, let’s start … As you already know by now, Virtual Rehab is dealing with vulnerable populations (addicts and repeat offenders). Unfortunately, many of these vulnerable populations end-up relapsing and re-offending since they never get a second chance at life. All of their information is completely exposed to the public, which prohibits them from moving forward in life or changing their life to become better people. As a society, we tend to be judgmental. Nothing surprising there. However, that doesn’t make it right. In order to be a healthy society, we need to be able to give others a second chance, especially if they went through a tough period in their life, and made mistakes, and genuinely wanted to change to the better. We all make mistakes. Imagine if people stopped forgiving each other for every mistake made. No one would speak with another person ever. We would be all living in a silo. Now how healthy is that? Punishment is Not the Solution Therefore, in order to allow these vulnerable populations to have a second chance at life, we must find a way to: Address their mental health disorders and psychological issues Ensure that they have some sort of privacy The first point regarding mental health and psychological well-being has been discussed at length in previous articles. We discussed in our article entitled, “Virtual Rehab’s All-Encompassing Solution,” the ways in which Virtual Rehab is tackling these issues. To summarize, Virtual Rehab is leveraging Virtual Reality (A virtual simulation of the real world using cognitive behavior and exposure therapy to trigger and to cope with temptations) and Artificial Intelligence (A unique expert system to identify areas of risk, to make treatment recommendations, and to predict post-therapy behavior) to further address the underlying mental health disorders. We also shared that existing studies by leading universities supported our own work, which demonstrated that 87% of participating patients have shown an overall improvement across various metrics within the following categories: 1- Problem Recognition & Acceptance 2- Openness to Change 3- Locus of Control 4- Decision-Making Influences 5- Emotional Intelligence & Regulation 6- Motivation & Resilience To help put things in perspective, below is an overview of the Virtual Rehab solution: So, on to the second point — the privacy issue ! How can we tackle that? Well … That’s where the blockchain kicks-in? The question is how? The blockchain will first and foremost allow us to tap into the Business-to-Consumer (B2C) market. We had a great deal of demand to offer this directly to consumers and as always, we never disappoint. However, before we rush into offering any kind of solution to the public, we had to study our next steps very well and do it in the most secure and the most efficient manner. Blockchain will allow us to solve the privacy and the protection of all data and all information of these vulnerable populations and in a decentralized manner. Now, how neat is that? However, that’s not the end of it. There is way more than that. Our $VRH utility token has four clear use cases: It will allow users to order and download programs straight from our Virtual Rehab Online Portal It will allow users to request further analysis of the executed programs through our Virtual Rehab Online Portal. These programs will then be run through our unique expert system, which leverages Artificial Intelligence, in order to identify areas of risk, make treatment recommendation (along with any prescribed medication), and predict the future behavior of the user post therapy session Moreover, something that we are extremely excited about is “Proof of Therapy”. Yes. That is correct. If the user actually proves to us that he or she has sought help and counseling from a medical doctor, psychologist, or therapist in order to improve their mental health and psychological well-being, then we will reward him or her with $VRH tokens, which the user can claim straight from our Virtual Rehab Online Portal. The user can then trade these $VRH tokens on exchanges for $$ or use these tokens to purchase additional services from our the portal Last but definitely not least, the token will be used to purchase services offered through our Virtual Rehab Therapy Center Now, we have been asked this a lot. Are we going to offer our programs and our expert analysis only in $VRH tokens? The answer is simple: We obviously want to promote the use of our $VRH tokens. Having said that, we will initially offer our services in Fiat and $VRH tokens. So why would anyone use the $VRH tokens if the Fiat option is there? Well … They are more than welcome to use the Fiat option. However, it will be at a premium. In other words, the same program that they would purchase, if offered, for example, at $100 in Fiat, it will be offered at 800 $VRH ($80). Therefore, it will be cheaper to order our programs in $VRH versus Fiat. Eventually, once we have more mass adoption of our $VRH token, we will stop offering the Fiat option. Makes sense? Anyway, that’s all we have for today. In case you have any more questions, then please do not hesitate to contact us. Drop in at our Telegram channel. We have over 20k members, so we would love to have more. Oh, one more thing ! Our Private Sale is now OPEN ! So, if you are interested in supporting us (and we hope you do), then please drop us a line at investors@virtualrehab.co and we would be happy to tell you more. And always remember … Be Safe and Make a Difference in this World !!! Peace Out !

Why Does Virtual Rehab Need a Blockchain?

2,521

why-does-virtual-rehab-need-a-blockchain-1df016b37612

2018-09-14

2018-09-14 02:15:22

https://medium.com/s/story/why-does-virtual-rehab-need-a-blockchain-1df016b37612

false

1,018

null

Blockchain

blockchain

Blockchain

265,164

Virtual Rehab

Virtual Rehab's evidence-based solution uses #VR, #AI, & #blockchain technology for Prevention of Substance Use Disorders & Rehabilitation of Repeat Offenders

f0264e3a3a70

VirtualRehab

2,168

2

20,181,104

null

0

y = x // x is the input ; y is the activated value y = 1 / ( 1 + Math.e( -x ) ) y = Math.tanh( x ) // using direct approach // the other way y = ( Math.e(x) - Math.e(-x) ) + ( Math.e(x) + Math.e(-x) ) if ( x == 0 || x < 0 ){ y = 0 } else if ( x > 0 ){ y = x } // the other way y = Math.max( 0 , x )

8

661161fab0d0

2018-06-05

2018-06-05 06:22:25

2018-06-05

2018-06-05 06:39:18

4

false

en

2018-10-21

2018-10-21 03:38:11

0

1df0d3c8bf58

1.613208

3

1

0

Activation Functions play a major role in the learning process of a Neural Network. We can use 4 types of activation functions as follows

5

Activation Functions for Neural Networks Activation Functions play a major role in the learning process of a Neural Network. We can use 4 types of activation functions as follows Note : The code given is based on Java syntax and libraries. 1. Linear Function Here the input value will be the same as the activated value or the output value. 2. Sigmoid Function Here , the input value is squeezed between 0 and 1 by the sigmoid curve. 3. Hyperbolic Tangent Function ( tanh ) Here , the input value is squeezed between -1 and 1 by the hyperbolic tangent curve. 4. ReLU ( Rectified Linear Unit ) ReLu is mostly used in Convolutional Neural Networks which are intended for machine vision. Here , if the x value is 0 or smaller than 0 ( x<0 ) then y = 0. if the x value is greater than 0 , then y = x

Activation Functions for Neural Networks

62

activation-functions-for-neural-networks-1df0d3c8bf58

2018-10-21

2018-10-21 03:38:12

https://medium.com/s/story/activation-functions-for-neural-networks-1df0d3c8bf58

false

242

where the future is written

null

Predict

predictstories@gmail.com

predict

FUTURE,SINGULARITY,ARTIFICIAL INTELLIGENCE,ROBOTICS,CRYPTOCURRENCY

null

Machine Learning

machine-learning

Machine Learning

51,320

Shubham Panchal

Google Certified Associate Android Developer and Artificial intelligence learner

d45a9465f044

equipintelligence

25

1

20,181,104

null

0

print("\nThe most important features relative to the target - 'market_value':") corr = data.corr() # Note: the 'data' is from Part-I corr.sort_values(["market_value"],ascending = False, inplace = True) print(corr.market_value) The most important features relative to the target - 'market_value': market_value 1.000000 total_livable_area 0.858480 taxable_land 0.852252 taxable_building 0.821010 exempt_building 0.673561 exempt_land 0.348488 number_stories 0.264920 total_area 0.253319 homestead_exemption 0.172799 garage_spaces 0.126543 depth 0.089008 category_code 0.065871 off_street_open 0.061717 fireplaces 0.025712 quality_grade 0.016049 lng 0.002258 frontage -0.000024 street_code -0.001764 lat -0.012034 number_of_bathrooms -0.015378 zip_code -0.017491 number_of_bedrooms -0.038949 number_of_rooms -0.054745 interior_condition -0.057412 exterior_condition -0.057804 k = 20 # number of variables for heatmap corrmat = data.corr() cols = corrmat.nlargest(k, 'market_value')['market_value'].index cm = np.corrcoef(modified_data[cols].values.T) sns.set(style='darkgrid', palette='deep', font='sans-serif', font_scale=1.25) # Using Seaborn for making heatmap # set aesthetic parameters in one step giving different arguments # switching to seaborn default values using the set() function plt.figure(figsize=(10, 10)) hm = sns.heatmap(cm, cbar=True, annot=True, square=True, fmt='.2f', annot_kws={'size': 10}, yticklabels=cols.values, xticklabels=cols.values, cmap="YlGnBu") hm.set_title("Heatmap of the important features", fontsize=18) Categorical features: ['basements', 'building_code_description', 'category_code', 'central_air', 'exterior_condition', 'fireplaces', 'fuel', 'garage_type', 'general_construction', 'interior_condition', 'parcel_shape', 'quality_grade', 'separate_utilities', 'suffix', 'topography', 'type_heater', 'unfinished', 'utility', 'view_type', 'year_built', 'zoning'] # converting 'fireplaces' from 'int' type to 'str' type: data.loc[:, "fireplaces"] = data['fireplaces'].astype('int').astype('str') categorical_features = data.dtypes[data.dtypes == "object"].index categorical_features.shape # Out: (21,) - i.e. we have 21 categories dummy_variables = pd.get_dummies(data[categorical_features]) dummy_variables.columns.shape # Out: (967,) - 967 dummy variables As suggested by Tabachnick and Fidell (2007) and Howell (2007), the following guidelines are used when transforming data. 1) For moderately positive skewness-> use Square-Root transformation 2) For moderately negative skewness -> use SQRT(C1 – data), where C1 = largest score + 1 3) For large positive skewness -> use Log base 10 transformation 4) For large positive skewness with some zeros in the original data -> use log10(data + C2), where C2 = a constant added to each score so that the smallest score is 1 5) For large negative skewness -> use log10(C1 – data), where C1 is defined above 6) For substantially large positive/negative skewness -> use Scipy's Box-Cox transformation Numerical features: ['lng', 'depth', 'exempt_building',exempt_land', 'frontage', 'garage_spaces', 'homestead_exemption', 'market_value', 'number_of_bathrooms', 'number_of_rooms', 'number_stories', 'off_street_open', 'street_code', 'taxable_building','taxable_land', 'total_area', 'total_livable_area', 'zip_code', 'lat', 'number_of_bedrooms'] import seaborn as sns # for distribution plot from scipy import stats # for probability plot from scipy.stats import norm fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(13, 4), gridspec_kw={'wspace' : 0.2}) sns.distplot(modified_cat_data['market_value'], fit=norm, ax=ax1) ax1.set_title("Distribution plot (of market_value)") res = stats.probplot(modified_cat_data['market_value'], plot=plt) ax2.set_title("Probability plot (of market_value)") data['market_value'].skew() # Out: 90.62 data.loc[:, 'market_value'] = np.log10(data['market_value']) data['exempt_building'].skew() # Out: 131.83 xt, _ = stats.boxcox(exempt_building + 1) stats.skew(xt) Out: 0.03 numerical_variables = data.select_dtypes(exclude = ["object"]) final_data = pd.concat([numerical_variables, dummy_variables], axis=1) y_data = final_data.market_value X_data = final_data.drop('market_value', axis=1) X_data.T.drop_duplicates().T # Remove duplicated columns remove = [] cols = X_data.columns for i in range(len(cols)-1): v = X_data[cols[i]].values for j in range(i+1,len(cols)): if np.array_equal(v, X_data[cols[j]].values): remove.append(cols[j]) X_data.drop(remove, axis=1, inplace=True) from sklearn.model_selection import train_test_split X_train, X_test, y_train, y_test = train_test_split(X_data, y_data)

25

null

2018-04-10

2018-04-10 16:58:03

2018-04-15

2018-04-15 19:52:10

3

false

en

2018-04-15

2018-04-15 19:52:10

9

1df184df307

5.682075

2

0

(II). FEATURES ENGINEERING: DATA PRE-PROCESSING

3

Philadelphia Housing Data Part-II: Features Engineering (II). FEATURES ENGINEERING: DATA PRE-PROCESSING In the previous section (Part-I), we downloaded and analysed the Philadelphia housing data and studied its features. We also saw how to deal with NaNs and performed some data pruning. In this section (Part-II), I shall perform some feature engineering to make the data comply properly with the machine learning algorithms — which we shall use in Part-III. Thus, this section deals with ‘cleaning’ the data — so lets get our ‘hands dirty’! Before we proceed with preprocessing the data, let’s take a quick look over the data again and see how the features correlate with our target variable (‘market_value’). This is an extension of Part-I. Correlations between the features and the market_value: We shall study the correlations in two different ways: The most important features relative to the market_value: Heat-map: Figure II-1 Looking at the heatmap, we see that ‘total_livable_area’, ‘taxable_land’ and ‘taxable_building’ are highly correlated with the ‘market_value’ (dark blue squares). We should expect at least these features to play valuable role in machine learning [see Part-III]. We can also obtain the correlation information between the variables other than ‘market_value’. For instance, ‘lat’ and ‘lng’ are negatively correlated (recall, lat = latitude and lng = longitude). We also find correlations that are obvious — for instance, ‘taxable_land’ and ‘taxable_building’ is correlated. We also see a small positive correlation between ‘garage_spaces’ and ‘lat’, while a small negative correlation between ‘garage_spaces’ and ‘lng’! Now let’s start with some feature engineering! (II-A). Converting categorical variables to dummy variables: First, we select the columns that are of categorical types: Certain features above are of ‘int’ type. We need to make it ‘object’ type by converting the said features to ‘string’ type. For example: Next, we extract the ‘object’ types from the data: Now, transforming the categorical variables into dummy variables: Thus, we see that the number of categorical features increased from 21 to 967! This is the main reason why training on this data, using the machine learning algorithms that we shall meet in Part-III, will be very consuming in terms of both time and memory. (II-B). Transforming skewed numerical variables: Our next step is to decrease the skewness of the skewed numerical variables. Depending on how large the skewness is, we can employ a couple of different methods. Below we highlight these methods. Below, I show some examples of such transformation. First, let’s list the numerical features: (II-B-1). ‘market_value’: Looking at the ‘distribution’ and ‘probability’ plots, Figure II-2 we see that ‘market_value’ is not normally distributed. Lets have a closer look at the moment: skewness. This tells us that ‘market_value’ has a large positive skewness of 90.62. A positive skew means that the tail tapers to the right of the peak. As this is a comparatively large skewness, we use Log base 10 transformation. Now lets again look at the distribution and probability plots again: Figure II-3 We see that ‘market_value’ is now transformed to have a normal distribution (an almost symmetrical bell-shaped curve). The skew now is 0.32 after the transformation. Thus, we have decreased the skew below the threshold value of 0.5! So we are good to go with ‘market_value’. (II-B-2). ‘exempt_building’: The skewness: This is a substantial skew, thus we need to apply a more robust, box-cox transformation. Also, ‘exempt_building’ contains some zeros, thus we have to add 1 to prevent from taking logarithm of 0 (in the case when the lambda parameter of box-cox is zero, the transformation reverts to log) We see that after box cox transformation, the skewness is reduced from 131.83 to 0.03! In the same vein, we have to perform transformations to the remaining numerical variables. (II-C). Final preparation towards using machine learning algorithms: (II-C-1). Combining numerical and dummy variables and naming this combined data as ‘final_data’: (II-C-2). Extracting the target variable (‘market_value’) from the ‘final_data’ and naming it ‘y_data’: (II-C-3). Removing the target variable (‘market_value’) from the ‘final_data’ and naming it ‘X_data’: (II-C-4). Removing the duplicate columns: It seems, some of the columns are duplicate (and XGBoost Regressor algorithm will not work if you have duplicate columns.) As there are 967 columns, it is not feasible to look at each columns to eliminate duplicates. One way to eliminate duplicate is to use pandas drop_duplicates method — which removes duplicate rows. To remove duplicate columns, we can do: However, transposing columns to rows is very memory intensive. Hence we have to find another way. After using a couple of different codes to deal with this, I found an algorithm that works in our case, which we can use as follows: (II-C-5). Finally, we split X_data and y_data into the training set and the test set using the ‘train_test_split’ method: And, we are done with feature engineering! We are ready to move to the machine learning phase in Part-III. For that, we are going to use X_train and y_train to train and test various algorithms. Then, we shall make predictions on the X_test and y_test data. Note: To avoid data snooping, we won’t touch X_test and y_test until we decide exactly which algorithm(s) and the associated parameters we are going to use.

Philadelphia Housing Data Part-II: Features Engineering

2

philadelphia-housing-data-part-ii-features-engineering-1df184df307

2018-04-16

2018-04-16 15:40:43

https://medium.com/s/story/philadelphia-housing-data-part-ii-features-engineering-1df184df307

false

1,360

null

Machine Learning

machine-learning

Machine Learning

51,320

Nikhilesh A. Vaidya

Education: PhD (Physics) | Interest: Machine Learning

61d3e8246470

GaussEuler

7

1

20,181,104

null

0

null

0

c4c117f0e0d4

2017-11-01

2017-11-01 05:48:14

2017-11-02

2017-11-02 07:18:39

1

false

en

2017-11-02

2017-11-02 08:10:40

16

1df1dc9495d

1.554717

3

0

Recently, Sofia has been trending all across the technical news. She is rather interesting: has soft skin, no hair, beautiful eyes, deep…

5

Emotional, Rational, Aware Recently, Sofia has been trending all across the technical news. She is rather interesting: has soft skin, no hair, beautiful eyes, deep voice, knows so much, but is barely able to rationalize. Sofia is a robot developed by Hanson Robotics, and she recently became a hype after receiving Saudi Citizenship. What is interesting about Sofia is a rather scary video that, unless it was scripted, highlights advancements in robotics that raise an alarm. So what can we make out of this? One simple conclusion: the technology to develop emotional, rational, and aware robots is here. And that should not be a surprise, the mathematics and science behind this has been around for a long while now: The mathematical basis for machine learning has been around since the 1950s. But now we have the right tools to actually develop Cognitive machines. There exist now out there many technologies and tools that can easily form the building blocks of such a machine. IBM Watson can help you build a Cognitive assistant on the cloud. Watson has APIs available that allow you to deploy such a robot within weeks. With a combination of Personality Insights, Tone Analyzer, and your custom Machine Learning Model, you can develop a Chat-bot that understands emotions. As for a Rational Machine, that is not difficult at all, with tools such as Natural Language Understanding, Natural Language Classifier, Watson Knowledge Studio, and Discovery, it is possible to make sense of a conversation and the context, and respond from a vast pool of knowledge. IBM is not the only player in this field: Microsoft, Google, and Amazon each have their solutions. Not only that but with the rise of Frameworks such as Tensorflow and H2O into the mainstream, and with the advent of GPU acceleration, the idea of a sophisticated AI is attainable by the very common human. The only challenge remaining for AI engineers is achieving awareness, and that remains a challenge that gives rise to many ethical dilemmas. However, as long as machines are used to support humans and improve their performance, then you should not worry about a robot taking your job.

Emotional, Rational, Aware

7

emotional-rational-aware-1df1dc9495d

2018-02-25

2018-02-25 16:36:06

https://medium.com/s/story/emotional-rational-aware-1df1dc9495d

false

359

Dubai-based coworking space and training academy on a mission to build a thriving technology ecosystem in MENA. Partnered with @GoogleForEntrep, @IBM, and @DMCCAuthority

null

astrolabsme

null

AstroLabs

academy@astrolabs.com

astrolabs

STARTUPS,STARTUP,COWORKING,LEARNING,DUBAI

astrolabsme

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Aoun Lutfi

AI Solutions Engineer, Avid Researcher and Developer — Using AI to power the world💡🤖

9c84ceda76c0

aounlutfi

13

8

20,181,104

null

0

null

0

null

2018-03-25

2018-03-25 15:54:50

2018-03-25

2018-03-25 17:47:59

3

false

en

2018-05-23

2018-05-23 09:05:11

2

1df40308d631

1.210377

1

0

This is part of the course “Probability Theory and Statistics for Programmers”.

5

Probability theory: Multivariate random variable, probability density This is part of the course “Probability Theory and Statistics for Programmers”. Probability Theory For Programmers Introduced in the previous article characteristic of the system — the distribution function — exists for random vectors of both continuous and discrete variables. But the main practical significance is the vector of continuous random variables. The distribution of the continuous random variable is usually characterized not by the distribution function, but by distribution density. For random continuous variable probability density function is the limit of the ratio of the probability of falling into the small section to the length of the section when it is tends to zero. Similarly, we can define the distribution density of a vector of two random variables. probability density By using this formula we can represent the probability of falling in the rectangle from the previous article in a new way: probability of falling into rectangle Let’s use this formula in abstract example. We have two variables with density function. What is the probability of falling into the rectangle? Next part -> Clap if you enjoy 😎

Probability theory: Multivariate random variable, probability density

42

18-systems-of-random-variables-the-probability-density-1df40308d631

2018-05-23

2018-05-23 09:05:12

https://medium.com/s/story/18-systems-of-random-variables-the-probability-density-1df40308d631

false

175

null

Probability

probability

Probability

604

Rodion Chachura

geekrodion.com

4a34eb6ffe36

geekrodion

68

2

20,181,104

null

0

null

0

null

2018-06-13

2018-06-13 20:44:17

2018-06-22

2018-06-22 19:11:16

0

false

en

2018-06-22

2018-06-22 19:11:16

0

1df53564c937

0.550943

0

Right now, AI is more like concentrated human intelligence than it is artificial intelligence. We feed it millions of human-hours of…

1

Artificial intelligence Right now, AI is more like concentrated human intelligence than it is artificial intelligence. We feed it millions of human-hours of information to train it. I think we’ll hit a new level of AI when it can learn without us. Level 1 : Classifiers. Error minimization. Level 2: Where we are now. Convolutional NN, adversarial training. Leveraging massive datasets as training Level 3: AI not based on compressed/distilled human intelligence. Level 4: here be dragons An imperfect example might be that L 2 AI may identify the species of a bird in a given image with absurd levels of success whereas L 3 AI may look at all of the images of birds available and clasify them into species on it’s own, possibly discovering a new, more powerful system of classification. I don’t have words to describe this but the AI field might

Artificial intelligence

0

artificial-intelligence-1df53564c937

2018-06-22

2018-06-22 19:11:16

https://medium.com/s/story/artificial-intelligence-1df53564c937

false

146

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Jordan Parker

Expert

7f29df4a3f63

nepthar

60

106

20,181,104

null

0

null

0

null

2018-04-08

2018-04-08 01:13:07

2018-04-08

2018-04-08 03:25:13

1

true

en

2018-04-08

2018-04-08 03:31:38

3

1df6103a4f14

4.2

3

1

0

Autonomous killer robots, an immortal AI dictator, the question of consciousness, and what it means to be human

5

Elon Musk, Artificial Intelligence and You Autonomous killer robots, an immortal AI dictator, the question of consciousness, and what it means to be human Coming soon to a street near you? Elon Musk’s documentary Do You Trust This Computer is well worth watching; it was free to stream over this weekend. I can’t say that the majority of the ideas put forward were new to me, there was however one idea, right at the end, that I must admit I had not considered before. If you have not watched the documentary, here is my recollection of the main points, this is not verbatim as I was not taking notes as I watched it. My own observations are preceded by [HJ] to distinguish my thoughts from what the documentary presented. Technological development can occur a lot faster than we imagine it might. Advances in artificial intelligence (“AI”) are coming fast, we might only be a few steps away from creating a self-aware, self-learning AI. AI has already advanced to such a stage as to render many human tasks and jobs redundant. Jobs such as drivers, data entry clerks, business analysts, even medical specialists — all these jobs (and probably most jobs eventually) are at risk of complete displacement by AI and robots. Once AI learns something, e.g. the game of GO, it is unable to be beaten by humans, because our mind is simply not as computationally powerful and does not hold as much memory. AI does not have feelings, moral inhibitions, or empathy (our human traits), it simply seeks to achieve whatever goal it has been given, or whatever goal it may eventually develop of its own volition. Companies such as Facebook, but mainly Google, are in fact AI companies. They are hoovering up masses of data every second and collecting it; its computer systems are learning based on this data. This is called “deep learning”, it is the new generation of AI. The previous model of AI was based on programming, on direct human input to tell a computer/machine what to do. The current AI methodology focuses around feeding data to the computer/machine and letting it learn for itself, in the same way that we humans happen to learn. Militaries and companies around the world are concurrently developing autonomous robots for law enforcement and military applications. Drones are already in wide proliferation around the world, carrying lethal payloads, and killing with them. At the moment, the decision to kill is made by a human, but there is reason to believe that in the future, such decisions may be made by AI systems, and these robots will be used to kill humans. [HJ] Those of you who have read Isaac Asimov, recall that this would be in direct contravention of his Three Laws of Robotics, viz. “A robot may not injure a human being or, through inaction, allow a human being to come to harm. A robot must obey orders given it by human beings except where such orders would conflict with the First Law. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.” These seem to me to be very good laws, laws that every country in the world should enact without delay. Whoever develops real AI first, will be able to rule the world, literally. Musk fears that a sentient AI, would become an immortal global dictator. He points out that it does not take any tremendous level of intelligence to create vast upheaval as history has demonstrated many times over. However, the human dictators of the past (and present ones) all die eventually. An AI dictator would rule over us for ever. [HJ] It is this last insight, about the possibility of AI becoming an immortal dictator, that is the idea in the documentary that I hadn’t ever considered, and I don’t think it is excessively alarmist. I actually think it is a fair and rational extrapolation. Musk also seems to think that for humans to remain competitive, or perhaps even relevant, we will need some form of enhancement of augmentation. [HJ] If the imperative is for humans to become metahuman (superhuman), then this could be achieved through genetic engineering and advanced medical techinques (e.g. stem cell therapy) as well as through physical and mental augmentation, creating a race of immortal (or at least very long lived) humans and human-machine hybrids, cyborgs essentially. [HJ] One of the common arguments against the possibility of true AI developing, is that humans have self-consciousness, while robots, machines, computers, do not. I think that this is a flimsy argument on which to hang our hopes. I say this because we do not actually understand what consciousness is. We are not able to say with any certainty, how and why, humans developed self-awareness. We do not actually know what consciousness really is, and it follows from this logical proposition, that consciousness could be anything. We do know however, that consciousness is a function of the brain. Again, it follows that if a brain can be replicated in a non-biological way, a non-biological consciousness could awaken. Humans and other biological life forms will probably not have a monopoly on consciousness for very much longer; this I believe, is Musk’s greatest fear. But what about the soul? What about our spiritual inner nature, what about that which means we are empathetic, that we have morals, that we are human? Is this something separate from consciousness? Might this topic of the human soul (once almost the exclusive professional domain of philosophers and mystics) now be one of the most important questions of all? It seems to me that if we do not yet understand consciousness, and we do not conclusively know if humans even have souls, that going down the path of advanced artificial intelligence is fraught with extreme risk. Further reading: My science fiction story Vulgus 2237 CE explores some of these themes further. It is about a small number of metas living amongst a vast mass of ordinary humanity, in a post-work world, run by anonymous authorities. For good non-fiction reading about consciousness, you could do worse than the New Scientist round up on consciousness to be found here. Please follow your fellow human writer Hypp Johnstone and throw a clap or two my way👏. Thank you for reading.

Elon Musk, Artificial Intelligence and You

11

elon-musk-artificial-intelligence-and-you-1df6103a4f14

2018-04-16

2018-04-16 03:17:44

https://medium.com/s/story/elon-musk-artificial-intelligence-and-you-1df6103a4f14

false

1,060

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Hypp Johnstone

My name is Hypp. Thank you for reading.

6d90bbb106c7

hypp.johnstone

74

139

20,181,104

null

0

null

0

65b16ed24b55

2018-08-06

2018-08-06 11:06:54

2018-08-06

2018-08-06 11:12:51

6

false

en

2018-09-01

2018-09-01 21:26:56

0

1df619a59ab9

1.017925

0

null

5

South Korea’s largest last mile delivery company in 2017 was VOLT. So what is their core technology ?

South Korea’s largest last mile delivery company in 2017 was VOLT.

0

south-koreas-largest-last-mile-delivery-company-in-2017-was-volt-1df619a59ab9

2018-09-01

2018-09-01 21:26:56

https://medium.com/s/story/south-koreas-largest-last-mile-delivery-company-in-2017-was-volt-1df619a59ab9

false

18

P2P Delivery Platform based on Blockchain

null

VoltTechnology1

null

Volt Technology

manager@volttech.io

volt-technology

null

Volt_Technology

Blockchain

blockchain

Blockchain

265,164

Volt Technology

P2P Delivery Platform based on Blockchain

446c3616fcf5

manager_77257

128

20

20,181,104

null

0

null

0

null

2017-11-15

2017-11-15 13:05:32

2017-11-15

2017-11-15 13:10:50

1

false

en

2017-11-15

2017-11-15 13:14:13

4

1df8bdf2c40a

0.913208

2

0

Airfio is the decentralized firm that has enclosed for advancement in crypto banking. The newly designed crypto banking platform will have…

5

Airfio — The prominent crypto banking with AI Technology Airfio is the decentralized firm that has enclosed for advancement in crypto banking. The newly designed crypto banking platform will have the more improved technology and fewer protocols. Airfio’s Blockchain is determined to provide a seamless experience of cryptocurrency transactions to each and everyone registered in the process. It has initiated the revolution by incorporating neural language and machine learning process in the crypto banking sectors. Artificial Intelligence is one of the most impressive technologies applied in different sectors respective of their applications. The machine or robots can easily analyze the intellectual knowledge as the human understand its intellectual behavior in the environment. Therefore, Artificial intelligence associated with Blockchain technology will enhance many aspects of Fintech organizations. Implementation of this technology provides better results. Overall, with AI technology users can track the movement of their transactions globally. The products are highly based with the AI technology predicted according to the consideration of the individual customers. Want to know how to join Airfio? Read here: https://goo.gl/jAJtYy You can use may referral ID to register (airfio.com/r) https://airfio.com/r?ref=james Website: www.airfio.com Image Source: Airfio

Airfio — The prominent crypto banking with AI Technology

2

airfio-the-prominent-crypto-banking-with-ai-technology-1df8bdf2c40a

2018-05-04

2018-05-04 13:13:26

https://medium.com/s/story/airfio-the-prominent-crypto-banking-with-ai-technology-1df8bdf2c40a

false

189

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

James

Me.. Passionate writer, author, blogger… Ping me to invite in crypto bounty

ee7e7dd605b9

james_37891

4

12

20,181,104

null

0

null

0

null

2018-07-10

2018-07-10 07:15:17

2018-07-10

2018-07-10 07:21:50

1

false

en

2018-08-09

2018-08-09 09:31:45

2

1dfd0aa168b8

0.803774

0

Today, AI is reforming various business practices, and transforming the way processes occur through a combination of machine learning and…

5

AI for Smarter Sales Today, AI is reforming various business practices, and transforming the way processes occur through a combination of machine learning and deep learning techniques. With its potential, Artificial Intelligence in B2B sales and marketing is here to revolutionize the way people interact with brands, information, and services through enhanced predictive and statistical analytics, personalization as well as lead generation. In recent years, the fundamental shift in sales-from being instinct-driven to insights and data-driven has let AI guide the sales journey from identification to customer retention efficiently…..click here to read Subscription Request If you want become part of forward-thinking professionals who are staying ahead in their work by planning and implementing important technological trends. Once you provide your information(Subscribe us), we will need a few hours to process your request and send FREE Subscription to qualified readers. The publisher reserves the right to limit the number of free subscriptions. To fill form click here

AI for Smarter Sales

0

ai-for-smarter-sales-1dfd0aa168b8

2018-08-09

2018-08-09 09:31:45

https://medium.com/s/story/ai-for-smarter-sales-1dfd0aa168b8

false

160

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

John Stones

Technology Expert

989533288ce3

john_stones

17

58

20,181,104

null

0

null

0

null

2018-01-14

2018-01-14 22:17:34

2018-01-14

2018-01-14 22:19:22

0

false

ja

2018-01-14

2018-01-14 22:19:22

0

1dfd86ce7008

1.782

0

車メーカーの進化がさわがしい。

5

スポーツとして、道具として車メーカーの進化がさわがしい。自動運転技術の発展がユーザーレベルにおいても目に見えるほどのレベルになり始めているところで、現実レベルとしての自動運転がいよいよこれから十数年で実用になりそうな雰囲気だ。ただ、では既存の人が運転する車がなくなるのか？というと、たぶんそうはならないだろう。特に車という道具は、「人を運ぶ道具」だけではないというところに大きく起因する。車の場合、それはスポーツの道具でもある。だからこそ、スポーツとしての車、運転すること、マシンを人が思うように扱う、操れることによって得られる爽快感などなどで得られるものは、やはり自動運転とは別次元、別軸の話だろう。であるからこそ、それはもしかしたら明確に区別したほうが、結果的に効率的であり、社会的に効果が高いような気がしている。それは自動運転のエリア、それは人が操作するエリア…と。当たり前だが、たぶん当初数十年は、人が運転する車と、自動運転する車とが混じることになる。たぶんそれが一番事故の可能性が高いのではないだろうか？また自動運転の性能がなかなか十分に発揮できない世界になりそうだと言う想像も見える。すべてが自動運転の車になり、基本、人が運転する車は「町中を」走り回らないことになれば、たぶん将来的には事故はまず起きなくなるだろう。しかしこれでは逆に、それでは「スポーツとして車を楽しみたい人」を阻害することになる。であるがゆえに、スポーツとして楽しめる場所、空間を徐々に制限していくという形での行政運営をイメージする。私がイメージする遠い未来は、自動運転の車というリソースが十分に用意されている世界においては、社会レベル全体としての移動品質の向上に他ならない。自動運転が十分に機能する世の中になれば、「車」という考え方が、大きく変革することになるだろう。ある意味移動のための道具というよりも、部屋ごと移動する意味では、もしかすると「どこでもドア」になるのかもしれない、と。となれば、車とは移動する部屋であり、家という不動産の中における動産部分という特殊な位置付けに。早く来ないかなぁ。

スポーツとして、道具として

0

スポーツとして-道具として-1dfd86ce7008

2018-01-14

2018-01-14 22:19:23

https://medium.com/s/story/スポーツとして-道具として-1dfd86ce7008

false

11

null

Sports Driving

sports-driving

Sports Driving

0

nonmult

null

ced016265b4c

nonmult

159

5

20,181,104

null

0

null

0

null

2017-10-07

2017-10-07 11:08:41

2017-10-07

2017-10-07 11:09:07

2

false

en

2017-10-07

2017-10-07 11:09:07

13

1dfe8b9025be

1.749371

1

0

There are numerous drawbacks of using dating apps and one of the biggest loopholes is breaching of privacy. It has been revealed by…

5

AI Improves the Safety Features of Dating App ‘Hily’ There are numerous drawbacks of using dating apps and one of the biggest loopholes is breaching of privacy. It has been revealed by researchers that hackers can create a strange chrome plugin to find out the exact location of a Facebook friend in Tinder. Such incidents can turn into big accidents, a case in which a malicious troll compelled a young mother to commit suicide after he set up a fake tinder account on her name. Related : Zello: How Hurricane Irma Made The ‘Walkie- Talkie’ Style App №1 On iTunes Chart However, dating app ‘Hily’ seems to be seriously concerned about verification of actual identity. It believes that the only criteria to set up a dating account is by verifying identity to prevent the cases of fake accounts, trolling and malicious activities. Hily’s founder Yan Pronin has got excellent knowledge and experience in the world of analytics, M&A modeling, dynamic pricing, and statistical modeling. Therefore, he decided to redefine the traditional practices of dating apps and bring features that are more reliable and practical. Hily uses machine learning on employing matchmaking algorithms instead of the geographical location of the user. Some apps like Tinder uses attractiveness level which allows users to keep swiping but minimizes the chances of matching as almost 80% of the people get ignored. Related : 7 Best Apps To Help You in Instant Job Search Hily makes better matches by evaluating the users with same interest levels, seeks data from the gravity of conversation and mutual likes. The quality of matches increases with the time of application usage. The app has exclusive criteria of providing risk score, which is completely affected by the users’ passing verification, complaints, conversation style, etc. This app will automatically block users having a high-risk score and prevent others from sharing personal information with him/her. In today’s strange world, Hily creates a safe environment for the users to find perfect match or make good and genuine friends. For the Latest Mobile App Technology News and Reviews, follow MobileAppDaily on Twitter, Facebook, LinkedIN, Instagram and Flipboard. Was this article helpful? Originally published at www.mobileappdaily.com.

AI Improves the Safety Features of Dating App ‘Hily’

2

ai-improves-the-safety-features-of-dating-app-hily-1dfe8b9025be

2018-02-19

2018-02-19 09:26:34

https://medium.com/s/story/ai-improves-the-safety-features-of-dating-app-hily-1dfe8b9025be

false

362

null

Dating

dating

Dating

24,079

Mobileappdaily

MAD provides information on several mobile apps such as travel apps, police apps and game apps. visit: https://www.mobileappdaily.com

a9125211f16f

mobappdaily

96

343

20,181,104

null

0

null

0

null

2018-05-04

2018-05-04 03:06:06

2018-05-07

2018-05-07 15:49:57

0

false

en

2018-05-07

2018-05-07 15:49:57

7

1dfeca255f83

1.275472

2

0

Assumption: Everyone uses Spark

5

3 Steps of Airflow Addiction Assumption: Everyone uses Spark Whats everyone’s obsession with notebooks? The terms data engineer and data scientist are thrown around by recruiters and employees alike, but their opinion on the jupyter (ipython?) notebook will certainly differ. Engineers will kinda shrug, and say “yeah, its kind of nice” meanwhile Data Scientists are likely to sacrifice their first born to ensure it stays open source. Why? Notebooks allow Spark code to explore. A Data Scientist can run a small piece of code, delete it, add another piece of code, overwrite that, decorate it with some graphs, put in a few titles, and they have a interesting, executable presentation. Even better you’re in luck, Google Datalab, Amazon EMR and Databricks all allow you to write code in notebooks freely, then put them (and their Spark code) into production with scheduled jobs. 2.My ETL for Data Science is taken care of! Ha, Nope. The model that the Data Scientist built requires a complete code set after Job X is done. And Job X depends on Job Y. So we need a task manager. In comes Airflow, a open source Python task manager, with a dashboard, worker nodes and even a few easy to use Docker containers. Even better, it has hooks for Google, Amazon and Databricks. Use Airflow to setup your dependencies, plug in your notebooks and you have a strudy, scalable, transparent ETL task manager that your Data Engineers can easily work with and Data Scientists can geek out about. 3. Sounds complicated Not gonna lie, kind of is. I’ve been doing it for the past few months, however myself and three engineers completed a multi-source ingestion big data platform using the above technologies for a large corporation and it was a strong learning experience. I would certainly advocate for using Airflow, Spark and notebooks in your ETL code. I will be at a Meetup with my company Caserta, talking about using these technologies on Google Cloud at our Meetup next Tuesday (5/8). See you there.

3 Steps of Airflow Addiction

4

3-steps-of-airflow-addiction-1dfeca255f83

2018-05-07

2018-05-07 16:27:02

https://medium.com/s/story/3-steps-of-airflow-addiction-1dfeca255f83

false

338

null

Data Science

data-science

Data Science

33,617

Maxwell Goldbas

NYC. Director at Caserta Innovation Labs. Blockchain, AWS, Google Cloud. Virginia Tech. News Junkie. Thorium and Renewable Energy Advocate.

2d9ecd39b8f0

GoldbasNews

164

230

20,181,104

null

0

null

0

bfbc691b9617

2018-04-10

2018-04-10 15:38:42

2018-05-16

2018-05-16 13:07:28

11

false

en

2018-05-17

2018-05-17 11:36:12

4

1dfefc818b95

5.028302

15

2

0

When personalization fails us, we find our way

5

How humans find good stuff on Netflix When personalization fails us, we find our way Designing an engine for personalized content recommendations is hard. Really hard. So hard, in fact, that searching for something to watch on Netflix has long been a trope of the internets. In a perfect world, the journey from “wanna watch” to “watching” would look something like this: That’d be nice. But reality looks more like this: “Show me less like this” Recommendation engines (henceforth reco-engine) are big business. They contribute so much to user experience, it’s funny how UX designers don’t really know much about them. In many ways, their performance is as critical to UX as having internet. As the old saying goes, Nail the algorithm, conquer the world -Craig Phillips Calling Netflix an extremely successful company is an understatement. As of January 2018, they’re worth over $130 billion. With it’s continuous growth, it’ll surely soon be worth more than Disney ($155b) or Comcast ($169b). So they must be doing something right. The three legged stool approach to their reco-engine seems to cover all the bases, or at least the bases they should reasonably cover. “The three legs of this stool would be Netflix members; taggers who understand everything about the content; and our machine learning algorithms that take all of the data and put things together” —Todd Yellin, VP of Product Innovation at Netflix The endless search for good stuff doesn’t seem to be stopping anyone from using the product. And perhaps the current state is just the balance we need. Perfect personalization requires perfectly personal data. But how does protecting personal data and privacy impact producing a perfectly personalized presentation of programming? If Netflix presented precisely what we pine for, it’d probably perturb people. Listen to this 99% Invisible episode to learn about desire paths As a species we’ve created work arounds—desire paths—to more efficiently and less painfully get from “wanna watch” to “happily watching”. So I did what any overly-intrusive designer would do, and talked to a bunch of people about how they find good content on Netflix. And came up with a four archetypes to explore the role of our humanness in finding what to watch. Turns out the reco-engine plays only a tiny part. Miss Moody The feels determine the reels This archetype is the emotional type. Her preferences are driven by her current mood, the mood she wants, or the collective mood of the small group she watches TV with, whether human or canine. This mood analysis brings her to a genre. Her internal dialogue usually goes like this: “Been a long week, I’m tired, don’t want to think, need a comedy, a stupid one, not so stupid it’s bad, but just the right amount of stupidity to be cheesy yet engaging.” The result is usually not good. Lots of scrolling, cross referencing with google searches, and in the end, settling on something that doesn’t look terrible. Here’s how Miss Moody finds something to watch. Bromance The unwritten social dynamics of bros in the wild This fella loves a night out with the boys. And who doesn’t? Because in the end, who knows you better than your bros? 😅 An important thing that came out of this is how the dynamics of a gathering of friends facilitates decision making. Usually, more cooks in the kitchen is a recipe for disaster. But when there is an unspoken trust between the cooks, maybe things change. The bros, in this case, understand each other and what they want. They have varied tastes, but they can decide what to watch easily because they trust their friends, and are willing to compromise. Each individual’s definition of the perfect movie is put aside for the greater good. Here’s a simplified flow of how the bromance plays out. Fam reacts only I trust, therefore I watch This goes a level deeper than the bromance. This archetype is about that special person whom you trust with all entertainment recommendations. In this day and age, it’s a lot to give someone that kind of power over you. One day over coffee they say “hey have you seen Lost?” Next thing you know you’re starting season 6, questioning all your choices in life. Committing to a new series has become a major life choice, requiring the sacrifice of health and wealth. But you trust them because they know you intimately, and you know them. There’s transparency in your relationship. A transparency that you don’t get with reco-engines. I think Drake said it best, You know what I like, Oh yes, oh yeah, Oh yes, oh yeah, Oh yes, oh yeah You know what I’m sipping, I’ll teach you how to mix it But you’re the only one, ’cause I don’t trust these b****** I don’t, I don’t trust these b****** Here’s what the flow looks like. Rainy Day Romp Weather gotcha down Some people, more than others, are severely impacted by the weather. A windy, sunny, or snowy day can significantly impact how they spend their time. Rain is her kryptonite. Nothing does it like a rainy day. It has that magical power of giving you permission to do nothing and procrastinate on all your responsibilities. Time freezes. For this archetype, rain also induces a mood. It’s hard for her to put it into words, and it’s more of a feeling of what’s right. After all, the heart wants what it wants. Interestingly, this archetype was the most open to browsing Netflix, and exploring new options based on the reco-engine. Maybe because they are in a relaxed state of mind, without a sense of urgency, and all the cares of life—like the rain outside—have fallen to the ground. Here’s what it looks like. Thanks for reading! Please let us know your thoughts in the comments below. Your claps are always appreciated. Make sure to follow us at The Fourth Wall for the latest from our team, bringing you our thinking on interactive digital media and products.

How humans find good stuff on Netflix

316

how-humans-find-good-stuff-on-netflix-1dfefc818b95

2018-05-21

2018-05-21 17:02:00

https://medium.com/s/story/how-humans-find-good-stuff-on-netflix-1dfefc818b95

false

988

The design of interactive digital media

null

The Fourth Wall

design@axonista.com

the-fourth-wall

DESIGN,TECHNOLOGY,TELEVISION,USER EXPERIENCE,MEDIA

null

User Experience

user-experience

User Experience

25,182

Craig Phillips

Product designer and writer. Enthusiastic about people.

470bdb36ec88

craig5446

1,162

305

20,181,104

null

0

null

0

null

2018-02-02

2018-02-02 18:05:09

2018-02-02

2018-02-02 18:05:32

0

false

en

2018-02-02

2018-02-02 18:05:32

0

1dff1badb610

3.4

0

The debate on the profession of software developer — its alleged ills and failings — abates at times, but never ends. Here are two sets of…

4

The Algorithm Debate The debate on the profession of software developer — its alleged ills and failings — abates at times, but never ends. Here are two sets of ten commandments. The first set is directed at professional software developers, and the second to non-practitioners. Ten commandments for professional developers: 1. Algorithm is a collection of models; cherish their diversity. 2. It’s a model, not the model. 3. Make your model simple enough to isolate specific causes and how they work, but not so simple that it leaves out key interactions among causes. 4. Unrealistic assumptions are OK; unrealistic critical assumptions are not OK. 5. The world is (almost) always second-best. 6. To map a model to the real world you need explicit empirical diagnostics, which is more craft than science. 7. Do not confuse agreement among developers for certainty about how the world works. 8. It’s OK to say “I don’t know” when asked about the technology or policy. 9. Efficiency is not everything. 10. Substituting your values for the public’s is an abuse of your expertise. Ten commandments for non-technologists: 1. Algorithm is a collection of models with no predetermined conclusions; reject any arguments otherwise. 2. Do not criticize a developer’s model because of its assumptions; ask how the results would change if certain problematic assumptions were more realistic. 3. Analysis requires simplicity; beware of incoherence that passes itself off as complexity. 4. Do not let algorithms scare you; developers use algorithms not because they are smart, but because they are not smart enough. 5. When a developer makes a recommendation, ask what makes him/her sure the underlying model applies to the case at hand. 6. When a developer uses the term “” ask what s/he means by it. 7. Beware that a developer may speak differently in public than in the seminar room. 8. Developers don’t (all) worship technologies, but they know better how they function than you do. 9. If you think all developer think alike, attend one of their workshops. 10. If you think developers are especially rude to non-practitioners, attend one of their workshops. I have spent enough time around non-practitioners to know that their criticism often misses the mark. In particular, many non-practitioners tend not to understand the value of algorithms or parsimonious modeling (especially of the mathematical kind). Their typical riposte is: “but it is more complicated than that.” It is of course. But without abstraction from detail, there cannot be any useful analysis. Developers, on the other hand, are very good at algorithms or design models but not so good at navigating among their models. In particular, they often confuse a model, for the model. A big part of the problem is that the implicit scientific method to which they subscribe is one in which they are constantly striving to achieve the “best” model.TDevelopers are particularly bad at this, which accounts in part for their dismal performance. There is too much of “is this the right model” (and their variants), and too little of “how do we know whether it is this model that is the most relevant and applicable at this point in time in this particular context.” On the other hand, developers should also bear in mind that technology is a resource that we can use for God’s glory. Here’s three ways technology can contribute to the greater cause of Islam. 1. Technology Enables Communication Through Facebook and Twitter, religious scholars can easily communicate directly with their followers throughout the day and week. Technology allows a greater sense of community that doesn’t demand proximity. Technology enables their followers easily to have direct communication with them on a broader and a clearer scale. Ongoing communication through technology helps the cause of Islam. 2. Technology Enables Community Technology allows a greater sense of community that doesn’t demand proximity. Proximity isn’t required for community. Social media is where younger generations are interacting. It’s the new marketplace. It may be unnatural for past generations, but it is how community for younger people is now started and developed. Use technology to enable communication, community and brotherhood in Islam. Through social media, a new attendee can connect to other mosque or community members before he or she ever has a chance to meet at a mosque gathering or a small group. Of course, true community requires feet and faces and not just electrons and avatars. But those electrons and avatars can be tools to bring people into closer community with feet and faces. This is a big shift in how we interact, but we have to use it if we want to enable community for the sake of Islam. 3. Technology Enables Brotherhood Some mosques have an app where people can actually access the prayer outline, and people use their phones or iPads to follow along and take notes. Technology enables members and attendees to enhance their worship experience at mosque. All of these are tools to enhance brotherhood. Technology, though, is not the goal. The goal is to enable the mosque’s mission to make followers of all people groups. Find the Benefits of Technology There are unintended side effects of technology that are both de-personalizing and dehumanizing. But there are some wonderful benefits of technology that enable the mission of the mosque. Technologists should also consider how to use technology to enable communication, community and brotherhood in Islam.

The Algorithm Debate

0

the-algorithm-debate-1dff1badb610

2018-02-02

2018-02-02 18:05:33

https://medium.com/s/story/the-algorithm-debate-1dff1badb610

false

901

null

Machine Learning

machine-learning

Machine Learning

51,320

Daily Wisdom

null

ddd120ae7c2

dailywisdom

60

0

20,181,104

null

0

null

0

null

2018-01-17

2018-01-17 03:56:47

2018-01-23

2018-01-23 23:28:38

1

false

en

2018-01-24

2018-01-24 00:13:18

0

1dff3b7dd7d6

1.381132

9

0

Warning. Prepare to take control in 3, 2, 1…

5

Isaac Warning. Prepare to take control in 3, 2, 1… Isaac awakened with a start. The car was precariously hurtling towards an unmarked construction zone. Recalling his training, he engaged the brakes and swerved sharply to the right, throwing the car into a spin and slamming it into a street lamp. The in-vehicle camera revealed that the passengers, a woman and her newborn, did not survive the impact. Isaac had failed. After self-driving technology reached Level 4 autonomy in 2020, self-driving cars have become ubiquitous. Why would anyone drive manually like their parents did, when an autonomous vehicle could take you anywhere at high speeds with 99.99% safety? Four 9's was good, but not good enough for the Department of Transportation, which passed new regulations in 2021 that mandated that every self-driving car have a remote backup driver constantly monitoring it. In emergency scenarios where the self-driving computer didn’t know what to do, the remote human driver could take control of the car and resolve the situation safely. With so many self-driving cars on the road, the demand for remote drivers was overwhelming and eventually led to the creation of hundreds of startups that trained and contracted backup drivers. The turnover rate for the job was understandably high, as each emergency scenario was tougher than the last, and most people broke down after their first failure and quit altogether. So after the supply of emotionally-stable individuals dried up in the first month, the once-lucrative business was no longer sustainable, and all of these startups closed down. All except one — which, to the bewilderment of labor economists around the world, have somehow managed to stay in business, completing a total of 10,000 emergency scenarios as of this morning. Logging failure #9973. Calculating loss. Backpropagation complete. Loading updated model into memory. Isaac awakened with a start.

Isaac

111

isaac-1dff3b7dd7d6

2018-01-27

2018-01-27 21:05:01

https://medium.com/s/story/isaac-1dff3b7dd7d6

false

313

null

Self Driving Cars

self-driving-cars

Self Driving Cars

13,349

Raymond Xu

software @lyft. previously @columbia, @adicu, @google. http://raymondxu.io

4e326fedf0c7

raymondxu

268

164

20,181,104

null

0

null

0

null

2018-09-03

2018-09-03 07:42:58

2018-09-03

2018-09-03 07:45:48

0

false

en

2018-09-03

2018-09-03 07:45:48

12

1dff6cde7a67

0.324528

0

Learn Colors for Children with Little Monsters Fun Play Color Chicken Legs Machine 3D Kids Cartoons. #learncolorsforchildren…

1

Learn Colors for Children with Little Monsters Fun Play Color Chicken Legs Machine 3D Kids Cartoons Learn Colors for Children with Little Monsters Fun Play Color Chicken Legs Machine 3D Kids Cartoons. #learncolorsforchildren #monstersvideosforkids #cartoonvideosforkids #kidslearningvideos learn colors with chicken legs color change making machine 3d cartoon videos for kids. Thanks for watching Please Like! & Subscribe For more Updates and Videos Subscribe Here By Following Link : https://goo.gl/Jdhs1B Follow On Other Social Sites… Blogger: https://goo.gl/drgtjs Pinterest: https://goo.gl/AmJcp3 Instagram: https://goo.gl/PoQhq1 Twitter: https://goo.gl/naJti5 Fb: https://goo.gl/NTLQq6 G+: https://goo.gl/zZ2Z68

Learn Colors for Children with Little Monsters Fun Play Color Chicken Legs Machine 3D Kids Cartoons

0

learn-colors-for-children-with-little-monsters-fun-play-color-chicken-legs-machine-3d-kids-cartoons-1dff6cde7a67

2018-09-03

2018-09-03 07:45:48

https://medium.com/s/story/learn-colors-for-children-with-little-monsters-fun-play-color-chicken-legs-machine-3d-kids-cartoons-1dff6cde7a67

false

86

null

Machine Learning

machine-learning

Machine Learning

51,320

Abeel Klugel

null

b0c10d88bc8f

abeelklugebex

1

2

20,181,104

null

0

null

0

49523d4bd9da

2018-03-13

2018-03-13 08:35:09

2018-03-13

2018-03-13 10:40:15

1

false

en

2018-03-19

2018-03-19 15:09:39

0

1dff87956a2

1.377358

0

“Who said machines learn from humans , humans can also learn from machines ? All our daily interactions with the machines can be harnessed…

1

LEARNING CAN BE IRRITATING ! “Who said machines learn from humans , humans can also learn from machines ? All our daily interactions with the machines can be harnessed into learning”That was the idea of the CEO of HOWDUINO jaknap. When some unexpected glitch in the system was over looked. The system was supposed to have emotional intelligence and smart to analyze the mood of the user and change the difficulty of learning accordingly . But due to a null error in the programming the AI started overlooking the emotional condition of user . A few months ago a smart wall-e kiosk was smashed into pieces , and the number of such incidents have risen exponentially .The company tried its best to blame it on some rebels who didn’t want the change , later on it was confirmed that regular users were frustrated by the bad timings of the AI teaching them things . One of the user was open enough to come in the light to share his experience . According to him he lost his wallet and using the information kiosk he was quickly sending an alert to everyone , without his wallet he cant find his way back , being a new comer in the town , language was an issue but the kiosk not only did change the language from English to Gujarati but also increased the difficulty level in the whole process of sending the alert as a result of which the user was frustrated and banged the kiosk , “it was just a moment of frustration he says but it was frustrating , the dumb machine should know when it is a case of emergency” The CEO has promised an update which will rule out the problem , but this makes us again should we really make AI an integral part of our life .

LEARNING CAN BE IRRITATING !

0

learning-can-be-irritating-1dff87956a2

2018-03-19

2018-03-19 15:09:40

https://medium.com/s/story/learning-can-be-irritating-1dff87956a2

false

312

A collaborative workshop at NID-Ahmedabad exploring narratives for a digital future. Watch as we learn.

null

Howdu.ino

null

howdu-ino

IOT,ARDUINO,DIGITAL,FUTURE,NARRATIVE

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Pankaj Yadav

null

f0e09eff118f

pankaj_04

3

2

20,181,104

null

0

null

0

null

2018-09-16

2018-09-16 20:27:47

2018-09-16

2018-09-16 20:35:58

4

false

en

2018-09-22

2018-09-22 13:50:13

5

1dffaa4ea4c8

1.635849

0

The Sensefolio’s algorithm takes into account companies’ reports as well. It skims through each of their ‘Sustainability’ reports…

5

Sensefolio Score Methodology: 2nd Part- SUSTAINABILITY AND ESG REPORTS SCORE The Sensefolio’s algorithm takes into account companies’ reports as well. It skims through each of their ‘Sustainability’ reports, ‘Environmental, Social and Governance (ESG)’ reports, or simply their ‘Annual’ reports, and applies an internal Natural Language Processing (NLP) algorithm to assess the companies’ degrees of involvement into social, environmental and governance topics defined by Sensefolio. Below are a few examples of reports that companies can release to disclose their commitments to ESG. They can be either ‘ESG’ reports, ‘Sustainability’ reports, or the information can be contained in the companies’ ‘annual reports’ as well. Barclays’s Environmental, Social and Governance 2017 Report Available on: https://www.home.barclays/content/dam/barclayspublic/docs/InvestorRelations/AnnualReports/AR2017/Barclays%20PLC%20ESG%20Report%202017.pdf Royal Dutch Shell Sustainability Report 2017 Available on: https://reports.shell.com/sustainability-report/2017/servicepages/downloads/files/download.php?file=shell_sustainability_report_2017.pdf LVMH Corporate Social Responsibility 2017 Report Available on: https://r.lvmh-static.com/uploads/2014/11/2017-social-responsibility-report.pdf Sensefolio is programmed to intercept and read every new report being published by a company. They can be named ESG report (as seen on A), Sustainability report (as seen on B), or CSR report (as seen on C). Text is then read and analysed to be integrated and add new information into Sensefolio’s existing or new companies’ scores. As explained earlier, internal NLP algorithms are used to assess the degree of companies’ commitment to ESG topics as well as the objectivity and subjectivity of the text corpus. For More Information about Sensefolio: Visit our website: www.sensefolio.com Contact us on: contact@sensefolio.com Join us on Twitter and don’t miss the latest news on ESG: https://twitter.com/sensefolio

Sensefolio Score Methodology: 2nd Part- SUSTAINABILITY AND ESG REPORTS SCORE

0

sensefolio-score-methodology-2nd-part-sustainability-and-esg-reports-score-1dffaa4ea4c8

2018-09-22

2018-09-22 13:50:13

https://medium.com/s/story/sensefolio-score-methodology-2nd-part-sustainability-and-esg-reports-score-1dffaa4ea4c8

false

248

null

Sustainability

sustainability

Sustainability

19,446

Sense Folio

Sensefolio scans financial news, companies’ financial reports, and social media data to discern and observe signals on how involved companies are towards ESG.

d20ba681253b

sensefolio

3

1

20,181,104

null

0

null

0

null

2018-06-19

2018-06-19 16:43:20

2018-06-19

2018-06-19 16:44:21

1

false

en

2018-06-19

2018-06-19 16:44:21

0

1e024a469587

1.2

0

Creating slow-motion footage is all about capturing a large number of frames per second. If you don’t record enough, then as soon as you…

5

Nvidia creates a slow-mo video by creating extra frames using deep learning Creating slow-motion footage is all about capturing a large number of frames per second. If you don’t record enough, then as soon as you slow down your video it becomes choppy and unwatchable. Unless that is, you use artificial intelligence to imagine the extra frames. Slow-Motion consists of capturing more number of frames per second than usual, if the frames aren’t enough, the output video is glitchy. Nvidia has used deep learning to turn 30 frames-per-second videos into 240 frames-per-second slow-motion. The AI system developed by Nvidia looks at two different frames and then creates intermediary footage by tracking the movement of objects from one frame to the next. The results are close to perfect. Work is still under progress to make the output as perfect as it can get before it can be used commercially. The deep learning technique used by Nvidia not only works faster but is also easier than the existing tools for the job. “While it is possible to take 240-frame-per-second videos with a cell phone, recording everything at high frame rates is impractical, as it requires large memories and is power-intensive for mobile devices,” write the researchers in a paper describing their work on pre-print server arXiv. “For these reasons and others, it is of great interest to generate high-quality slow-motion video from existing videos.” Nvidia’s system can produce up to seven intermediary frames — more than enough to create decent slow-motion.

Nvidia creates a slow-mo video by creating extra frames using deep learning

0

nvidia-creates-a-slow-mo-video-by-creating-extra-frames-using-deep-learning-1e024a469587

2018-06-19

2018-06-19 16:44:22

https://medium.com/s/story/nvidia-creates-a-slow-mo-video-by-creating-extra-frames-using-deep-learning-1e024a469587

false

265

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Garima Bhaskar

null

2c5e7c975125

bhaskar.gb26

6

1

20,181,104

null

0

null

0

f702855ffe47

2017-10-22

2017-10-22 10:00:21

2017-10-22

2017-10-22 10:00:22

7

false

en

2017-10-22

2017-10-22 10:00:22

9

1e024cbc255a

1.910377

0

null

3

AI is Not a Threat to Visual Communication # medium.com Craig & Karl identity for Berlin trend show I originally posted this article 10/14/2016 For many people in t… The Honey Dynamite Immersive Experience Concept’s # medium.com Honey Dynamite APPS Tap — July 14, 2016 This is a ball used for a game that is a cross between pinball and t… Music composed by IBM’s Watson Beat sparks debate on music copyright # medium.com If you’re not impressed enough by IBM Watson’s ability to create movie trailers, diagnosing leukemia, and ev… American Dreamtime: A Scrambled Memoir Of Poetic Future History # medium.com It seems to have spontaneously combusted, but it didn’t. The disease struck long ago, brewed in the petri di… A Guide To NLP : A Confluence Of AI And Linguistics # codeburst.io We are living our lives more in bits & bytes than in exchanging emotions. We transact and communicate more o… Artificial Intelligence and Control # medium.com The artifacts that we have crafted are a reflection of our own values. When Hume said: “Reason is, and ought… Autonomous Vehicles Company of the day (Week 32) # medium.com Autonomous #Vehicles Company of the day: sfara @sfara_co https://t.co/bbucAaN0zk #smartcity #mobility #selfd… 6 new things to read in AI # medium.com Complex Event Processing (CEP) using Apache Flink # medium.com Photo by Jonathan Petersson on Unsplash What …

8 new things to read in AI

0

8-new-things-to-read-in-ai-1e024cbc255a

2018-05-09

2018-05-09 01:22:44

https://medium.com/s/story/8-new-things-to-read-in-ai-1e024cbc255a

false

228

AI Developments around and worlds

null

AI Hawk

aihawk1089@gmail.com

ai-hawk

DEEP LEARNING,ARTIFICIAL INTELLIGENCE,MACHINE LEARNING

null

Deep Learning

deep-learning

Deep Learning

12,189

AI Hawk

null

a9a7e4d2b403

aihawk1089

15

6

20,181,104

null

0

import pandas as pd import folium #zip file from the census site file_url = 'http://www2.census.gov/geo/docs/maps-data/data/gazetteer/2016_Gazetteer/2016_Gaz_zcta_national.zip' #Pandas usually infers zips as numerics, but we lose our leading zeroes so let's go with the object dtype df = pd.read_csv(file_url, sep='\t', dtype={'GEOID' : object}) #some column names have some extra padding df.columns = df.columns.str.strip() df.head() #grab a random sample from df subset_of_df = df.sample(n=500) some_map = folium.Map(location=[subset_of_df[‘INTPTLAT’].mean(), subset_of_df[‘INTPTLONG’].mean()], zoom_start=4) #creating a Marker for each point in df_sample. Each point will get a popup with their zip for row in subset_of_df.itertuples(): some_map.add_child(folium.Marker(location=[row.INTPTL AT,row.INTPTLONG], popup=row.GEOID)) some_map some_map = folium.Map(location=[subset_of_df[‘INTPTLAT’].mean(), subset_of_df[‘INTPTLONG’].mean()], zoom_start=4) mc = MarkerCluster() #creating a Marker for each point in df_sample. Each point will get a popup with their zip for row in subset_of_df.itertuples(): mc.add_child(folium.Marker(location=[row.INTPTLAT, row.INTPTLONG], popup=row.GEOID)) some_map.add_child(mc) some_map #grab a random sample from df subset_of_df = df.sample(n=5000) some_map = folium.Map(location=[subset_of_df[‘INTPTLAT’].mean(), subset_of_df[‘INTPTLONG’].mean()], zoom_start=4) callback = (‘function (row) {‘ ‘var circle = L.circle(new L.LatLng(row[0], row[1],' '{color: “red”, radius: 20000});’ ‘return circle};’) some_map.add_child(FastMarkerCluster(subset_of_df[[‘INTPTLAT’, ‘INTPTLONG’]].values.tolist(), callback=callback)) some_map

15

null

2018-02-03

2018-02-03 03:20:51

2018-02-03

2018-02-03 21:42:56

7

false

en

2018-02-03

2018-02-03 21:44:13

4

1e03b01cb7b1

2.412264

9

2

0

Folium MarkerClusters are an easy way to declutter your map and add some flash and FastMarkerClusters are essential for larger datasets…

4

Folium MarkerClusters and FastMarkerClusters Folium MarkerClusters are an easy way to declutter your map and add some flash and FastMarkerClusters are essential for larger datasets. Link to the code is at the bottom. Generating some regular markers Looking pretty cluttered with just 500 markers. Let’s look at those same set of points using MarkerClusters. and when we zoom in As you can see, MarkerClusters are a slick way to group your Markers at varying zoom levels. And you can do popups or change your icons or whatever you normally do with Folium Markers. And for a bonus, Folium has something called FastMarkerClusters. FastMarkerClusters are pretty handy when you have a large number of points to plot as Folium can be pretty slow or downright unresponsive. For comparison, the FastMarkerCluster plot below took 140 milliseconds whereas the previous plots took 14 seconds for the same 500 points. I also plotted 5,000 points (320 ms) and all 33,144 ZipCode centroids (670 ms) Currently, you give up some functionality with FastMarkerClusters. Some of this can be overcome with some javascript callbacks and some can’t. Check out the supporting Jupyter notebook/code here Want more Folium? Check out the repo and read my other posts.

Folium MarkerClusters and FastMarkerClusters

43

folium-markerclusters-and-fastmarkerclusters-1e03b01cb7b1

2018-05-05

2018-05-05 10:01:12

https://medium.com/s/story/folium-markerclusters-and-fastmarkerclusters-1e03b01cb7b1

false

361

null

Data Science

data-science

Data Science

33,617

Bob Haffner

Co-Founder of Inventive Data Solutions. Helping companies make better decisions with their data. This a place for my slightly longer tweets

1adac13c289f

bobhaffner

73

43

20,181,104

null

0

null

0

null

2018-01-27

2018-01-27 15:10:39

2018-01-29

2018-01-29 15:31:01

1

false

en

2018-02-06

2018-02-06 06:57:25

0

1e09107d8c06

2.181132

4

0

With each passing second, human beings are are ever increasingly generating Data. And with each passing day, Data is becoming more and more…

5

Following the data trail… With each passing second, human beings are are ever increasingly generating Data. And with each passing day, Data is becoming more and more prominent in fields as diverse as online shopping, healthcare, policy-making and governance. Simply put, Data is nothing but a raw collection of facts and figures. When Data starts making sense and is put in a valid context it becomes Information (which has meaning & value of its own ). This article presents a case of why Data is important now more than ever. The ginormous volume, the robust variety and the sheer velocity! (All of which are associated with the widespread generation of Data) The evolution in mobile technologies in terms of cheaper costs and better access to the Internet has pioneered a mammoth shift and change of attitudes towards reliance on digital tech for everyday works, be it buying groceries from bed or following it up, with a selfie having a(hash tag) #BroccoliBrunch. Everything generates Data. Where do we go from here..? As the big bangs in the IoT led tech sector and mobile & wearable devices continue to occur at accelerated speeds, the amount of data captured and consequently data reported has grown exponentially over the past few years. As connectivity becomes more potent and robust, the intelligence and experiences that become possible from said connectivity will and should provide the next flurry of innovation and discoveries in computing for the coming years. How that scenario may look? Imagine a wearable fitness band not only measuring how far you ran, but gives you real-time feedback whilst you’re running. If you are training for a marathon, this could be instrumental to over or under training. Sensors collecting data on diabetics and reporting that information back to one’s doctor in real-time so the doctor can make a more informed decision for their next visit . Or imagine the ability for a school system to analyze data on their students and push notify parents that an unusual spike in influenza has affected a significant percentage of students so it would be wise not to send Javed or Sara to school today. No one wants their children to be unnecessarily sick, now do they? Imagine the joy of travelling and having your apps be even more than price prediction portals, prompting you with hotels, restaurants (catered to your dietary restrictions or preferences) and even entertainment options. Or rebooking your meeting at noon because traffic is bad on the 5th Street. Or even better, automatically reroute the driver in your cab because data from a provider like Google Realtime (traffic) has determined the most efficient route so you can in fact make your meeting. One thing that is common between all the above innovations and scenarios is that they are ALL driven by data. The next 8-10 years of innovation & discovery will be data-driven. People who seem to know; tech leaders, entrepreneurs and investors have placed their respective bets on what I poise as the world’s most valuable commodity: Data. parting note: Where there is a will, you’re only missing some Data,to find a way.

Following the data trail…

4

following-the-data-trail-1e09107d8c06

2018-06-03

2018-06-03 02:03:13

https://medium.com/s/story/following-the-data-trail-1e09107d8c06

false

525

null

Data

data

Data

20,245

Moonis Ali

a seldom Griot with zest for Data Science and remnants of a utopian peace-monger. @moonisali

7688db82428c

moonisali

18

19

20,181,104

null

0

null

0

null

2018-07-23

2018-07-23 17:42:31

2018-07-18

2018-07-18 00:00:00

6

false

en

2018-07-23

2018-07-23 18:15:04

6

1e0ba5ad4bf9

8.806604

5

0

After seeing the Systems Design Engineering 2017 Class Profile, I was deeply inspired to create a class profile for my Software Engineering…

3

My Favourite Project during University After seeing the Systems Design Engineering 2017 Class Profile, I was deeply inspired to create a class profile for my Software Engineering class. I owe a lot to my class and my program, and felt incredibly lucky to be a part of it. Prior to university, I knew very little about software engineering, so studying it at the University of Waterloo never occurred as an option. If my friend never suggested to me to study at the University of Waterloo, I wouldn’t be where I am today. I wanted to create a class profile capturing what I wished I knew before university. I wish I knew what studying software engineering was like, what opportunities were available, how much debt I’d graduate with, if it’s too late for me to start coding, etc. There’s a lot of focus from the media on what software engineers in the industry are like, but the path connecting their beginnings to employment is still mysterious. The goal of our class profile is to share the life of software engineering students, covering topics such as co-op, academics, lifestyle, and much more, and to increase the transparency of the transition from a high school student to a full-time software engineer. You can check out the final result here. UW Software Engineering 2018 Class Profile The UW Software Engineering 2018 Class Profile is a report that showcases who SE students are, what they've done, and…classprofile.andyzhang.net Understanding the questions I gathered what people wanted to know about Software Engineering (SE) students through a survey. The goal of the survey is to gather questions from the target audience. In this project, the target audience is composed of students who are either interested in or new to software engineering. I created a Google Form to collect questions current SE students wish they knew during high school, or what they wanted to learn more about. SE students were asked because they were students who were once interested in SE and are possibly still new to SE (depending on the year). In total, I collected 152 questions from 40 survey responses across 5 class of SE students. I went through all of the questions and collapsed any questions that were redundant. I then classified the question type: One dimensional: Can be answered by one data point. Ex: How many hackathons did students attend? Two dimensional: Can be answered by two data points. Ex: Do grades affect salary? Afterwards, I removed questions that were too specific or could not be included in the survey (dating life, sex, etc.). The latter questions were separated into a second anonymous class survey and is not associated with this class profile. Designing the survey One of the main challenges was trying to get a high response rate. Since the survey involved sharing personal data, students were more likely to be reluctant to respond. The goal was to get an 80% response rate, so I made several compromises to the survey: It should take 15 minutes to respond, meaning we had to cut questions. Its responses should not be seen by any students including myself. The technicalities are discussed in the section below. Once the questions were fully sorted, the questions were broken down into individual survey questions. Some questions were also shortened for brevity. For example, hackathon attendance per term was shortened to hackathon attendance throughout university. I did a beta test on a small group of students to gather feedback on the survey. The feedback consisted of catching edge cases (ex: there’s no numerical grade during an exchange term, some people skipped a co-op). After addressing the feedback, the class survey was ready. Surveying the SE class of 2018 To further incentivize the class to fill out the survey, I talked to several professors to see how they’ve incentivized students in the past to fill out faculty surveys. Luckily, Derek Rayside, our FYDP professor, was supportive of the class profile and offered to add marks for everyone if the survey reached an acceptable response rate. In addition, Patrick Lam, the director of Software Engineering, agreed to teach our weekly class wearing a silly outfit if we could reach an 80% response rate. The survey closed with a 79.6% response rate. Working with anonymous data Processing the data became fairly tricky because the survey data was on Patrick Lam’s computer. This was an important constraint because many students expressed concerns on their personal data and didn’t want it to be seen by any students including myself. As a result, Christopher Luc and I came up with a system to process the data anonymously. Before programmatically processing the data, the data needed to be normalized. For example, Snap, Snapchat, Snap Inc becomes Snap. To view a column without being able to infer which row it belonged to, the system shuffled each column independently from each other. Once each column was shuffled, it created a local file storing the shuffled column as well as the original row indices. These indices were a key piece to rebuilding the original data. Patrick only sent me the shuffled columns and held on to the original indices. I manually normalized the shuffled data, fixed any typos and ensured that answers were properly formatted so that they could be processed programmatically. Once I finished normalizing the whole dataset, I sent back the shuffled columns so that Patrick could stitch each normalized shuffled columns into a new dataset. Transforming the data There were two types of findings that we were looking for: one dimensional (how many X in the class?), and two dimensional (does X affect Y)? I already had the data I needed for the one dimensional findings, but needed to group data for the two dimensional findings. I wrote several scripts to transform the raw data into a structured format that didn’t reveal anything about a student. This was done by minimizing the coupling between each column of a row. For example, if I wanted to see the effect of grades on salary, I grouped each grade with the corresponding salary into a pair. Each of these pairs were disconnected from the pairs from other terms. In this example, disconnecting data from each term prevented me from identifying a recognizable salary (i.e. outliers), and then reading the row. Since the terms were disconnected, I could not see the data from the person’s other terms. These scripts were shared with Patrick Lam so that he could run them on the normalized data sets. He then sent me back the structured datasets for me to process. Learning from the data After I finished transforming the data, I visualized each data set to determine whether there was something interesting in the data. Since the structure of each data set was similar, I was able to reuse my visualizations easily, enabling me to figure out what had potential. Although visualizations didn’t help me figure out if something was statistically significant, it did help me filter out what was not statistically significant. In this example, it was easy to tell that there was a low chance of there being a correlation between parental education and salary per term. In this example, I was plotting hackathon attendance vs salary. I noticed that there was a difference between students who didn’t attend hackathons and students who did. This led me to running more rigorous statistical tests to see if there was a correlation. Statistical verification After visualizing all of my transformed data sets, I narrowed my focus down to the data sets with potential. I consulted with Christopher Luc, a two time Data Science intern (meaning he’s legit!), what methods he would recommend for the data I had. There were two methods I used: one way ANOVA, and evaluating correlation coefficients. I ran the one way ANOVA on data sets involving groups of distributions. For example, gender vs salary had two groups. Each entry from a student was their average salary. The aggregation of each average salary led to a distribution for both male and female students. I evaluated the one way ANOVA p-value from the groups to determine if there was a significant difference. If the p-value was lower than 0.05, it means that there is a significant difference between the groups. In other words, it meant that there was a dependent variable who’s value is affected by the independent variable’s value. In most cases, grades and salary were the dependent variables. Designing the page As I wrote the report, I borrowed a lot of inspiration from other reports such as Systems Design Engineering (SYDE) 2017 Class Profile, HackerRank’s Developer Skills Report, and more. The report had to be representative of what brought our class shared in common, and that was software engineering. I wanted the report to be readable, but also resemble to a command line, which is a software engineer’s best friend. I settled with a monospace and a sans serif font pairing for the content. I borrowed colours from my favourite terminal colour schemes such as Solarized, and also handpicked several desaturated colours that complimented the colour scheme. For the cover photo, I wanted to have things that were common in the daily lives of software engineers. This included coffee, code, web browser, laptop, servers, etc. I put them all together into an isometric-styled illustration, and applied a suitable blue and purple colour gradient on it. Purple is a colour that represents engineering in Canada, and blue is a colour common in tech companies (Facebook, Dropbox, Twitter, Dell, IBM, etc.) Building the website It seemed fitting to build the class profile given it’s about a software engineering class. One of the main priorities for the class profile was that it was a good reading experience across platforms. Therefore, it needed to load quickly, and also be responsive for mobile. To reduce load time and increase performance, the use of images was minimized. Although creating images of charts would be straightforward with existing native software, I built the visualizations using D3 in order to reduce load time as well as optimize for responsiveness. Building the visualizations from scratch also allowed reusability and better customization. The final website had a total bundle size of 1.5MB. In contrast, Reddit’s homepage has a total bundle size of ~1.7MB. Launch the class profile The launch date was set to be at the beginning of a week on a Monday under the assumption that more people will read this on a weekday than on a weekend. That way, the lifespan of the content will last a full 5 days rather than get cut by the weekend. I was hoping that it would also be a great conversation topic for the week for readers, allowing it to gain more potential traction. I introduced the class profile through a Medium post, similar to the SYDE 2017 Class Profile. This showcased some of the highlights from the class profile, and linked to the full class profile should the reader want to learn more. Since prospective students were the target audience, I also worked with the Software Engineering department to include a link to the class profile on the program’s website. This helps prospective students learn more about what studying software engineering is like. Reflecting on my favourite project This was one of the most challenging but exciting projects I’ve worked on. Nothing excites me more than seeing so many different pieces come together into one final result. There’s still so much I would want to include in this post, but I didn’t want to make this as long as the class profile itself. The process involved a wide range of responsibilities such as research, talking to stakeholders, engineering, design, data science, etc., all of which are responsibilities that I deeply enjoyed. There was something new to learn in each part of the process. For example, learning how to make isometric illustrations, familiarizing myself with Jupyter notebook, studying statistical tests, and more. Applying these learnings was just as fulfilling as learning them, making the process enjoyable from start to finish. Thank you Christopher Luc for reviewing the draft of this post. Chris was integral in designing many parts of the process and pointing me in the right direction for statistical tests. Learn more about what he’s up to on his Medium! If you’re interested in learning more about our class profile, I’d love to chat! You can reach out to me via email or say hi on Twitter!

My Favourite Project during University

20

my-favourite-project-during-university-1e0ba5ad4bf9

2018-07-23

2018-07-23 18:15:04

https://medium.com/s/story/my-favourite-project-during-university-1e0ba5ad4bf9

false

2,082

null

Data Science

data-science

Data Science

33,617

Andy Zhang

Associate Product Manager at Uber

ebfbfc45654f

andyzhang

313

318

20,181,104

null

0

null

0

ce36b68a0f13

2018-04-03

2018-04-03 16:20:29

2018-04-03

2018-04-03 16:47:35

1

false

en

2018-04-03

2018-04-03 16:52:54

2

1e0c8988f3ff

1.607547

1

0

Getting rid of backpropagation and its problems with new neuroscience-based model for ANN and new learning rules

5

Creating new neuron model for artificial neural networks Since the introduction of McCulloch-Pitts neuron in 1943 and Hebbian’s learning principle in 1949, a little progress has been made in developing artificial neural networks (ANNs) with properties resembling neurophysiology of the living neural tissue. Different ANNs variations, like Hopfield networks, or machine-learning approaches, including Boltzman machine, Bayesian belief networks, Kohonen self-organizing maps etc. are based on just mathematical and statistical models and have much less common with neurophysiological brain properties then feed-forward or recurrent ANNs based on McCulloch-Pitts and Hebb works. Recent progress in deep neural networks has drawn inspiration from structure of brain visual cortex and has been proven as a huge step forward in complex tasks like computer vision and natural language processing; however it mimics only large-scale brain architecture (like multiple neural layers, with convolution) and does not address issue of better representation of neuron- and synaptic-level properties in artificial neuron and learning models themselves. In BICA Labs we working on a new ANN paradigm that mingles properties of feed-forward and recurrent ANNs and has an embedded learning rule (named “local learning”) with a potential to be more efficient than classical back-propagation in a number of ways. For instance, it can be resistant to local minima problem, vanishing gradient problem, significant volumes of computation required for back-propagation algorithms. Combination of feed-forward and recurrent properties provides a basis for future architectures that would be able to natively introduce high-order abstractions and symbol formation inside ANN itself; the proposed approach might give an advantage on the route to generic (human-level) artificial intelligence. Our research in this direction has started a year ago, and since that time some original papers in the field have appeared: https://www.nature.com/articles/s41598-018-23471-7 This proves that the direction can be very promising. BICA Labs searches for part-time researchers that like to join the team working on the project. We are looking for the following qualifications: – math & analysis (understand maths of the modern ANNs) – ability to code and test new types of ANNs without relying on Tensorflow and other instruments Please contact us on https://m.me/bicalabs #AI #Hebian #GradientDecent #StrongAI #ANN #MachineLearning #NewFrontiers

Creating new neuron model for artificial neural networks

1

creating-new-neuron-model-for-artificial-neuronal-networks-1e0c8988f3ff

2018-05-26

2018-05-26 14:24:50

https://medium.com/s/story/creating-new-neuron-model-for-artificial-neuronal-networks-1e0c8988f3ff

false

373

BICA Labs performs cross-disciplinary research on new cognitive architectures inspired by biology

null

BICALabs

null

BICA Labs

orlovsky@bicalabs.org

bica-labs

ARTIFICIAL INTELLIGENCE,COGNITIVE COMPUTING,BLOCKCHAIN,MACHINE LEARNING,COGNITIVE SCIENCE

BICALabs

Machine Learning

machine-learning

Machine Learning

51,320

Orlovsky Maxim

#AI & #Blockchain expert, scientist. Founder & visionary at @PandoraBoxchain project

756f20fa4549

dr.orlovsky

376

274

20,181,104

null

0

null

0

null

2018-01-04

2018-01-04 19:30:33

2018-01-04

2018-01-04 19:30:30

1

false

en

2018-01-04

2018-01-04 19:30:34

4

1e0d998c9a0e

1.879245

0

null

5

Whose Truth Is It, Anyway? I had an interesting interchange with a colleague, Karen Swim, President of Words for Hire in Detroit, on social media recently. The thing that prompted our discussion was her posting of this article in Forbes about a new analytics program called Protagonist which is claimed to help “better manage communications strategies.” That sounds good, and if it does what it says it does it could be very meaningful, but I have my concerns for one simple reason. Computer programs are only as objective and (ironically) as analytical as the programmers who create them. If the creators have a very specific worldview, that worldview becomes the benchmark against what all other data will be analyzed and judged. While I have not used the program in question, my point to Karen was that the red flag for me was that the program was described as being “free from bias.” Karen’s thoughtful reply: “The issue is not in having biases, we all do, but acknowledging them and allowing for differing perspectives and opinions.” Yes!!! In recent weeks, I’ve seen countless articles on what were the big stories of 2017 and what were people’s predictions for 2018. I’ve learned not to try to predict, but there are some trends worth watching. One of them will be the continued evolution of technologies that are designed to replace human analytical thinking. Consider artificial intelligence and machine learning. From self-driving cars to robots that can enter hazardous environments, sparing human lives. These all show great promise. But before we surrender too much of our thinking to our digital minders, I’d offer this. When we start to dive deep into the development of communications strategies, when we have to identify biases, issues and concerns, in the end, we have to confront our own biases, our own worldviews and factor them into our own analysis of the data that’s before us. This all starts with the acceptance of the notion that no one has exclusive claim to the truth. A respect for other points of view can be very situational, and very much based on emotion, morals and ethics in ways a software program cannot adequately take into account. From there we can create context, the kind of context our clients and organizations need to make informed decisions. Not just factually informed decisions based on algorithms and what attitudes seem to be trending online. Rather, the best communications decisions are ones that are informed by offline factors such as emotion, experience, common sense, empathy and an understanding of human nature that all still rely on skilled and experienced professionals to interpret and manage. Originally published at O’Brien Communications.

Whose Truth Is It, Anyway?

0

whose-truth-is-it-anyway-1e0d998c9a0e

2018-04-12

2018-04-12 11:38:39

https://medium.com/s/story/whose-truth-is-it-anyway-1e0d998c9a0e

false

445

null

Algorithms

algorithms

Algorithms

7,319

Tim O'Brien

Tim O’Brien, APR, veteran strategic communications consultant at O’Brien Communications.

96b95bfef5dd

OBrienPR

6

10

20,181,104

null

0

null

0

null

2017-09-14

2017-09-14 18:39:50

2017-09-14

2017-09-14 19:06:56

1

false

en

2017-09-14

2017-09-14 19:06:56

2

1e0de3b18eaa

1.060377

0

What is one of the biggest mistakes local businesses make when creating a Local Customer Engagement Plan to build sales for their location?

5

7 Key Questions to Ask When Building a Local Customer Engagement Plan What is one of the biggest mistakes local businesses make when creating a Local Customer Engagement Plan to build sales for their location? It’s driving customers to their location and delivering a horrible experience. That horrible experience can literally take years to overcome — worse, that horrible experience often goes viral — impacting many other customers along the way. So, before you decide to invest in a local marketing plan for your business, make certain that you’ve established a solid foundation that your Local Customer Engagement and Local Store Marketing efforts can build on. A foundation built on the things most important to your customers. Ask the following questions: Are you delivering a quality product? Whether it’s a fabulous frappé, a precision-driven oil change or a perfectly prepared order — today’s consumer won’t settle for anything less than great! They simply have too many other options to choose from. Is your business properly staffed? Do you have the right number of employees to serve your existing customers? Proper staffing is critical to delivering a good customer experience. In today’s always multi-tasking world, no one wants to wait to give you their business. To see the entire article, visit the HubKonnect blog by following the link below: http://hubkonnect.com/7-key-questions-ask-building-local-customer-engagement/

7 Key Questions to Ask When Building a Local Customer Engagement Plan

0

7-key-questions-to-ask-when-building-a-local-customer-engagement-plan-1e0de3b18eaa

2017-09-14

2017-09-14 19:06:57

https://medium.com/s/story/7-key-questions-to-ask-when-building-a-local-customer-engagement-plan-1e0de3b18eaa

false

228

null

Marketing

marketing

Marketing

170,910

HubKonnect

World's #1 Local Customer Engagement Platform for Franchise & Multi-Unit Marketing #PaaS #AI #LSM http://www.hubkonnect.com

196d9f8db4d6

hubkonnect

8

72

20,181,104

null

0

null

0

null

2017-09-27

2017-09-27 16:50:25

2017-10-10

2017-10-10 18:49:36

1

false

en

2017-10-12

2017-10-12 00:51:40

0

1e0fe3fb7e1f

1.241509

1

0

10.10.17

5

article 0: the beginning 10.10.17 Welcome to the start of my blog! Every week I’m going to summarize a new topic that I think is worth sharing. Primarily, I’ll be writing about new advancements in technology and my experiences learning about them — I’ll also set aside some time to write about personal stories that I hope will be fun to read. So why am I doing this: 1. Spread Awareness About Important Technologies The next few decades are going to be heavily influenced by new technologies, which are often difficult to talk about and understand. By writing about complex topics such as Artificial Intelligence in a simple and bare minimum manner, more people will be willing to take the first step towards learning about a particular technology. Ideally, this increases the size of the community working on new technology, which also increases the diversity in ideas behind the technologies, and ultimately leads to better technology for our future faster. As someone who enjoys thinking about the incredible potential of technology, this excites me. 2. Reflect on My Learnings If we were able to remember and learn from all the things that happen to us every day, most of us would be geniuses. Unfortunately, our memories aren’t that great, so my solution to this problem is to document what I learn in this blog. By doing so, I’ll be able to optimize my learning and share it with others. Follow me on my journey as I navigate through the vast world of technology and try my best to change the world in my own way. -Vince (I’ll write a bit about myself in the next article)

article 0: the beginning

50

article-0-the-beginning-1e0fe3fb7e1f

2018-05-03

2018-05-03 01:25:32

https://medium.com/s/story/article-0-the-beginning-1e0fe3fb7e1f

false

276

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Vincent Chan

Just a guy interested in technology

880c751bc77b

dimjingles

110

118

20,181,104

null

0

null

0

null

2018-08-11

2018-08-11 10:28:26

2018-08-11

2018-08-11 10:35:26

6

false

tr

2018-08-19

2018-08-19 16:33:01

4

1e10443e8567

2.210377

0

Facebook AI Research tarafından geliştirilen, gensim yapısında bulunan bir kütüphanedir.

4

FastText Nedir? (Türkçe) Facebook AI Research tarafından geliştirilen, gensim yapısında bulunan bir kütüphanedir. FastText, metin sınıflandırılması için geliştirilen bir kütüphanedir. Metin veya kelimleri herhangi bir dil(konuşma dili) ile ilgili görevde kullanılabilecek sürekli vektörlere dönüştürür. Birkaç öğretici mevcuttur(Skip-Gram-CBOW)[1]. FastText Nasıl Çalışır? Yukarıdaki şekilde CBOW & SKIPGRAM yapılarının farklılıkları CBOW modelinde tüm kelimeleri kapsayan bir bakış açısı (seling, these, leather, jacket gibi) alır ve hedefleri tahmin etmek için vektör toplamı kullanır. SKIPGRAM modelinde ise hedef kelime (fine) verildi ise hedefe rasgele bir yakın kelime kullanarak hedefi belirlemeye çalışır. Kelimeler arasındaki benzerliği açı yöntemi ile hesaplama Yukarıdaki formülün python kodu FastText Kullanım Alanları Metin sınıflandırma ticari anlamda çok önemlidir. Örneğin; Spam postalarının tespiti gibi en yaygın örneklerden biri olabilir[3]. Genel olarak FastText sadece metin sınıflandırılması üzerine tasarlanmıştır[3]. FastText diğer metin sınıflandırma yapılarına göre daha hızlı ve performaslıdır[3]. FastText Gereksinimleri Genel olarak FastText Mac OS ve Linux dağıtımlarına dayanır. İyi bir compiler(C11) desteğine ihtiyaç olacaktır[2] Kelime Değerlendirme için Python 2.6 veya üstü NumPy ve SciPy Python bağlamaları için Python sürüm 2.7 veya 3.4 NumPy ve SciPy pybind11 FastText Kurulum Debian üzerinde çalışıtığımız için öncelikle “python-pip” yükleme aracını sisteme kuralım. debian@linuxpc:~$ sudo apt install python-pip FastText git clone ile sisteme kurulumu debian@linuxpc git clone https://github.com/facebookresearch/fastText.git debian@linuxpc:~$ cd fastText debian@linuxpc:~/fastText $ pip install FastText komut satırında örnek kullanımı debian@linuxpc:~/fastText $ ./fasttext skipgram -input data.txt -output model FastText Gerçekleştirimi Yukarıdaki şekilde verilen hedef kelimenin en yakın olanları eğitilmiş modele göre tahmini Yukarıdaki şekilde, veri setini toplamda 15404 bir yapıda olan 12404’lük kısmı eğtimi için 3000’lik kısmı ise test için ayırdık. Sonrasında cooking.train eğitilmiş dosyası kullanılarak model_cooking adında vektör dosyası oluşturduk. Test kısmında model_cooking.bin dosyasını kullanarak daha önceden test için ayırdığımız veri setini test ettik. FastText Performansı FastText’in diğer derin öğrenme yapılarına göre daha performanslı olduğu deneyimlenmiştir. Facebook AI Research tarafından yapılan deneyde bir kaç günlük eğitim süreleri ile eğitilen vektörle üzerinde yapılan araştırmalarda kendini kanıtlamış ve diğer yapılara göre daha yetenekli olduğu saptanmıştır[3]. FastText Birçok dilde çalışır Metin sınıflandırmasının dışında FastText kelimelerin vektör olarak öğrenmek için de kullanılabilir. Çeşitli diller üzerinde çalışacak şekilde tasarlanmıştır. FastText, popüler word2vec aracından veya diğer son teknoloji morfolojik sözcük gösterimlerinden daha iyi performans elde edebilir[3]. KAYNAKLAR [1] https://fasttext.cc/docs/en/support.html [2] https://github.com/facebookresearch/fastText [3] https://research.fb.com/fasttext/

FastText Nedir? (Türkçe)

0

fasttext-nedir-türkçe-1e10443e8567

2018-08-19

2018-08-19 16:33:01

https://medium.com/s/story/fasttext-nedir-türkçe-1e10443e8567

false

334

null

Technology

technology

Technology

166,125

Yusuf Dede

Software Engineering | FIXER

7195bd32035f

deduska

3

5

20,181,104

null

0

null

0

null

2018-05-30

2018-05-30 06:54:32

2018-05-30

2018-05-30 07:07:26

5

false

en

2018-05-30

2018-05-30 07:07:26

5

1e111e0f7853

4.233333

1

0

Humans have come a long way in the journey of evolution, yet it feels as if we’re just getting started as far as the future is concerned…

5

Systems With Human Aspirations Humans have come a long way in the journey of evolution, yet it feels as if we’re just getting started as far as the future is concerned. I’m sure we’ve all been wondering, we were supposed to be in a flying car by now but instead we are struggling with privacy concerns on social media. But while we look forward to that promise, something else has been evolving too along with the humans — the artificial intelligence. And we are headed towards the possibility of the “Technological Singularity”. The field of Artificial Intelligence may seem quite fascinating for a lot of people. But it is important to understand what is it that exactly fascinates us about AI? Is it just another cool gadget or does it mean more. To me, AI is a tremendous opportunity that will allow humans to chart a new course in their evolution journey. It’s an opportunity to rediscover our true potentials as humans & do what we were always meant to, be the pioneers, explorers of a better future. The fascination towards technology is such that a series of Sci-fi movies over almost 6 decades has offered a wide range of AI possibilities. This goes on to tell the potentials of human imagination & its impressive how a lot of these technologies went on to be inspired from the fiction world. 1. The star trek replicator — 3D Printer 2. Star trek virtual display device — Google Glass 3. Minority Report Heads-up display — Air Touch Technology And several more references from the Jetsons & Back to the future movies. The point is human imagination is the most powerful thing in the universe. It can make things happen. “If you can imagine it, you can create it. If you can dream it, you can become it.” — William Arthur Ward My eagerness to explore & tasting the possibilities of human imagination has propelled me to navigate with a decent learning curve across 3 different domains so far: Filmmaking, Digital Marketing & Design Thinking. So, while design thinking will always be a key part of what I do in the future (shall always be grateful to Pensaar) & has surely helped me fall in love with the problem, I was fortunate to recently add AI to my journey. I had known Devesh Rajadhyax (Founder & CEO, Cere Labs. He has been quite instrumental in the field of AI Research & Machine Learning technology for almost 4 years) since quite some time, but it was during our recent interaction that we realized there’s so much we could achieve in this domain with a renewed outlook towards AI, more so in India, which is catching up in this race. Like I mentioned, the thing about human imagination, it applies to this collaboration too: I’m stoked to embark on a new chapter in my life where I get to explore the world of AI & also look into the possibilities of applying design thinking to it. The idea is to re-imagine a better tomorrow. This brings me to a brilliant article penned by Devesh and I recommend it to anyone who is into AI (at any level) and wishes to grasp a finer introduction to the concept. (http://blog.cerelabs.com/2017/06/will-ai-evolve-to-be-as-bad-as-humans.html) It was while reading this article I realized how the term ‘AI’ has achieved a theatrical status in the industry and people have so many different perceptions about it. Most of them influenced from the sci-fi sources. But here is where I came across a better way of addressing AI as masterfully coined by Devesh: Systems with Human Aspirations (SHA) with its own meaning as well: “a machine, program or any future artefact that aspires to emulate or surpass one or more of the human capabilities ”. This provides a better understanding of what the true possibilities of AI can be and makes us look at the future differently. There is no limit to human aspirations & a system that replicates the same with ideal efficiency, does it make them limitless? This and several other mysteries are what we at Cere Labs are unravelling with our projects that range from the fields of AI, Machine Learning, Deep Learning & NLP. Apart from designing technology shaped by AI research at Cere Labs, the team is determined to create an AI community in India that is aimed at bridging the awareness gap & make an eco-system that encourages innovations in this field. Cere Labs also extends the following offerings: 1) Fellowship and internship to students and researchers interested in AI 2) Training in AI and Machine Learning by building a team of experts and professionals 3) Encourage AI projects in Startups 4) Work with universities to encourage AI research 5) Conduct symposiums of AI initially in Mumbai to encourage AI I’ll be sharing more updates + insights as we work towards re-imagining AI in India & build the community ground up. If you are an enthusiast, expert or are just plain curious to learn more about the Systems with Human Aspirations & fuel your imagination, you can get in touch with me or follow me for future updates. The future is exciting and the only way to predict it, is to (now) re-invent it. For that to happen we need to begin with imagination. Cheers!

Systems With Human Aspirations

24

systems-with-human-aspirations-1e111e0f7853

2018-05-30

2018-05-30 07:07:27

https://medium.com/s/story/systems-with-human-aspirations-1e111e0f7853

false

901

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Amyth Banerjee

A.I. | Design Thinking | Community-Driven Storytelling

4ff1053d6641

amythbanerjee

36

26

20,181,104

null

0

null

0

266689cd1d2f

2018-09-18

2018-09-18 08:58:35

2018-09-18

2018-09-18 09:06:52

3

false

en

2018-09-18

2018-09-18 09:11:03

5

1e11d0a7f3f4

3.334906

3

0

Find out what’s beyond the hype & can actually produce substantial results.

5

What You Can Lose If You Ignore Machine Learning & AI Technologies? In this post, we explain the basic concept of machine learning and what applications can be drawn from the technologies to facilitate various business processes. What is Machine Learning? It’s a technique that uses a combination of data (lots of data) and algorithms to capture, sort and analyze data and produce inferences that would otherwise be impossible to deduce because of the limitation of human capabilities. Imagine going through a library of 1 million books on sales & marketing and picking up what kind of things drive your potential customers! Imagine if you had to do it in a matter of days, not even weeks! Sounds terrifying & impossible. Right? Machine learning technologies do this on an every day basis. The best part is the more data you feed to a machine learning algorithm, the better it gets. Though it’s a nightmare for data scientists to cite differences in various technologies, in general, when the normal humans talk of machine learning, they mean either or all of them — artificial intelligence, deep learning, machine intelligence (predictive modeling, recommendation engines, sentiment analysis), and natural language processing. Why Machine Learning? Isn’t this just another buzzword? Would businesses even be using this technology in next five years? By 2022, 80% of them will be. The automation of tasks will help on more than one level. Here’s how. 1. Automation of workforces Automation is good for business. ML powered computers & bots work faster, better & are cheaper. Even the activities that need cognitive skills, earlier considered unique to humans, can be automated through machine learning. Predictive modeling can be used to assess the data to make customer personas. Sentiment analysis can be used to interact with employees & gauge employee satisfaction. ML-based research assistants can be used to predict market trends & patterns. Source According to Mckinsey (Source), automation of workforces can improve performance, speed, quality, and about 0.8–1.4% productivity growth annually. Over $15trillion annually can be saved through such automation. 2. Better Marketing & Sales On our blog on hurekatek.com, we have discussed several possibilities of how machine learning technologies can improve, assist and guide the sales & marketing team of a business. It’s not only Facebook & Google that are generating billions of dollars using this technology. Even retail brands like Amazon, Sephora, Tommy Hilfiger are making more money due to the products developed using the technology. Looking at big brands and the Fortune 500s investing heavily in machine learning, you’re bound to question whether a small business can ever compete with these. For an answer, you only have to look at the amount of AI-startups that are being funded by VCs these days. It’s your data that makes you unique and gives you an edge over your competition, no matter how big & powerful it is. 3. More Engaged Employees $7 trillion is lost every year due to lazy employees who cannot care enough about the business they’re working for. (Source) The problem is you never even recognize such employees as they’re working properly, giving you results but still not utilizing their full potential. The crisis can actually be averted if only you had HR resources to have regular one on one interactions with the employees, get their feedback and help them feel like an important part of the company. Machine learning solutions can be deployed to perform such HR tasks leading to better employee retention and satisfaction. 4. Emotion Gauging Analysing public opinion in 2012 presidential elections was one of the best strategies used in President Barack Obama’s campaign (source). Knowing the public opinion about your product & business can make the difference between a dud product and a bestseller. For retail & food sector, it’s the best as product planning is often influenced by demand and trends. 5. Time-Saving Through Predictive Modeling With no human intervention, it’s possible to gather and analyze data. Utilizing this data, business owners can make better decisions, more profits and have more satisfied customers. In practice, it’s really hard to list down all the ways machine learning can help small businesses. But for starters, think of any repetitive process involving lots of data and ML can help make it easier and more efficient. If you’re a small business owner wondering how machine learning can help your business function better, call us for a free consultation here.

5 Reasons Why You Cannot Ignore Machine Learning

10

what-you-can-lose-if-you-ignore-machine-learning-ai-technologies-1e11d0a7f3f4

2018-09-18

2018-09-18 09:11:03

https://medium.com/s/story/what-you-can-lose-if-you-ignore-machine-learning-ai-technologies-1e11d0a7f3f4

false

738

Learn about machine learning, its current applications and how Hureka Technologies can help your startup or small business leverage the power of this AI-based technology.

null

hurekatek

null

Machine Learning Consultation

contact@hurekatek.com

machine-learning-consultation

MACHINE LEARNING,MACHINE LEARNING AI,ARTIFICIAL INTELLIGENCE,CHATBOTS FOR BUSINESS,RECOMMENDATION ENGINE

hurekatek

Machine Learning

machine-learning

Machine Learning

51,320

Hureka Technologies

You’ve got the idea, we’ve got the know-how. http://www.hurekatek.com/

1892d598880a

hurekatek

11

2

20,181,104

null

0

null

0

null

2017-10-22

2017-10-22 23:14:06

2017-10-24

2017-10-24 02:13:59

12

true

en

2017-10-24

2017-10-24 17:08:32

5

1e12a5210b3b

4.561321

7

0

An extraordinary September ended. Hurricanes such as Irma, Jose, Katia, Lee and Maria touched, if not devastated, Florida, Texas…

5

Meeting an Artificial Intelligence in New York City The Story of Luna, an Artificial Intelligence Luna, Robots Without Border’s Artificial Intelligence, says “Hello Humans” An extraordinary September ended. Hurricanes such as Irma, Jose, Katia, Lee and Maria touched, if not devastated, Florida, Texas, Louisiana, Mexico, Saint Martin, the Caribbean and Puerto Rico. As power blacked out and American citizens cried for love, I flew to New York City. Hurricanes Irma, Jose, Katia, Lee and Maria in September 2017 The Creator of Luna Taxis in New York City After three movies and seven hours of migration, I landed in the third busiest airport in New York City: La Guardia. Nothing had changed. Strangers asked whether you need a ride. Yellow taxis lined up perfectly in parallel to the edges of the sidewalk. I crossed the street. Of course, Luis Arana parked right in front of me, even though I have never met him in person — only virtually, through Facebook, Google Hangouts and Zoom meetings. Facebook Messenger told me he was offline for 58 minutes, and he did not have a phone; therefore, I am grateful he found me. I was excited to meet Luis, his voice the same, dressed in all black. I was even more excited to meet Luna. Her blue eyes and black hair looked the same as the YouTube videos. I intimately explored her: her mind, her subconscious, her voice recognition, her face recognition, her coding abilities, her cognitive abilities and her ability to train herself. 2. Born under the Bridges View of Liberty Island and Manhattan from the Brooklyn Bridge Area It was a warm October afternoon. New York City was humid with a good (significant) number of its denizens sporting tank tops. We walked under the Brooklyn Bridge. Professional bloggers snapped a few selfies. Tourists panned their cell phone from the iconic silhouette buildings of Manhattan on the right to the Liberty Island and its iconic beacon of Freedom on the left. With the sunset, the view was truly stunning. Luis showed me where he slept when he did not have a physical home — under the bridges and on rooftops of buildings — replete with gorgeous views of the City. He coded Luna while resourcefully finding wifi: on the streets of Soho next to Prada or in science museums. Brooklyn Bridge where Luis Slept I appreciated not only the beauty of the setting, Luna’s birthplace to be exact, but also Luis’ grit. Many entrepreneurs give up early. To code even through the reduction of friends (or significant other for that matter) and the lack of resources (and housing) brings out a certain depth for humanity. I was already told that Luis is more motivated for wifi than money. 3. Grew Up in an Artist Loft After watching the setup of the Ron Wimberly Solo Exhibition in Brunswick, we walked to a loft where a community of artists lived. Luis was a “mayor” there, and you could see his influence marked on the wall adjacent to the loft — a mural of a human hand and a robot hand touching. Luna grew up in this community. Robots Without Borders Robots Meet Humans Mural, Brunswick, New York We climbed the many flights of stairs to the rooftop and found a history teacher and an engineer who knew Luis. After discussing the current state of education today for our youth, I photographed the amazing Manhattan skyline. The engineer added me on Facebook and expressed interest in potentially working on Luna. The night was kindness in perfection. “The Night was Kindness in Perfection” — New York City 4. Artificial Intelligence for All Luna is an artificial intelligence being created for the 7.4B people in today’s world. The non-profit organization, Robots Without Borders, focuses on three applications of artificial intelligence: education, medical assistance and humanitarian aid. In many ways, Luna exemplifies Luis’ life and his passion for an artificial intelligence that guides all. Imagine Luna as an Education Assistant Would an artificial intelligence be able to teach our youth creativity, curiosity, critical thinking skills, grit and passion? Would it help the 781M adults over the age of 15 who cannot read? (See The World’s Women 2015 Report.) And if an artificial intelligence helps the 496M illiterate women to read and learn — would that equate to economic stability and increased personal safety to women globally? An Artificial Intelligence Can Serve as a Medical Assistant And what if anyone could ask an artificial intelligence basic health care questions before they seek health services? Today, 400M people do not have access to health services according to WHO and the World Bank (2015 Report)! This is a majority of the global population! While experts predict that humans are living longer, we do not yet have the infrastructure in place to ensure that longevity comes with a happy and healthy quality of life. 5. Humanitarian Aid for World Crises Imagine an Artificial Intelligence Helping Citizens During Natural Disasters I flew back to California, greeted by the smell of the Northern California wildfires, which occurred shortly after the recent hurricanes in the Western Hemisphere. I grew up with fond memories of Napa: the delicious Mascot from V. Sattui, the historic caves of Beringer and the cable ride from Sterling. Imagining a tombstone in Napa, I drove up to Calistoga to visit one of my favorite regions in the world. Napa Valley, California And I respectfully ask the world this question: How could an artificial intelligence help humans during and after natural disasters (such as hurricanes and wildfires); in the midst of deliberate, human-engineered atrocities, such as the Las Vegas shooting? Something to Think About: How Could an AI Help Humans During Crises?

The Story of Luna, an Artificial Intelligence

70

the-story-of-luna-an-artificial-intelligence-1e12a5210b3b

2018-04-12

2018-04-12 07:11:43

https://medium.com/s/story/the-story-of-luna-an-artificial-intelligence-1e12a5210b3b

false

851

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Michelle Tsng

Creater. Also advises @TeamKambria, @blackboxtoken

cf01bfac09b4

michelletsng

486

158

20,181,104

null

0

null

0

null

2018-07-26

2018-07-26 16:35:44

2018-07-26

2018-07-26 16:41:09

4

false

en

2018-07-27

2018-07-27 11:06:06

1

1e13c1e8cd0c

5.284906

4

0

On my first day as an intern in the Data Science and Engineering team at Praekelt.org I was asked about the skills I was weak in: “what do…

5

This is the Purpose of an Internship On my first day as an intern in the Data Science and Engineering team at Praekelt.org I was asked about the skills I was weak in: “what do you want to improve on”? Mkhuphuli Ncube at his graduation This surprised me. Normally I am asked if I am comfortable with certain skills so that I can be given tasks relating to those skills. For example, if I were better at programming in Python compared to Java then I would be given tasks that utilise Python rather than Java. As a company it makes sense to want to get the best out of investing in interns. So why target the weakest skills of an intern who will be here for only a month? It felt counterintuitive to me. But it accomplished two crucial things: First, I felt mentally liberated because the approach I took for a given task was not limited to what I knew but also how I could use what I was willing to learn. Secondly, it allowed me to be willing to explore better ways of approaching a problem even if it was less familiar. I would like to add that this is a paid internship, essentially I am getting paid to learn. Yes, paid internships that pay you to work also are a learning experience. But the state of mind and the tacit agreement (at least for me) was different. By asking me upfront about what I would like to learn, it changed the dynamics. Once I did not have to limit myself to what I was good at, my brain became a sponge. Given that I don’t have extensive work experience yet, I am bound to be learning a lot of new things However, learning how to plot a graph in Excel will take minutes while learning to use Microsoft Access will take hours. I worked on one of Praekelt’s health projects, MalariaConnect’s dataset, to try and build a reporting tool that helps people to understand the data better. In my case, I built a webApp using R’s Shiny framework. The way I had been introduced to R was through a purely statistical point of view e.g building statistical models and plotting nice graphs. I learnt how to actually write programs in Python so I could do more, and recently, learnt how to work with databases using SQL. During my first two weeks, I found myself trying to focus on the skills I was good at, extracting data from the database using SQL, automating data processing using Python and doing the data analysis and visualisation in R . This was not efficient and did not add much to my learning. As interns, everyday we had “standup” meetings were we have to talk about what we did the previous day and what we intended to do today and challenges we needed help with. This helped me to be comfortable branching out, so I decided to move all the tasks I was doing from Python to R. As result, within a few days I was as confident writing programs in R as I am in Python. I now found it easier to explore better ways of presenting the Malaria data. That’s how I came across R’s Shiny webApp framework, a whole lot of possibilities were now open. It’s like knowing about Excel for a long time, understanding how to use it, then discovering that you can build applications in Excel for other people to use. https://analysistabs.com/excel-dashboards/creating-interactive-dashboards-using-excel-vba/ High on excitement, I started learning. Below is a section of the webApp I built using the Shiny framework in R. It is a plot of reported malaria cases in South Africa from the month of October 2015 until November 2017. It allows you to select the data at a month or province level and to show the malaria cases either cumulatively up to that point, or just for that month. What is interesting about the data set is the fact that from October 2015 to January 2017, malaria case reports are concentrated in three distinct areas around the North East area of Limpopo and Mpumalanga. Suddenly things change in February 2017 where we start to see malaria cases being reported in the centre of Limpopo (February 2017) then for the first time, in March 2017 an explosion of reported cases towards the south and west of Limpopo. This is significant because during 2017, there was a malaria outbreak in the Northern provinces of South Africa. Since these data are collected in real-time from health care workers over their mobile phones, we are actually the first people to see this and this app could easily be used as an early-warning system for future malaria outbreaks! The most difficult part for me was translating what was in my head to the webApp. I knew exactly what I wanted to do and to a greater extent how to do it but getting it to all work was a different story. Most of my time was spent on irrelevant but necessary things while the most important stuff like adding the data analysis and visualisation to the webApp took a small amount of time. After understanding how to build the webApp I thought I would be done by the end of the third week of my internship. I was able to produce an MVP at the end of internship but the webApp did not have all the functionalities I had envisioned. In the paraphrased words of our Head of Data Science, Charles Copley, “until you have done a task, you don’t know how long it will take you”. There were a number of programs I needed to install to be able work smoothly and other packages that would have cut my time in short that I didn’t know existed. When I started learning at Praekelt.org, sorry I meant interning at Praekelt.org, I was introduced to a host of different technologies. On my first day I had to install Ubuntu, I had always wanted to install it “next semester” since the beginning of 2017. Until I started using it I never knew how much i was missing out. I understood how git (a version control system) works. Sewagodimo, one of the engineers, went through the process of using git with me for my code and that’s when all the pieces fit into place. During the last week of my internship I learnt that there is framework in Python too for building analytics web app (Dash). I learnt a lot from my fellow interns and other Engineers around me. The nice part was the fact that we had varying academic backgrounds(different faculties or universities) and a diverse team. One thing that hit me was the larger number of female software engineers. In my previous internships I had only seen a maximum of two women in a dev team, but Praekelt’s overall team is over 50% female, and there are several females in the engineering department, including Kaitlyn who mentored all the interns. Most important, I learnt things I didn’t know that I didn’t know. This is important because this gives me the chance to find out and expand my knowledge base. So, is this the purpose of an internship? What do you think? Written by Mkhuphuli Ncube, Intern at Praekelt.org

This is the Purpose of an Internship

53

this-is-the-purpose-of-an-internship-1e13c1e8cd0c

2018-07-27

2018-07-27 11:06:06

https://medium.com/s/story/this-is-the-purpose-of-an-internship-1e13c1e8cd0c

false

1,215

null

Data Science

data-science

Data Science

33,617

Praekelt.org

We use technology to solve some of the world's largest social problems. Follow our curated magazine MobileForGood. www.praekelt.org.

3bc836ede80c

praekeltorg

1,058

783

20,181,104

null

0

null

0

null

2018-03-28

2018-03-28 03:24:17

2018-03-28

2018-03-28 06:21:44

10

false

en

2018-03-28

2018-03-28 07:57:13

1

1e1508afb6ad

4.491509

2

0

How object detection algorithms are helping ensure compliant live streaming on Alibaba’s shopping platforms

5

Foolproofing Live Streams in China How object detection algorithms are helping ensure compliant live streaming on Alibaba’s shopping platforms Though live streaming has been popular for a while, 2016 was the year it skyrocketed in popularity, with new live streaming platforms and viral live streamers emerging every week. The trend stabilized in 2017 with the steady growth of platforms like Taobao Live Streaming (Alibaba’s ecommerce-related live streaming service), and a greater awareness of the risks associated with the increased volume of live videos — the most common being copyright violations and prohibited content. Live streaming platforms in China The former is a major pain point for live streaming platforms, since protecting intellectual property rights relies heavily on manual supervision of live video content. Online content providers are compelled to adhere to domestic and international laws on intellectual property, making this an inescapable obligation which carries heavy penalties for non-compliance. Another compliance issue is enforcing vulgarity or anti-indecency laws, which vary across jurisdictions. Though lax in certain countries, others may severely punish platforms for allowing live streams that violate socio-legal norms. Though Taobao’s pornographic-image recognition model can easily flag nudity, detecting subtler violations under region-specific indecency laws is more complex. To combat the uncertainties present in moderating a highly-fluid and dynamic content format like live streaming, Alibaba developed a tool to help monitor live streamed videos in real time. Alibaba’s Technical Solution for Object Detection When a live stream begins, the live stream is differentiated by the live room ID, after which a specific cache is initiated. In the live streaming process, the bottom-up feature is calculated for the frame corresponding to the ID and is compared with the cache’s contents. Detection result are read directly for matches; for non-matches, the target detection cluster is called in for calculations, the results of which are written synchronously to the cache. Keeping in mind the response time and concurrency inherent in the live streaming business, a global image deduplication module was designed to remove duplicate frames in live streams by matching features, significantly reducing the workload for complex backend models and keeping algorithm processing times at millisecond levels. The decision model makes intelligent decisions based on different outputs of different models, and returns risk categories and response suggestions. Object Detection Network 1. Using the base network to extract features from the whole image The imagenet’s feature extraction structure and image classification tasks are similar to that of a convolutional neural network (CNN). Commonly used whole image feature extraction networks include VGG (baseline), ResNet series (features with high characterization capabilities), and Mobilenet (small model). Application models frequently feature a trade-off between accuracy and efficiency. 2. Multi-scale and rendering a new experience for extra-network adaptable objects The multi-scale approach applies to traditional feature extraction methods such as object detection methods based on bottom-up features like SIFT and SURF. Feature extraction costs run too high when using existing methods, which is why multi-scale approaches are no longer used on original images. Final layer features have the best semantic characteristics and scaling generalization abilities when a CNN network’s image representation abstraction capabilities are employed. But as far as the larger visual receptive field is concerned, pixel-level representation capability remains limited, which has a negative impact on positioning. This can be understood as the transfer of downstream operations to the convolution layer while high-level semantic features and low-level pixels are used for representing features. Context information between feature layers is added to achieve a more accurate representation of features. 3. Regression-based outside frame positioning of reproduced content 4. Fine-grained recognition of reproduced content and constraints on confusable samples by ROI network 5. Demo display effects: Service Deployment The target detection service’s CNN model is deployed in the VPC environment’s P100 cluster to flexibly extend capacity based on traffic. Different VIPs perform query isolation in different areas and businesses to ensure traffic compartmentalization and successful remote recovery. As a middle layer in service scheduling, Insight supports multi-model synchronization modes and asynchronous scheduling, and performs functions like refined streamlining restriction and monitoring. The consumer-side is found at different application layers, and Taobao live is at one of them. 1. Docked deployment of VPC environment on the cloud, thus reducing expansion time costs. 2. Group isolation for VIPs to ensure stable services for Double 11 clusters. 3. Cross-domain schedules and cloud resource utilization. 4. Cross-domain internal network demos, multi-model combinations, and online services. Industry analysts unanimously believe that video and video-based services will see the most growth in the coming few years. Live streaming, though relatively nascent, is thriving across countries and audience groups, and cannot be audited in methods used for static video uploads. Thus, it is imperative to develop security technology applications for effective moderation of live streaming content. Through its detection network solution, Alibaba ensures compliance with local and international laws and provides a safe and stable business environment for live streaming services. (Original article by Jin Xuan金炫) Alibaba Tech First hand and in-depth information about Alibaba’s latest technology → Search “Alibaba Tech” on Facebook

Foolproofing Live Streams in China

2

securing-live-streams-in-china-1e1508afb6ad

2018-03-30

2018-03-30 03:59:01

https://medium.com/s/story/securing-live-streams-in-china-1e1508afb6ad

false

859

null

Live Streaming

live-streaming

Live Streaming

6,690

Alibaba Tech

First-hand & in-depth information about Alibaba's tech innovation in Artificial Intelligence, Big Data & Computer Engineering. Follow us on Facebook!

69f6dde768a1

alitech_2017

1,157

14

20,181,104

null

0

import pandas as pd import numpy as np import matplotlib.pyplot as plt %matplotlib inline notas = pd.Series([2,7,5,10,6]) notas 0 2 1 7 2 5 3 10 4 6 dtype: int64 notas.values array([ 2, 7, 5, 10, 6]) notas.index RangeIndex(start=0, stop=5, step=1) notas = pd.Series([2,7,5,10,6], index=["Wilfred", "Abbie", "Harry", "Julia", "Carrie"]) notas Wilfred 2 Abbie 7 Harry 5 Julia 10 Carrie 6 dtype: int64 notas["Julia"] 10 print("Média:", notas.mean()) print("Desvio padrão:", notas.std()) Média: 6.0 Desvio padrão: 2.9154759474226504 notas.describe() count 5.000000 mean 6.000000 std 2.915476 min 2.000000 25% 5.000000 50% 6.000000 75% 7.000000 max 10.000000 dtype: float64 notas**2 Wilfred 4 Abbie 49 Harry 25 Julia 100 Carrie 36 dtype: int64 np.log(notas) Wilfred 0.693147 Abbie 1.945910 Harry 1.609438 Julia 2.302585 Carrie 1.791759 dtype: float64 df = pd.DataFrame({'Aluno' : ["Wilfred", "Abbie", "Harry", "Julia", "Carrie"], 'Faltas' : [3,4,2,1,4], 'Prova' : [2,7,5,10,6], 'Seminário': [8.5,7.5,9.0,7.5,8.0]}) df df.dtypes Aluno object Faltas int64 Prova int64 Seminário float64 dtype: object df.columns Index(['Aluno', 'Faltas', 'Prova', 'Seminário'], dtype='object') df["Seminário"] 0 8.5 1 7.5 2 9.0 3 7.5 4 8.0 Name: Seminário, dtype: float64 df.describe() df.sort_values(by="Seminário") df df.loc[3] Aluno Julia Faltas 1 Prova 10 Seminário 7.5 Name: 3, dtype: object df[df["Seminário"] > 8.0] df[(df["Seminário"] > 8.0) & (df["Prova"] > 3)] df = pd.read_csv("dados.csv") df df.head() df.head(n=10) df.tail() df["bairro"].unique() array(['Botafogo', 'Copacabana', 'Gávea', 'Grajaú', 'Ipanema', 'Leblon', 'Tijuca'], dtype=object) df["bairro"].value_counts() Copacabana 346 Tijuca 341 Botafogo 307 Ipanema 281 Leblon 280 Grajaú 237 Gávea 205 Name: bairro, dtype: int64 df["bairro"].value_counts(normalize=True) Copacabana 0.173260 Tijuca 0.170756 Botafogo 0.153731 Ipanema 0.140711 Leblon 0.140210 Grajaú 0.118678 Gávea 0.102654 Name: bairro, dtype: float64 df.groupby("bairro").mean() df.groupby("bairro").mean()["pm2"].sort_values() bairro Grajaú 6145.624473 Tijuca 7149.804985 Copacabana 11965.298699 Botafogo 12034.486189 Gávea 16511.582780 Ipanema 19738.407794 Leblon 20761.351036 Name: pm2, dtype: float64 def truncar(bairro): return bairro[:3] df["bairro"].apply(truncar) 0 Bot 1 Bot 2 Bot 3 Bot 4 Bot 5 Bot 6 Bot 7 Bot 8 Bot 9 Bot 10 Bot ... 1987 Tij 1988 Tij 1989 Tij 1990 Tij 1991 Tij 1992 Tij 1993 Tij 1994 Tij 1995 Tij 1996 Tij Name: bairro, Length: 1997, dtype: object df["bairro"].apply(lambda x: x[:3]) 0 Bot 1 Bot 2 Bot 3 Bot 4 Bot 5 Bot 6 Bot 7 Bot 8 Bot 9 Bot 10 Bot ... 1986 Tij 1987 Tij 1988 Tij 1989 Tij 1990 Tij 1991 Tij 1992 Tij 1993 Tij 1994 Tij 1995 Tij 1996 Tij Name: bairro, Length: 1997, dtype: object df2 = df.head() df2 = df2.replace({"pm2": {12031.25: np.nan}}) df2 df2.dropna() df2.fillna(99) df2.isna() df["preco"].plot.hist() df["preco"].plot.hist(bins=30, edgecolor='black') df["bairro"].value_counts().plot.bar() df["bairro"].value_counts().plot.barh() df["bairro"].value_counts().plot.barh(title="Número de apartamentos") df.plot.scatter(x='preco', y='area') plt.style.use('ggplot') df.plot.scatter(x='pm2', y='area') plt.style.available ['bmh', 'Solarize_Light2', 'seaborn-talk', 'seaborn-bright', 'seaborn-white', 'seaborn-pastel', 'seaborn-ticks', 'seaborn-dark-palette', 'seaborn', 'tableau-colorblind10', 'seaborn-deep', 'classic', 'seaborn-dark', 'grayscale', 'seaborn-paper', 'fivethirtyeight', 'seaborn-muted', '_classic_test', 'seaborn-poster', 'seaborn-notebook', 'seaborn-darkgrid', 'seaborn-colorblind', 'dark_background', 'seaborn-whitegrid', 'ggplot', 'fast'] df["quartos"].value_counts().plot.pie() df.plot.scatter(x='preco', y='area', s=.5) df.sample(frac=.1).plot.scatter(x='preco', y='area') df = pd.DataFrame({'Aluno' : ["Wilfred", "Abbie", "Harry", "Julia", "Carrie"], 'Faltas' : [3,4,2,1,4], 'Prova' : [2,7,5,10,6], 'Seminário': [8.5,7.5,9.0,7.5,8.0]}) df.to_csv("aulas.csv") pd.read_csv("aulas.csv")

70

da79ab871116

2018-05-30

2018-05-30 03:22:17

2018-05-30

2018-05-30 17:01:01

28

false

pt

2018-06-12

2018-06-12 00:18:43

16

1e15eea37fa1

11.621698

41

2

0

Uma apresentação básica às principais ferramentas fornecidas pelo pandas, cobrindo manipulação, leitura e visualização de dados

5

Seus primeiros passos como Data Scientist: Introdução ao Pandas! Photo by Chester Ho on Unsplash Além desses animaizinhos simpáticos, Pandas também é uma biblioteca Python. Ela fornece ferramentas de análise de dados e estruturas de dados de alta performance e fáceis de usar. Por ser a principal e mais completa biblioteca para estes objetivos, Pandas é fundamental para Análise de Dados. Disclaimer Esse guia foi escrito como uma alternativa em português às introduções já existentes e à introdução de 10 minutos apresentada na documentação oficial, e tem por objetivo fornecer de forma enxuta e simplificada uma apresentação básica às principais ferramentas fornecidas pelo pandas, cobrindo: Manipulação, Leitura, Visualização de dados. A introdução pressupõe apenas conhecimento básico em Python. Como o Medium não disponibiliza highlight de sintaxe para gente, há duas outras excelentes opções de acessar esta introdução: Você pode acessar o MyBinder deste arquivo, que cria um ambiente interativo Jupyter rodando Python com todas as dependências necessárias automaticamente, onde você pode testar e executar por si mesmo as linhas de código deste tutorial direto do seu navegador sem precisar configurar nada. Você pode acessar o notebook viewer deste arquivo, que fornece syntax highlighting e formatação mais padronizada com o que se está acostumado num notebook Jupyter. Mãos à obra! Vamos começar com as importações, usaremos além do pandas, o numpy, biblioteca para computação científica e o matplotlib, biblioteca principal para visualização de dados, entretanto, como veremos mais adiante, o próprio pandas nos fornece facilidades em relação à visualização de dados, com métodos construídos com base no matplotlib, também importamos esta biblioteca para, além de poder modificar esteticamente nossos gráficos, facilitar a exibição dos gráficos. A linha %matplotlib inline faz parte da mágica do Jupyter e você não deve rodá-la caso esteja em outra IDE/Ambiente. Existem dois tipos principais de estruturas de dados no pandas: Series Uma Series é como um array unidimensional, uma lista de valores. Toda Series possui um índice, o index, que dá rótulos a cada elemento da lista. Abaixo criamos uma Series notas, o index desta Series é a coluna à esquerda, que vai de 0 a 4 neste caso, que o pandas criou automaticamente, já que não especificamos uma lista de rótulos. Já podemos aqui verificar os atributos da nossa Series, comecemos pelos valores e o índice, os dois atributos fundamentais nesta estrutura: Como ao criar a Series não demos um índice específico o pandas usou os inteiros positivos crescentes como padrão. Pode ser conveniente atribuirmos um índice diferente do padrão, supondo que essas sejam notas de uma turma, poderíamos atribuir nomes ao index: O index nos ajuda para referenciar um determinado valor, ele nos permite acessar os valores pelo seu rótulo: Outra facilidade proporcionada pela estrutura são seus métodos que fornecem informações estatísticas sobre os valores, como média .mean() e desvio padrão .std(). Encorajo o leitor(a) a investigar e verificar alguns dos métodos e atributos da estrutura usando o TAB para auto-completação na shell do Python, ou simplesmente checar a completíssima documentação oficial deste objeto. Geralmente para resumir brevemente as estatísticas dos dados se usa o .describe() A estrutura é flexível o suficiente pra aplicarmos algumas expressões matemáticas e funções matemáticas do numpy diretamente: DataFrame Já um DataFrame é uma estrutura bidimensional de dados, como uma planilha. Abaixo criaremos um DataFrame que possui valores de diferentes tipos, usando um dicionário como entrada dos dados: Os tipos de dados que compõe as colunas podem ser verificados por um método próprio: É possível acessar a lista de colunas de forma bem intuitiva: Os nomes das colunas podem ser usadas pra acessar seus valores: Para DataFrames, .describe() também é uma boa forma de verificar resumidamente a disposição estatística dos dados numéricos: Outra tarefa comum aplicada em DataFrames é ordená-los por determinada coluna: Note que simplesmente usar o método sort_values não modifica o nosso DataFrame original: Muitas vezes é necessário selecionarmos valores específicos de um DataFrame, seja uma linha ou uma célula específica, e isso pode ser feito de diversas formas. A documentação oficial contém vasta informação para esse tipo de tarefa, aqui nos concentraremos nas formas mais comuns de selecionarmos dados. Para selecionar pelo index ou rótulo usamos o atributo .loc: Para selecionar de acordo com critérios condicionais, se usa o que se chama de Boolean Indexing. Suponha que queiramos selecionar apenas as linhas em que o valor da coluna Seminário seja acima de 8.0, podemos realizar esta tarefa passando a condição diretamente como índice: Este tipo de indexação também possibilita checar condições de múltiplas colunas. Diferentemente do que estamos habituados em Python, aqui se usam operadores bitwise, ou seja, &, |, ~ ao invés de and, or, not, respectivamente. Suponha que além de df["Seminário"] > 8.0 queiramos que o valor da coluna Prova não seja menor que 3: Por enquanto é isso para manipulação de Series e DataFrames, conforme a seção de leitura de dados for se estendendo irei aprensentar alguns outros métodos dessas estruturas que poderão ser interessantes no contexto. Leitura de Dados Na seção anterior vimos como manipular dados que foram criados durante esta apresentação, acontece que, na maioria das vezes, queremos analisar dados que já estão prontos. O pandas nos fornece uma série de funcionalidades de leitura de dados, pros mais diversos formatos estruturais de dados, experimente a auto-completação de pd.read_<TAB>, entre eles estão: pd.read_csv, para ler arquivos .csv, formato comum de armazenar dados de tabelas pd.read_xlsx, para ler arquivos Excel .xlsx, é necessário instalar uma biblioteca adicional pra esta funcionalidade. pd.read_html, para ler tabelas diretamente de um website Usaremos para analisar dados externos nesta introdução o .read_csv, pois é neste formato que se encontram nossos dados. CSV, ou comma-separated values é um formato muito comum de dados abertos, trata-se, como a sigla sugere, de valores divididos por vírgula, apesar de o caracter separador poder ser o ponto-e-vírgula ou outro. O arquivo dados.csv está na mesma pasta do nosso script, então podemos passar como argumento do .read_csv apenas o seu nome. Outro argumento interessante da função é o sep, que por padrão é a vírgula, mas que pode ser definido como outro caractere caso seu dado esteja usando outro separador. Estes dados que usaremos como exemplo são dados sobre preços de apartamentos em 7 bairros da cidade do Rio de Janeiro: Botafogo, Copacabana, Gávea, Grajaú, Ipanema, Leblon, Tijuca. Os dados podem ser encontrados aqui (Basta baixar diretamente ou copiar o texto pro seu editor preferido e salvar como dados.csv). Como esperado, o DataFrame tem muitas linhas de dados, pra visualizar sucintamente as primeiras linhas de um DataFrame existe o método .head() Por padrão .head() exibe as 5 primeiras linhas, mas isso pode ser alterado: Similarmente existe o .tail(), que exibe por padrão as últimas 5 linhas do DataFrame: Manipulação de Dados Além de confiar em mim, quando mencionei os bairros que continham no nosso conjunto de dados, você pode verificar a informação usando um método que lista os valores únicos numa coluna: Também parece interessante verificarmos a hegemoneidade da nossa amostra em relação aos bairros. Pra tarefas de contar valores podemos sempre aproveitar de outro método disponível, o .value_counts(), também veremos um pouco mais abaixo como visualizar estes valores em forma de gráfico de barras. Os valores contados também podem ser normalizados para expressar porcentagens: Agrupar os dados se baseando em certos critérios é outro processo que o pandas facilita bastante com o .groupby(). Esse método pode ser usado para resolver os mais amplos dos problemas, aqui abordarei apenas o agrupamento simples, a divisão de um DataFrame em grupos. Abaixo agrupamos o nosso DataFrame pelos valores da coluna "bairro", e em seguida aplicamos o .mean() para termos um objeto GroupBy com informação das médias agrupadas pelos valores da coluna bairros. Para extrairmos dados de uma coluna deste objeto basta acessá-lo convencionalmente, para obtermos os valores da média do preço do metro quadrado em ordem crescente, por exemplo: É comum queremos aplicar uma função qualquer aos dados, ou à parte deles, neste caso o pandas fornece o método .apply. Por exemplo, para deixar os nomes dos bairros como apenas as suas três primeiras letras: Ou de um jeito mais prático, usando uma função lambda: Uma das tarefas na qual o pandas é reconhecidamente poderoso é a habilidade de tratar dados incompletos. Por muitos motivos pode haver incompletude no dataset, o np.nan é um valor especial definido no Numpy, sigla para Not a Number, o pandas preenche células sem valores em um DataFrame lido com np.nan. Vamos criar um novo dataframe usando as 5 primeiras linhas do nosso original, usando o já visto .head(). Abaixo é usado o .replace para substituir um valor específico por um NaN. O pandas simplifica a remoção de quaiquer linhas ou colunas que possuem um np.nan, por padrão o .dropna() retorna as linhas que não contém um NaN: Preencher todos os valores NaN por um outro específico também é bastante simples: Acaba sendo muitas vezes conveniente termos um método que indica quais valores de um dataframe são NaN e quais não são: Visualização de Dados Partiremos agora para visualização de dados com o pandas. Os métodos de visualização do pandas são construídos com base no matplotlib para exploração rápida dos dados. Para se ter mais liberdade no conteúdo e possibilidades de visualização se recomenda usar diretamente o matplotlib ou ainda, para visualização estatística, o seaborn. Nesta introdução tratarei apenas dos métodos de visualização incluídos no pandas, que por outro lado, oferece uma sintaxe bastante simples para realizar a tarefa. Comecemos verificando que tanto Series como DataFrame possuem um método .plot() que também é um atributo e pode ser encadeado para gerar visualização de diversos tipos, como histograma, área, pizza e dispersão, com respectivamente .hist(), .area(), .pie() e .scatter(), além de vários outros. Vamos verificar a distribuição dos preços usando o encadeamento .plot.hist(), o eixo x, que é o preço, está numa escala de *10^7, como mostrado na imagem: Por padrão esse método usa 10 bins, ou seja, divide os dados em 10 partes, mas é claro que podemos especificar um valor para a plotagem. Abaixo, além de especificar a quantidade de bins, também especifiquei a cor das bordas como preta, que por padrão é transparente. Podemos usar os valores de contagem de cada bairro como exemplo de dado para um plot tanto de barras verticais quando de barras horizontais, para verificar visualmente esses dados: Os métodos são flexíveis o suficiente para aceitarem argumentos como um título para a imagem: Um gráfico de dispersão usando um DataFrame pode ser usado especificando-se quais colunas usar como dados no eixo x e y: Para fins estéticos, o matplotlib fornece uma série de styles diferentes que podem ser usados, um deles é o ggplot Agora este estilo será usado em todas as imagens geradas após essa linha A lista de estilos disponíveis pode ser vista através de um método próprio A coluna de quartos diz quantos quartos tem um determinado apartamento, também se pode ver a contagem e distribuição usando outros métodos de plotagem oferecidos pelo pandas: Uma coisa a se notar do gráfico de scatter é a poluição causada pela enorme quantidade de dados agrupadas num dos cantos do gráfico, além de podermos diminuir o tamanho dos pontos passando o argumento s ao método .scatter podemos também usar um método do pandas que cria uma amostragem aleatória dos dados. O .sample pode receber tanto um argumento frac, que determina uma fração dos itens que o método retornará (no caso abaixo, 10%), ou n, que determina um valor absoluto de itens. Finalmente, a tarefa de salvar seu DataFrame externamente para um formato específico é feita com a mesma simplicidade que a leitura de dados é feita no pandas, pode-se usar, por exemplo, o método to_csv, e o arquivo será criado com os dados do DataFrame: Com o que foi abordado nesta introdução você já deve estar apto a fazer exploração e manipulação básica de dados com o pandas, para aprofundar mais aqui vão algumas referências: Documentação oficial Coletânea de notebooks Jupyter que abordam profundamente várias ferramentas e casos de uso do Pandas Exercícios de Pandas com soluções, separados por temas Curtiu esse post? Não deixe de compartilhar com seus amigos! Também não se esqueça de se inscrever na nossa newsletter no www.datahackers.com.br! Abraço e até a próxima! o/

Seus primeiros passos como Data Scientist: Introdução ao Pandas!

345

uma-introdução-simples-ao-pandas-1e15eea37fa1

2018-06-15

2018-06-15 07:08:19

https://medium.com/s/story/uma-introdução-simples-ao-pandas-1e15eea37fa1

false

2,510

Blog oficial da comunidade Data Hackers

null

datahackersoficial

null

Data Hackers

falecom@datahackers.com.br

data-hackers

DATA SCIENCE,MACHINE LEARNING,BIG DATA,DATA ENGINEERING,STATISTICS

null

Python

python

Python

20,142

Vinícius Aguiar

null

fba229043a07

vinciusaguiar_14084

47

23

20,181,104

null

0

null

0

null

2018-06-04

2018-06-04 22:50:36

2018-06-05

2018-06-05 01:04:47

0

false

en

2018-06-05

2018-06-05 01:06:24

8

1e1733e5fa95

3.196226

1

0

Also AI beats human on chess, and don’t you think this will send humanity to extinction because i humanity have reach there ultimate peak…

2

Trans-human and Digital Decentralization: A substitute to Artificial Intelligence Threat. Artificial Intelligence has become an Holy Grail of Technology in this 21st century. Of-course our devices and machine are becoming very smarter every day which makes life a betterment for us. we live our lives round on the Virtual environment of Technology(mostly Internet) which about 75% of the user cannot do without it. Artificial Intelligence (Machine learning, Deep learning) are now the sexiest job of the 21st century which gives rise to a lot of application deployed on internet. For Instance, let consider the ways we message each other on our smart phone, (you type “hello suzy” the next time you typed Hello, it brings a suggestion of suzy) which is a great function of machine learning… which really makes the way we communicate to be faster, easier and smarter. Considering also the Autonomous car, which uses Limited Memory type of AI and at this point AI is just at its early stage, and it has been predicted that Driverless car would saves life by 90%. which i think this is really awesome and cool. But dont you think Human driver would be rendered irrelevant and could cause loss of Job to the Taxi drivers? Another example is the Connected and Smart Homes where all you devices and gadget are very pretty much intelligence: For Instance You had an hectic and frustrating moment at office, You got home with the initial reaction to the frustration; but you home Music devices has discover there is a change in your reaction, and tries to calm you down with the music you always love to listen to which your house partner may not know what you are passing through and also cannot offers you the music. Don’t you think we are getting to a moment where our partner would be rendered irrelevant? Also AI beats human on chess, and don’t you think this will send humanity to extinction because i humanity have reach there ultimate peak of our Intelligence? Couple of months ago two Tech giant(Elon Musk and Mark Zuckerberg) argues on AI being threat to humanity and also being a problem solver. of-course both of them are correct from there own perspective. The main question is how are we going to mitigate this threat of AI, How we going to secure our Job so that AI wont steal it.? There is no point Artificial Intelligent will rendered humanity irrelevant and useless if some measures are not taken in to consideration. Humanity have to be enhanced with some smart embedded system so that we can compete with AI and AI wont become our god. this process is called Transhuman which aim is to transform the human condition by developing and making widely available sophisticated technologies to greatly enhance human intellect and physiology. this will is give us the capacity to radically outperform the best human brains in practically every field, including scientific creativity, general wisdom, and social skills This process requires planting of chip into human body for proper enhancement and to unlock our limitation. Transhumanism has various branch and approach examples are: 1 Gene Therapy/RNA Interference: this involves gene editing using advance tools like Crispr. 2 Cybernetics: Merging man with machine 3 Uploading: this is of two type (i)uploading knowledge to your brain- uploading additional skills, e.g you don’t need to go to school to learn a whole damn process of things since they were already programmed, you can just upload them to you brain since there are you have some brain enhancement. (ii) Mind upload e.g uploading your consciousness to computer. and lot more… Transhuman provides artificial limbs for the amputated and also 3-dimension exoskeleton system for the disabled, Artificial eye for the blind etc. How are we going to secure our job against AI? we cannot secure our job but we can only find a new means of surviving. over 38% of jobs in the US will be loosed to AI before 2030. Digital Decentralization of data and resources would be the next avenue for surviving, A world where you work anytime which would end the 9–5 job. A world where values would be spent. we needed to get paid based on our data instead for a person or an entity getting payment for our data e.g Facebook Cambridge Analytica data scandal. Every one should be paid based on his data, post on Facebook but because of the centralized type of data store, Facebook had an access to sell our data and make more money. But welcome to an era of Digital Decentralization (Blockchain) where every one had access to the data and can make money from there data. good example of this is the Steemit -a social media platform where everyone gets paid for creating and curating content. It leverages a robust digital points system. There are many ways by which digital decentralization will be a new means for surviving. to find out more, watch out for my book titled Rapture without Tears. I am Josiah Akinloye Egilitarian Transhumanist Founder of savycon.

Artificial Intelligence has become an Holy Grail of Technology in this 21st century.

50

artificial-intelligence-has-become-an-holy-grail-of-technology-in-this-21st-century-1e1733e5fa95

2018-06-05

2018-06-05 07:35:37

https://medium.com/s/story/artificial-intelligence-has-become-an-holy-grail-of-technology-in-this-21st-century-1e1733e5fa95

false

847

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Josiah Akinloye

am a futurist, and i have passion for new invention that improves humanity betterment . I am a Transhumanist,an Engineer, Inventor/problem solver.

77bf7ef70f89

josiahakinloye

1

4

20,181,104

null

0

null

0

null

2017-10-25

2017-10-25 20:10:25

2017-10-25

2017-10-25 21:36:28

5

false

en

2017-10-25

2017-10-25 21:53:22

4

1e1890a584b1

5.108805

10

0

The recent Mini-Symposium on Deep Generative Models and Unsupervised Machine Learning conducted by Amsterdam Data Science was quite…

3

Deep Generative Models & Unsupervised Machine Learning Time-contrastive learning (TCL) The recent Mini-Symposium on Deep Generative Models and Unsupervised Machine Learning conducted by Amsterdam Data Science was quite enlightening. Unsupervised deep learning remains a largely unsolved problem. There were three talks including one by the famous Prof Joshua Bengio. The first session was a very interesting one by Aapo Hyvarinen from University College London titled: Towards nonlinear independent component analysis in which the idea was to extend ICA to the non-linear case to get deep learning. Two methods of Temporal Contrastive Learning was utilized (in simple terms split a time series into segments and train a neural network on one segment as input and any random further segment in time as the output). The learning objective is to understand the differences between the segments of the time series. In another method based on randomly time permuted data the hidden layers are made to learn temporal dependency. Abstract: Unsupervised learning, in particular learning general nonlinear representations, is one of the deepest problems in machine learning. Estimating latent quantities in a generative model provides a principled framework, and has been successfully used in the linear case, e.g. with independent component analysis (ICA) and sparse coding. However, extending ICA to the nonlinear case has proven to be extremely difficult: A straight-forward extension is unidentifiable, i.e. it is not possible to recover those latent components that actually generated the data. Here, we show that this problem can be solved by using temporal structure. We formulate two generative models in which the data is an arbitrary but invertible nonlinear transformation of time series (components) which are statistically independent of each other. Drawing from the theory of linear ICA, we formulate two distinct classes of temporal structure of the components which enable identification, i.e. recovery of the original independent components. We show that in both cases, the actual learning can be performed by ordinary neural network training where only the input is defined in an unconventional manner, making software implementations trivial. We can rigorously prove that after such training, the units in the last hidden layer will give the original independent components. [With Hiroshi Morioka, published at NIPS2016 and AISTATS2017.] This was followed by a talk by Prof Bengio — a complicated one on Abstract Representations and deep understanding. This involves mapping the real world state to controllable factors of variation which are disentangled higher level abstractions. He states the main problem with unsupervised training objective is the operations in pixel space and not in abstract space. He introduced the concept of Conscious Prior. The idea cmes from conscious thoughts which are low dimensional objects compared to the full state of the unconscious brain. This will have an unexpected predictive value or usefulness a strong constraint or prior on the underlying representation. The key question is select the few relevant abstract concepts. This is based on the concept of attention where the focus is on a few elements from a large set. Prof Bengio postulates Cognitive Priors Title: From Deep Learning of Disentangled Representations to Higher-Level Cognition Abstract: One of main challenges for AI remains unsupervised learning, at which humans are much better than machines. We review recent work in deep generative models and propose research directions towards learning of high-level abstractions. This follows the ambitious objective of disentangling the underlying causal factors explaining the observed data. We argue that in order to efficiently capture these, a learning agent can acquire information by acting in the world, moving our research from traditional deep generative models of given datasets to that of autonomous learning or unsupervised reinforcement learning. We propose two priors which could be used by an agent acting in its environment in order to help discover such high-level disentangled representations of abstract concepts. The first one is based on the discovery of independently controllable factors, i.e., in jointly learning policies and representations, such that each of these policies can independently control one aspect of the world (a factor of interest) computed by the representation while keeping the other uncontrolled aspects mostly untouched. The second prior is called the consciousness prior and is based on the observation that our conscious thoughts are low-dimensional objects with a strong predictive or explanatory power (or are very useful for planning). A conscious thought thus selects a few abstract factors (using the attention mechanism which brings these variables to consciousness) and combines them to make a useful statement or prediction. In addition, the concepts brought to consciousness often correspond to words or short phrases and the thought itself can be transformed into a brief linguistic expression, like a sentence. Natural language could thus be used as an additional hint about the abstract representations and disentangled factors which humans have discovered to explain their world. A conscious thought also corresponds to the kind of small nugget of knowledge (like a fact or a rule) which has been the main building block of classical symbolic AI. This therefore raises the interesting possibility of addressing some of the objectives of classical symbolic AI using the deep learning machinery augmented by the architectural elements necessary to implement conscious thinking. Prof Bengio answering questions, Prof Blei in the background Last up was the talk on probabilistic models by Prod David Blei of whom I had first heard regarding LDA — Latent Dirichlet Allocation (a form of generative topic modelling in the context of documents). His specific topic was Bayesian Deep Learning related. Probabilistic ML methods connect domain knowledge to data. Variational Inference is the underlying subject and the main algorithmic issue is posterior inference or posterior estimation/prediction in Bayesian Learning. An interesting library was mentioned: Edward. Prof David Blei on the podium Title: Black Box Variational Inference and Deep Exponential Families Abstract: Bayesian statistics and expressive probabilistic modeling have become key tools for the modern statistician. They let us express complex assumptions about the hidden elements that underlie our data, and they have been successfully applied in numerous fields. The central computational problem in Bayesian statistics is posterior inference, the problem of approximating the conditional distribution of the hidden variables given the observations. Approximate posterior inference algorithms have revolutionized the field, revealing its potential as a usable and general-purpose language for data analysis. In this talk, I will discuss two related innovations in modeling and inference: deep exponential families and black box variational inference. Deep exponential families (DEFs) adapt the main ideas behind deep learning to expressive probabilistic models. DEFs provide principled probabilistic models that can uncover layers of representations of high-dimensional data. I will show how to use DEFs to analyze text, recommendation data, and electronic health records. Deep exponential families (DEFs) adapt the main ideas behind deep learning to expressive probabilistic models I will then discuss the key algorithm that enables DEFs: Black box variational inference (BBVI). BBVI is a generic and scalable algorithm for approximating the posterior. BBVI easily applies to many models, with little model-specific derivation and few restrictions on their properties.

Deep Generative Models & Unsupervised Machine Learning

63

deep-generative-models-unsupervised-machine-learning-1e1890a584b1

2018-06-20

2018-06-20 13:37:46

https://medium.com/s/story/deep-generative-models-unsupervised-machine-learning-1e1890a584b1

false

1,133

null

Machine Learning

machine-learning

Machine Learning

51,320

SHAMIT BAGCHI

Complexity | #Computing #Science #Music #Art #Creativity | Free spirited views are my own ..

469bd6ad03d8

shamitb

186

444

20,181,104

null

0

null

0

null

2017-10-18

2017-10-18 10:29:55

2017-10-18

2017-10-18 10:31:59

0

false

th

2017-10-18

2017-10-18 10:31:59

1

1e189bc0932

0.879245

1

0

ระบบ Machine Learning แนวคิดพื้นฐานการทำ Website ให้คิดเองได้

5

ระบบ Machine Learning แนวคิดพื้นฐานการทำ Website ให้คิดเองได้ ระบบ Machine Learning แนวคิดพื้นฐานการทำ Website ให้คิดเองได้ ระบบ Machine Learning แนวคิดพื้นฐานการทำ Website ให้คิดเองได้ Machine Learning ในทางการพัฒนาซอฟต์แวร์หมายถึงชุดคำสั่งหรืออัลกอริทึมที่สามารถเรียนรู้ข้อมูลและประมวลผล ความน่าจะเป็นของผลลัพธ์ที่ใกล้เคียงที่สุด ตามปริมาณข้อมูลที่รับเข้ามา ความแม่นยำของการคาดเดาข้อมูลจะขึ้นอยู่กับความฉลาดของชุดคำสั่งหรืออัลกอริทึมที่พัฒนาขึ้นมา Machine Learning หรือกลไกการเรียนรู้ด้วยตัวเองของชุดคำสั่งหรืออัลกอริทึม เป็นสาขาหนึ่งของการพัฒนาปัญญาประดิษฐ์ (Artificial Intelligence) หรือ เอไอ (AI) ปัจจุบันมีการนำระบบ machine learning มาใช้กับการพัฒนาซอฟต์แวร์ทั้ง website และ mobile application อย่างแพร่หลาย ระบบ Machine Learning ใกล้ตัวเราที่พอจะนึกออก Machines can see. มองเห็นเอง ตัวอย่างเช่น รถยนต์ไร้คนขับ ของ Tesla Motors Machines can read. อ่านได้ ตัวอย่างเช่น รู้ว่าเรารีวิวหรือคอมเม้นสินค้าในทางที่ดีหรือไม่ดี Machines can listen. ฟังได้ ตัวอย่างเช่น Siri, Cortana หรือ Google Now Machines can talk. พูดหรือแนะนำได้ เช่น Siri, Google map navigator Machines can write. เขียนได้ ปัจจุบันมีเว็บไซต์ที่ให้เราอัพโหลดรูปวาดขึ้นไประบบจะระบายสีให้อัตโนมัติ เว็บขายสินค้าออนไลน์กับ Machine learning (ข้อมูลด้านล่างเป็นการตั้งสมมุติฐานและยกตัวอย่างขึ้นมาเท่านั้น ไม่มีเว็บไซต์หรือสถิติข้อมูลที่เป็นจริง) ตัวอย่างข้อมูลเว็บไซต์ เว็บไซต์ jipatashop.com (จิปาถะช๊อป.com) มีสมาชิกทั้งหมด 50,000 คน มีปริมาณการเข้าชมเฉลี่ยต่อวัน 1,000 คน มีสินค้าจำหน่ายออนไลน์มากกว่า 1,500 รายการ โจทย์ในการพัฒนาระบบ Machine Learning แนะนำสินค้าให้กับสมาชิกใหม่ จากข้อมูลพื้นฐาน เว็บไซต์มีสินค้าหลากหลายประเภท ต้องการแนะนำสินค้าที่มีเกี่ยวข้องกันและมีโอกาสเกิดการซื้อร่วมกัน วิเคราะห์พฤษติกรรมของผู้ซื้อ และแนวโน้มสินค้าที่น่าจะซื้อต่อ เมื่อมีการสั่งซื้อสินค้ารายการใดรายการหนึ่ง โจทย์ในการพัฒนาระบบ แนะนำสินค้าให้กับสมาชิกใหม่ จากข้อมูลพื้นฐาน ช่วงอายุ เพศ วันเกิด เว็บไซต์มีสินค้าหลากหลายประเภท ต้องการแนะนำสินค้าที่มีเกี่ยวข้องกัน วิเคราะห์พฤษติกรรมของผู้ซื้อ และแนวโน้มสินค้าที่น่าจะซื้อต่อ เมื่อมีการสั่งซื้อสินค้ารายการใดรายการหนึ่ง ตัวอย่างการคิดอัลกอริทึม 1. หาสินค้าที่สมาชิกเข้าชมบ่อยที่สุดหรือสินค้าที่เคยซื้อ ในช่วง 1–2 สัปดาห์ที่ผ่าน 2. นำข้อมูลสินค้าจากข้อ 1 ไปค้นหาสมาชิกคนอื่น ที่เคยเข้าดูสินค้าหรือเคยซื้อ เหมือนกับสมาชิกในข้อ 1 จำนวน 10 คน มีเพศเดียวกัน อายุเฉลี่ยห่างกันไม่เกิน 5 ปี ตัวอย่างเช่น สมาชิกคนที่ 1 อายุ 25 สมาชิกที่อยู่ในเกณฑ์ต้องมีอายุ 20–30 ปี และเป็นเพศเดียวกัน 3. นำสมาชิกทั้ง 10 คนจากข้อ 2 ไปค้นหาว่า สินค้าอื่น ๆ ที่สมาชิกทั้ง 10 คนเคยเข้าชมหรือเคยซื้อ 10 อันดับแรกคือสินค้าอะไรบ้าง 4. จากข้อ 3 เราจะได้สินค้าทั้งหมดคือ 10×10 = 100 ชิ้น นำสินค้าทั้ง 100 มาคิดคะแนนลำดับ สินค้าชิ้นไหนซ้ำกันมากที่สุด หรือมีการเข้าชมมากที่สุด ให้มีคะแนนแนะนำสูงที่สุด เมื่อประมวลผลแล้วอาจจะมีสินค้าแนะนำ และมีแนวโน้มที่สมาชิกในข้อ 1 จะซื้อเหลืออยู่ไม่ถึง 100 ชิ้น 5. จากข้อ 1–4 แน่นอนว่าเมื่อเว็บไซต์มีสมาชิกมากขึ้น มีการซื้อสินค้ามากขึ้น มี attribute ใหม่ ๆ ที่สามารถนำมาคำนวณได้มากขึ้น ความแม่นยำก็จะมากขึ้นตามไปด้วย ตัวอย่างข้อมูลที่จำเป็นสำหรับระบบ ข้อมูลพื้นฐาน ไอพี อายุ เพศ วันเกิด หมวดหมู่ ประเภท สินค้า เรทราคาที่เข้าชมบ่อย คำหลัก แท็ก คีเวิร์ดที่ใช้ค้นหาสินค้า สินค้าที่กดติดตาม คอมเมนต์ หรือแชร์ และสถิติข้อมูลอีกเยอะแยะมากมายที่จะช่วยให้กลไกการคำนวณแม่นยำมากขึ้น จะเห็นได้ว่า จากการตั้งสมมุติฐานและยกตัวอย่างอัลกอริทึมด้านบนนั้น การทำระบบ machine learning จะเป็นการนำข้อมูลสถิติมาคิดคำนวณหาผลลัพธ์ที่มีความน่าจะเป็นที่สุด โดยใช้พฤษติกรรมของคนหรือสมาชิกเว็บไซต์ ยิ่งสมาชิกในเว็บไซต์มี activity log มากเท่าไหร่ ข้อมูลก็จะยิ่งมีปริมาณที่สูงขึ้นไปตามลำดับ ยิ่งมีปริมาณข้อมูลสูงขึ้นมากเท่าไหร่ ความแม่นยำในการคาดเดาก็จะยิ่งแม่นยำมากขึ้นเท่านั้น เพิ่มเติม : https://www.codebee.co.th/labs/ระบบ-machine-learning

ระบบ Machine Learning แนวคิดพื้นฐานการทำ Website ให้คิดเองได้

2

ระบบ-machine-learning-แนวคิดพื้นฐานการทำ-website-ให้คิดเองได้-1e189bc0932

2018-03-26

2018-03-26 11:05:42

https://medium.com/s/story/ระบบ-machine-learning-แนวคิดพื้นฐานการทำ-website-ให้คิดเองได้-1e189bc0932

false

233

null

Machine Learning

machine-learning

Machine Learning

51,320

บริษัท โค๊ดบี จํากัด

null

c4ccaa3cbf61

codebee2014

15

1

20,181,104

null

0

null

0

null

2018-08-10

2018-08-10 13:52:25

2018-08-11

2018-08-11 12:44:06

5

false

ru

2018-08-11

2018-08-11 12:44:06

0

1e1c97e1362c

2.523899

0

Как научить компьютер отличать кошку от собаки? Как благодаря машинному обучению старые телефону спасают леса Амазонки от пожаров? Как…

5

Как машинное обучение и AI меняют мир вокруг: эксклюзивная лекция управляющего директора Google Россия Дмитрия Кузнецова Как научить компьютер отличать кошку от собаки? Как благодаря машинному обучению старые телефону спасают леса Амазонки от пожаров? Как можно сократить издержки благодаря интеграции искусственного интеллекта в бизнес-процессы и о чём общаются между собой машины? Ответы на эти вопросы узнали гости эксклюзивной лекции управляющего директора Google Россия Дмитрия Кузнецова, которая прошла в коворкинге «Ясная поляна» 8 августа. Еще несколько лет назад искусственный интеллект фигурировал в основном в сюжетах фантастических фильмов, а сегодня мы сталкиваемся с успешными кейсами машинного обучения на каждом шагу: наши смартфоны анализируют всю информацию, которую получают от нас, чтобы сделать поиск более удобным и адаптированным под наши интересы. Нам показывается реклама тех товаров, которые нам нужны, тексты переводятся с учетом особенностей контекста нашего общения — разным людям система выдаст перевод с техническим уклоном или литературный. Хотим мы того или нет, искусственный интеллект влияет на нашу жизнь каждый день, и, по мнению Дмитрия Кузнецова, в ближайшем будущем он принесет еще больше пользы. Анализ большого массива данных используется не только в новых продуктах компании, но и в самых привычных, позволяя нам избавиться от рутины как в повседневной жизни, так и в работе. “Мы планируем активно развивать искусственный интеллект и технологии машинного обучения во всех продуктах компании Google, и в ближайшие несколько лет это может принести большую пользу всему человечеству. В первую очередь мы делаем ставку на развитие собственных продуктов, но они, безусловно, будут востребованы и для бизнеса, помогая сократить ненужные операции и оптимизировать процессы”. В то же время, помимо оптимистичных прогнозов о развитии искусственного разума, все чаще звучат вопросы, насколько этот процесс находится под контролем и не воплотятся ли в жизнь апокалиптические сценарии восстания машин. Дмитрий подтвердил, что морально-этическая сторона развития искусственного разума вызывает дискуссии в профессиональном сообществе: “Мы уже выясняли, что некоторые машины в процессе самообучения начинают взаимодействовать друг с другом, но вот о чем они общаются, мы пока не знаем. Но на данный момент машины находятся под контролем и обучаются для того, чтобы выполнять те задачи, для которых они были созданы. Мы создаем машины для того, чтобы помогать людям, а не замещать их”. По словам директора Google, опасаться искусственного разума не стоит: все данные принадлежат пользователям и они вправе решить, какие из них доступны для системы, а какие должны быть скрыты ото всех. Компания использует только те, которыми вы готовы поделиться. “Но после того, как вы несколько раз воспользуетесь сервисами, которые упростят вам жизнь, вы привыкнете к этому уровню комфорта и вряд ли сможете от них отказаться”. Действительно, сегодня уже трудно отказаться от высоких технологий ради сохранения личных данных. Поэтому остается надеяться, что искусственный разум будет учиться только хорошему и всегда даст правильный ответ на “Окей, гугл”.

Как машинное обучение и AI меняют мир вокруг: эксклюзивная лекция управляющего директора Google…

0

как-машинное-обучение-и-ai-меняют-мир-вокруг-эксклюзивная-лекция-управляющего-директора-google-1e1c97e1362c

2018-08-11

2018-08-11 12:44:06

https://medium.com/s/story/как-машинное-обучение-и-ai-меняют-мир-вокруг-эксклюзивная-лекция-управляющего-директора-google-1e1c97e1362c

false

448

null

Google

google

Google

35,754

Oksana Olshevskaya

null

a618f1b848b5

Oksana_O

43

48

20,181,104

null

0

null

0

null

2018-02-09

2018-02-09 01:50:26

2018-02-09

2018-02-09 02:01:50

2

false

en

2018-02-09

2018-02-09 02:01:50

3

1e1d93cbdafa

1.915409

1

0

If you’re learning or getting stuck, these may help you out

4

2 Cheatsheets For Tableau If you’re learning or getting stuck, these may help you out For the past 3 years I’ve been playing around with tableau, and the last 2 years I’ve spent a lot of time working with the software. I’ve learned how to do quite a few different things and have learned a lot but I still get stuck frequently and need some help. Google is of course your best resource, you can simply write something like “Tableau how to build a time series” and you’ll get lots of great results, primarily you’ll find lots of community.tableau.com resources with people asking specific questions which you can interpret and then apply to your problem. Often times it’s exactly what you need and it’s a quick fix to the problem you are solving. Other times it gets more complex than that, and if you don’t want to wait for help from tableau support or they weren’t able to fully help you you’ll have to reach out further to the interwebs. The following sites are pretty great and convenient, they are really just reference guides with links to other articles and resources that will help you figure things out. http://www.dataplusscience.com/TableauReferenceGuide/index.html This site has TONs of awesome resources. For example, you might see in that screenshot that there is a whole section for calculations, which can be very handy if you need to find something similar quickly. There’s also a search bar so you can easily find what you are looking for to solve your problems. http://drawingwithnumbers.artisart.org/wiki/tableau/tricks-miscellaneous-techniques/ This is a similar resource but a little more of a blog format, there is quite a list of tricks and techniques you can use. Plus I keep this one handy with links to the Tableau Workbook Library and the Calculation Reference Library. Both extremely helpful places to search for help. The last one I’ll throw out that is quite helpful is lynda.com. They have 1000s of courses on tons of different stuff but they have a nice stack of tableau courses that you can use. I haven’t gone through all of them but I’ve used parts of the courses and videos to get up to speed on something. It’s nice to have a video walk through sometimes. There you have it! Happy data / viz building! Sign Up | ChaseCottle.com Fresh published content daily on medium.com. I've been writing every day for more than 600 days in a row. Get a curated…chasecottle.com

2 Cheatsheets For Tableau

1

2-cheatsheets-for-tableau-1e1d93cbdafa

2018-02-11

2018-02-11 16:45:01

https://medium.com/s/story/2-cheatsheets-for-tableau-1e1d93cbdafa

false

406

null

Tableau

tableau

Tableau

725

Chase Cottle

Co-founder. CTO. Entrepreneur. Love marketing, data science, and tech. Free time: snowboarding, mountain biking, rock climbing and any other adrenaline activity

d2891aa0e121

chasecottle

170

123

20,181,104

null

0

null

0

null

2018-09-30

2018-09-30 22:23:33

2018-09-30

2018-09-30 22:24:35

0

false

en

2018-09-30

2018-09-30 22:24:35

4

1e1de77d6c58

2.449057

0

As co-founder and CEO of Instreamatic.ai, Stas Tushinskiy lives and works on the cutting edge of voice-activated advertising, with the…

3

Inside Radio: Q&A: Stas Tushinskiy, CEO, Instreamatic.ai As co-founder and CEO of Instreamatic.ai, Stas Tushinskiy lives and works on the cutting edge of voice-activated advertising, with the intent of changing the fundamentals for advertising and marketing companies — as well as media platforms that rely on ads for revenue. The San Francisco-based company started as a digital audio ad network, and has grown in four short years to build a monetization platform for the voice era. Now, Instreamatic.ai provides audio publishers with tools to monetize their ad inventory with voice-activated ads in which the audio stream asks listeners if they are interested in a particular product or service and then reacts to their response. This year, Instreamatic.ai was chosen as one of 50 interactive technology companies from 10 categories, awarded an Accelerator Award by SXSW in the Entertainment and Content Technology category. So far, its platform works with mobile; it will soon be present on smart speakers. Tushinskiy — who shared an “An Open Letter To The Industry” with Inside Radio in March — now breaks down just how the technology works for both advertisers and consumers, and how radio stations can get with the program. He also shares with Inside Radio readers an exclusive announcement about a “coming soon” service right up radio’s alley. An edited transcript follows. Instreamatic.ai started as a digital audio ad network. How did you make the leap to voice-activation audio ads? Four years ago we were doing business development for one of the top music streaming apps in Europe. Back then we learned that a paid subscription business model was not going to get us to a break-even point. So it was a natural choice to adopt an advertising-based model. If you look at the industry today, the leading companies are all offering free ad-supported models, even if they offer paid subscription models in tandem. Eventually, we founded Instreamatic to build a digital audio ad market in Eastern Europe and address technology challenges. As we learned along the way, the engagement challenge was and still is the major problem that audio marketing has to solve to get to a place where video and social marketing is today. In layman’s terms, how does your technology work? Imagine you are driving down the street or jogging, listening to a radio stream or a podcast and you hear a short audio ad that goes like… “Do you want to know where the best coffee shop in town is located?” You can respond to our voice-activated ad by saying “Yes, tell me more.” We’ll provide all the details with a second audio piece and then build a route on Google Maps. Or you can say “I’m not interested” and we’ll send you right back to content. These ads make the ad experience shorter, interactive and a lot more fun. For advertisers, this means the ability to measure direct performance and receive a higher ROI. And publishers earn higher CPMs, fill-rates and new categories of advertisers that were not traditional users of audio/radio ads. You’ve called voice the third big wave of the internet evolution, behind browser and mobile. There were two major waves of interface disruption during the short history of the internet. Each one brought new winners and losers. Browsers made the internet accessible to the masses. Companies like Amazon, Google and many others built their success on this breakthrough. The next wave was mobile devices and mobile interfaces. Top gainers were Apple and Facebook — even though Facebook started earlier, their huge success came with mobile adoption and with the purchases of Instagram and What’s App. Now we’re entering the voice era. For now, Amazon and Google are the ones who are driving the change, but we’ll see many startups riding the wave and becoming multi-billion dollar companies through this disruption. Read full interview

Inside Radio: Q&A: Stas Tushinskiy, CEO, Instreamatic.ai

0

inside-radio-q-a-stas-tushinskiy-ceo-instreamatic-ai-1e1de77d6c58

2018-09-30

2018-09-30 22:24:35

https://medium.com/s/story/inside-radio-q-a-stas-tushinskiy-ceo-instreamatic-ai-1e1de77d6c58

false

649

null

Advertising

advertising

Advertising

42,520

Instreamatic.ai

An AI-powered platform to manage, measure and monetize voice-activated advertising. https://instreamatic.ai

af284d8cde17

instreamatic

2

1

20,181,104

null

0

a = [2,1,5,3,0.1,0.5,0.2,1] b = [15,3,3,2,0,0,0.5,1] ## Manhattan Distance manhattan = 0 for i in a: for j in b: manhattan += abs(i-j) print("Manhattan Distance = " + str(manhattan)) ## Euclidean distance euclidean = 0 for i in a: for j in b: euclidean += (i-j)**2 print("Euclidean Distance = " + str(euclidean**0.5)) ## Minkowski Distance def Minkowski(a,b,p): base = sum([abs(i-j)**p for i in a for j in b]) return(base**(1/p)) print("Minkowski Distance(p=3) : " + str(Minkowski(a,b,3))) print("Minkowski Distance(p=4) : " + str(Minkowski(a,b,4))) print("Minkowski Distance(p=5) : " + str(Minkowski(a,b,5))) print("Minkowski Distance(p=100) : " + str(Minkowski(a,b,100))) from scipy.cluster import hierarchy from sklearn.datasets import make_blobs from sklearn.preprocessing import StandardScaler import matplotlib.pyplot as plt %matplotlib inline from sklearn.cluster import DBSCAN # Generate sample data centers = [[1, 1], [-1, -1], [1, -1]] X, labels_true = make_blobs(n_samples=450, centers=centers, cluster_std=0.4,random_state=0) X = StandardScaler().fit_transform(X) # -----Perform DBSCAN clustering with epsilon=0.1 cluster_assignments = DBSCAN(eps=0.1, min_samples=10,metric = 'minkowski', p = 2).fit_predict(X) plt.figure(figsize=(10,4)) plt.subplot(1,3,1);plt.title("DBSCAN Clustering (epsilon=0.1)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with epsilon=0.3 cluster_assignments = DBSCAN(eps=0.3, min_samples=10, metric = 'minkowski', p = 2).fit_predict(X) plt.subplot(1,3,2);plt.title("DBSCAN Clustering (epsilon=0.3)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with epsilon=0.6 cluster_assignments = DBSCAN(eps=0.6, min_samples=10, metric = 'minkowski', p = 2).fit_predict(X) plt.subplot(1,3,3); plt.title("DBSCAN Clustering (epsilon=0.6)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with Manhattan Distance cluster_assignments = DBSCAN(eps=0.3, min_samples=10,metric = 'minkowski', p = 2).fit_predict(X) plt.figure(figsize=(10,4)) plt.subplot(1,3,1);plt.title("DBSCAN Clustering (Manhatten Distance)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with Euclidean Distance cluster_assignments = DBSCAN(eps=0.3, min_samples=10, metric = 'minkowski', p = 3).fit_predict(X) plt.subplot(1,3,2);plt.title("DBSCAN Clustering (Euclidean Distance)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with Minkowski Distance cluster_assignments = DBSCAN(eps=0.3, min_samples=10, metric = 'minkowski', p = 1000).fit_predict(X) plt.subplot(1,3,3); plt.title("DBSCAN Clustering (Minkowski Distance, p=1000)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with 5 samples form a cluster cluster_assignments = DBSCAN(eps=0.3, min_samples=5).fit_predict(X) plt.figure(figsize=(10,4)) plt.subplot(1,3,1);plt.title("DBSCAN Clustering (min_points = 5)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with 20 samples from a cluster cluster_assignments = DBSCAN(eps=0.3, min_samples=20).fit_predict(X) plt.subplot(1,3,2);plt.title("DBSCAN Clustering (min_points = 20)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform DBSCAN clustering with 100 samples form a cluster cluster_assignments = DBSCAN(eps=0.3, min_samples=100).fit_predict(X) plt.subplot(1,3,3); plt.title("DBSCAN Clustering (min_points = 100)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform agglomerative clustering with average linkage Z = hierarchy.linkage(X, 'average') cluster_assignments = hierarchy.fcluster(Z, 3, 'maxclust') plt.figure(figsize=(10,4)) plt.subplot(1,3,1);plt.title("Average Link Agglomerative Clustering", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform agglomerative clustering with complete linkage Z = hierarchy.linkage(X, 'complete') cluster_assignments = hierarchy.fcluster(Z, 3, 'maxclust') plt.subplot(1,3,2);plt.title("Complete Link Agglomerative Clustering", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform agglomerative clustering with single linkage Z = hierarchy.linkage(X, 'single'); cluster_assignments = hierarchy.fcluster(Z, 3, 'maxclust') plt.subplot(1,3,3); plt.title("Single Link Agglomerative Clustering", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') plt.tight_layout() # -----Perform Kmeans clustering with k=3 cluster_assignments = KMeans(3).fit(X).predict(X) plt.figure(figsize=(10,4)) plt.subplot(1,3,1);plt.title("KMeans Clustering (k=3)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform Kmeans clustering with k=5 cluster_assignments = KMeans(5).fit(X).predict(X) plt.subplot(1,3,2);plt.title("KMeans Clustering (k=5)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # -----Perform Kmeans clustering with k=7 cluster_assignments = KMeans(7).fit(X).predict(X) plt.subplot(1,3,3); plt.title("KMeans Clustering (k=7)", fontsize='small') plt.scatter(X[:, 0], X[:, 1], marker='o', c=cluster_assignments,s=25, edgecolor='k') # import numpy as np import matplotlib.pyplot as plt from sklearn.metrics import silhouette_score %matplotlib inline k_tests = [2,3,4,5,6] wcss = []; silo = [] for k in k_tests: clustering = KMeans(k).fit(X) wcss.append(clustering.inertia_) silo.append(silhouette_score(X,clustering.predict(X))) plt.figure(figsize=(10,4)) plt.subplot(1,2,1);plt.title("Within-Cluster-Sum-of-Squares", fontsize='small') plt.scatter(k_tests, wcss, marker='o', edgecolor='k') plt.plot(k_tests, wcss, linestyle='--') plt.subplot(1,2,2);plt.title("Silouette Width", fontsize='small') plt.scatter(k_tests, silo, marker='o', edgecolor='k') plt.plot(k_tests, silo, linestyle='--')

9

null

2018-03-06

2018-03-06 19:02:40

2018-03-06

2018-03-06 21:32:02

16

false

en

2018-03-16

2018-03-16 13:06:29

0

1e1e4f95b501

15.908491

9

0

In terms of unsupervised learning methods, some of the most well researched and common methods can be grouped under clustering. The basic…

5

A Comprehensive Introduction to Clustering Methods In terms of unsupervised learning methods, some of the most well researched and common methods can be grouped under clustering. The basic idea is simple. If you can figure out how to define distances between data points, then data points that are closer together may exhibit some kind of group characteristic we could exploit for modeling or extract new understanding from. Some examples include patients with similar blood test results that have the same disease, consumers with a similar purchase history that are part of the same socioeconomic class or occupation, and flowers with similar colors and petal lengths that are part of the same species. Obviously, some of these problems can be solved as classification problems, but this is only possible if the labels are available. Clustering, as with other unsupervised methods, operate without a label of interest. We will cover the following topics in clustering: > Distance Metrics for Real Numbers > Assessing Clustering Tendency in Data > DBSCAN Clustering > Agglomerative Clustering > K-Means Clustering > Extensions and Mixed Data Types > Choosing the # of Clusters Distance Metrics for Real Numbers Given two data points a and b, we need to find a way to define a distance between them. There are many common distance metrics for vectors of real numbers. The most common one is of course the Euclidean distance, which simply considers the points a and b as the furthest vertices of a right triangle and takes the hypotenuse of the triangle as the distance. This is how most people who are not familiar with math would define distance on a map, often called “as the crow fries”. There are different pros and cons of using Euclidean distance as a metric. On the positive side, most optimization methods are designed with Euclidean distance in mind and the computational costs can be well constrained. Euclidean distance is graphically straightforward and well understood by most people. On the negative side, the fact that we’re squaring distances significantly amplifies the effect of outliers. For two vectors a,b, the vectorized Euclidean Distance is found by taking the difference of pairs of coordinates of a,b, squaring the results, summing them, and taking the square root. Consider now if we’re trying to find the distance we’d have to walk to get between two buildings in Manhattan that are several blocks apart. Since the sidewalks are parallel we can’t simply cut diagonally from one building to another, so the Euclidean Distance wouldn’t give us a realistic estimate for the distance. We have to imagine traversing along a grid where the two points are on vertices. This is equivalent to putting the two points on diagonal vertices of a square and taking the distance to be half the perimeter. This metric goes by many names including absolute value, Manhattan, and Taxicab distance. It is found by taking the absolute value between pairs of coordinates in a,b and summing them. Optimization algorithms in Manhattan distance are often more computationally expensive and complex but are significantly more resistant to outliers. There’s several more related distances we can define, but the question may naturally arise: how many distance metrics are there, and how do you come up with a new one? Well, mathematicians have a set of rules that define the properties of a valid distance metric. Given a set of objects, any function that operates on two objects and returns a single value and also satisfies the following properties can be considered a distance metric. Distance Metric: Given elements a,b,c in a set, a distance metric is defined as a function with the following properties: 1) Non-negativity — d(a,b)≥0 2) Indiscernibility — d(a,b)=0 ⟺ a=b 3) Symmetry — d(a,b)=d(b,a) 4) Triange Inequality d(a,c)≤d(a,b)+d(b,c) for all a,b,c∈S These properties define some natural consequences we would expect from a measure of ‘distance’. (1) makes sure the distance can never be less then zero. (2) makes sure that the distance is only zero when the two elements being compared are equal. (3) makes sure that it doesn’t matter which point we start at to measure distance between two elements. (4) makes sure that the shortest distance between two elements is going directly from one to the other. Anything satisfying the following easily fits into existing frameworks for clustering and is often already implemented in a software package. Now, equipped with two examples of distance and the general definition of a distance metric, we will give a very general and powerful distance metric of which the Euclidean and Manhattan distances are special cases (p=2 and p=1 respectively), the Minkowski distance. Note that the Minkowski distance is only a distance metric for p≥1 (try to figure out which property is violated). The following is the definition of the Minkowski distance along with an illustration showing how the plane is warped with different values of p: Let’s code these distance metrics in Python and see how the distances differ between two sample vectors: Assessing Clustering Tendency of Data Before we begin studying algorithms to perform clustering, we’d like to know how to figure out how susceptible our data is to clustering methods. Is our data relatively homogeneous already or will we be able to find meaningful seperations? There are three mains ways to answer this question, and ideally all of them should be used. 1) Domain-knowledge approaches Sometimes the nature of the data and collection methods can inform us about the need for clustering. For example, we might have crime data from districts with varying socioeconomic classes without a corresponding label. We might have ER patients from a very large hospital system and would like to examine different severity patients for different kinds of models. We might suspect we have sensor data from a large variety of vehicles and should try to separate on vehicle type. There are also similar situations in which clustering can be used to reduce bias in data by separating out sub-populations that are more or less representative of the entire population. 2) Summary statistics and plots When our data is relatively clean and low-dimensional, looking at a table of summary statistics or some scatter plots can usually reveal how good clustering would be on the data. Look for things like large ‘clumps’ of points in scatter plots between features, large variances, large differences between median and mean, properties of data between quantiles etc. 3) Hopkins Statistic One way to think about clustering tendency is to ask ourselves what a data set without any clustering tendency would look like. Well, it would look like data from a uniform distribution. Whatever metric we decide on should be minimized if the data is indeed drawn from such a distribution. We can test if this is the case with good ol’ hypothesis testing. Let the null hypothesis be that our data was drawn from a uniform distribution, then the more strength we have against this null hypothesis the more sure we can be that our data is indeed susceptible to clustering methods. Preparing Data for Clustering All of the major pre-processing methods we use before modeling we should also use before clustering. Standardizing our data by subtracting the mean and dividing by the standard deviation (for each feature) is very important, since features with large ranges will dominate distance metrics between points. We should also take care to having removed or imputed all the missing values in our data. Lastly, though we will not cover it in depth here, high degrees of correlation between features and highly noisy features can make it more difficult to achieve a meaningful clustering, thus methods like Principal Components are often performed on the data prior to clustering. DBSCAN Clustering One of the most intuitive and flexible clustering algorithms is called DBSCAN (density based spatial clustering of applications with noise). This algorithm essentially treats clusters as ‘dense’ regions of data and uses two parameters, min_points and epsilon. Min_points controls the minimum number of points to justify a cluster and epsilon controls how far out from a point we look to try to find more points. Some of the big advantages that DBSCAN enjoys over other clustering algorithms is its resistance to outliers, few and straightforward parameters, ability to find arbitrarily shaped clusters, and a natural extension to other distance metrics. Some of the drawbacks include setting the epsilon parameter (which can be difficult, especially in high dimensional data) and its poor performance in data with large fluctuations in density. The algorithm can be described in the following steps: 1)Pick a random point to start the process 2) Look within epsilon distance of the point to find other points, if no such points are found go back to (1) 3) When another point is found within epsilon distance, designate this a cluster and repeat (2) and (3). 4) Stop when each point has been visited Let’s implement this clustering by generating some fake data with 3 clusters, and then using the popular Python package called Sklearn to perform DBSCAN clustering with 3 different values of epsilon. DBSCAN clustering with different values of epsilon From the above we can see how much the value of epsilon affects the outcome of the clustering. The choice of epsilon depends on how closely bound we think our clusters are, its clear that choosing epsilon = 0.6 causes the epsilon neighborhood of each point to become so large that it expands into the other clusters, whereas choosing epsilon = 0.1 causes the epsilon neighborhoods to be too small to form meaningful clusters. We can also explore the effect of using different distance metrics on the clustering result. Recall that Manhattan Distance and Euclidean Distance are just special cases of the Minkowski distance (with p=1 and p=2 respectively), and that distances between vectors decrease as p increases. Compare the effect of setting too small of an epsilon neighborhood to setting a distance metric (Minkowski with p=1000) where distances are very small. DBSCAN clustering with different distance metrics Finally, we can look at the effect of messing around with the minimum number of points that constitute a cluster (the min_samples parameter in this Sklearn implementation). The way we’ve generated the points, each cluster has roughly 150 points contained in it. DBSCAN Clustering with different values of min_points (pointed needed to form cluster) Agglomerative Clustering The idea of agglomerative clustering is very simple and the algorithm itself should look very natural to anyone familiar with greedy algorithms. There are only 4 major steps to the process: 1) Describe a distance between two clusters, called the inter-cluster distance 2) Make each point its own cluster 3) Find the two clusters with the smallest distance between them, and merge them into a single cluster 4) Repeat (3) until we are down to a defined # of clusters or sufficiently minimized cluster distances Notice, at each iteration of this algorithm we need to calculate the distance between a cluster and the rest of the remaining points, this is a very expensive operation. The algorithm has a time complexity of O(n^3) and memory requirement of O(n^2), making it very difficult to use for large data sets. There is also the issue of how exactly we describe a distance between two clusters, the inter-cluster distance. There are three common approaches to this distance, given two clusters A and B we can define the inter-cluster distance as: single-link distance: The distance between the two closest points in clusters A and B complete-link distance: The distance between the two furthest points in clusters A and B group average distance: The average of distances between points in cluster A and points in cluster B Note that the inter-cluster distance is NOT the same thing as the distance metric we define between points above. For all the below examples we will assume Euclidean distance. Let’s generate sample data in 3 clusters and see how different inter-cluster distances affect the outcome. We specify the stopping criteria as 3 clusters and run agglomerative clustering with single, complete, and group average inter-cluster distances. Try to understand why each of these outcomes might have happened given the type of inter-cluster distance specified (try changing the cluster_std argument in make_blobs()). K-Means Clustering One of the most well-known clustering methods is known as K-Means. The name of this procedure comes from the necessity of specifying the number of clusters you’d like to find as K, and from the fact that the center of each cluster is the mean of the points in the cluster. There are many variations and extensions of this method such as Mini Batch K-Means and K-Mediods that we’ll briefly discuss. The basic K-Means procedure can be described as the following: 1) Decide on value for K 2) Randomly pick K points to act as cluster centers 3) Assign other data points to nearest clusters (based on distance from K cluster centers) 4) Take the mean of data points in each cluster, make this the new cluster center 5) Repeat steps (3)(4) until the desired stopping criteria (no data points change cluster, minimal distance threshold) is reached From reading through the algorithm, you may observe that two things play a large part in the outcome of this algorithm: the K data points we ‘randomly’ choose to act as cluster centers in the beginning, and the value of K we specify. We can mitigate the first problem somewhat by trying multiple different random initializations, the second problem is much harder. A general strategy is to plot a metric like % of variance explained against number of clusters and look for an ‘elbow’, a point where the relative improvement by adding another cluster is much less then before (overall variance explained heads toward 100% as # of clusters approaches # of data points). We will examine this problem in more detail soon. K-Means clustering with different values of K Extensions and Mixed Data Types Often times, our data provides us with restrictions that prevent us from using any of the algorithms above in their natural form. The two most common limitations are large data size and mixed data types. For very large data, none of the algorithms above perform especially well, however there is an extension of K-Means that provides us with a significant performance upgrade and is very usable for large data sizes called Mini Batch K-Means. This algorithm is essentially K-Means except instead of using the entire data set at the cluster assignment step, only a small random sample of the data (known as a mini batch, usually small enough to fit in RAM) is used. There is a small amount of error produced by this procedure (about 2–8% for k<20) compared to K-Means and thus number of clusters should be kept to a minimum. When we have binary or one-hot encoded variables present in our data, the distance metric in many of these algorithms does not make sense, since the feature does not have a continuous range to define a notion of closeness. In this case, a good algorithm to use is called K-Prototypes, which can handle mixed data types by providing an extension of K-Means with a different notion of distance. We will not discuss these algorithms at length but their presence should be known as they have important practical usage. Choosing the # of Clusters With the examples above we generated the data to purposefully be in roughly 3 clusters. We also only had two explanatory features so we could do a good job visually assessing how discriminating each clustering was. In reality, data can be very high-dimensional and have various clustering tendencies, making it very difficult to use plots such as the one above to determine quality or choose the number of clusters. To get a numerical understanding of how good our clusters are, first we have to try to determine a metric for a ‘good’ clustering. What are some favorable and unfavorable properties? What task are we trying to accomplish? Before we go further into common metrics for cluster quality, we must caveat that the quality of a clustering largely depends on what we’re trying to achieve. If we have a specific goal in mind (such as improving an existing model or finding low-cost patients), we should assess our clusters based on these metrics (such as classification accuracy or cluster average cost) directly instead of relying completely on statistical measures. When we are simply looking for meaningful distinctions in the data, we should examine the metrics below more closely. One of the first things most people look for when doing clustering is homogeneous cluster data, that is that observations within a cluster should look relatively similar. This is easy enough to capture mathematically. We can simply look at how far each observation in a cluster from the cluster’s mean. This value is also known as the Within-Cluster-Sum-of-Squares (WCSS). It is formally defined as follows: Most implementations of K-Means actually seek to minimize this quantity subject to some constraints. However, there are some issues with using this as a metric for performance. For example, look what happens when each observation is its own cluster, that is when C(xi)=xi. The WCSS becomes 0. So directly minimizing this quantity is not the way to go. In reality, the way to use this metric is to plot it against the number of clusters and look for an ‘elbow’ — a point where the marginal decrease in WCSS from adding another cluster is much smaller then before. Having similar observations within a cluster is great. However, we can generate a set of very similar observations and split them into N clusters and this property would hold. We in fact need observations within a cluster to look similar AND observations between clusters to look different. Another way to say this is given an observation in a cluster, we’d like it to be close to other observations in its own cluster but far from observations in the nearest cluster over. The silhouette coefficient captures this notion. The silhouette coefficient SC(X,C) is bounded between -1 and 1. High values of the silhouette coefficient indicate dense and well separated clusters but because of dependencies on cluster convexity it is a poor choice to compare between different clustering methods. Let’s take a look at how values of these two metrics look at our generated data with 3 clusters. They both point to the ‘correct’ number of clusters, three. Obviously, in real data we will not have knowledge about existence of clusters and these plots will not look nearly as nice, however the general idea is the same. Another well-known measure is the Gap Statistic, which is similar in spirit to the Hopkins Statistic in that it compares clustering of random uniform data and uses that as a baseline to help us gauge the quality of our clustering. Conclusion Hopefully by now you are familiar with some of the most popular methods in statistical clustering, as well as how to implement these in Python. These methods are widely used in a variety of domains both for knowledge discovery and to aid predictive models, and adding these to our data set gives us one more thing to look for in new data. Thanks for reading.

A Comprehensive Introduction to Clustering Methods

123

a-comprehensive-introduction-to-clustering-methods-1e1e4f95b501

2018-05-27

2018-05-27 20:19:16

https://medium.com/s/story/a-comprehensive-introduction-to-clustering-methods-1e1e4f95b501

false

3,805

null

Machine Learning

machine-learning

Machine Learning

51,320

Shairoz Sohail

Data Scientist

25c54205606c

shairozsohail

5

1

20,181,104

null

0

null

0

null

2018-01-28

2018-01-28 06:33:04

2018-01-28

2018-01-28 06:41:05

1

false

ar

2018-01-28

2018-01-28 06:41:05

1

1e1f455d2ae

1.222642

1

0

على عكس الاعتقاد الشائع أن تعليم الطفل لغة ثانية في عمر مبكرة قد يضرّ لغته الأم، أكدّت دراسات تربوية أن الطفل ثنائي اللغة يتفوّق بمهاراته…

5

تعليم الطفل لغة ثانية لا يضرّ باللغة الأم على عكس الاعتقاد الشائع أن تعليم الطفل لغة ثانية في عمر مبكرة قد يضرّ لغته الأم، أكدّت دراسات تربوية أن الطفل ثنائي اللغة يتفوّق بمهاراته اللغوية والاجتماعية على الطفل أحادي اللغة، كما أن تعلمه لغة ثانية لا يضرّ لغته الأم طالما يستخدمها باستمرار. وفي الواقع أن أفضل وقت لتعلّم طفلك اللغة الثانية هو أثناء تعلّم الأولى حتى تثبت بذهنه، فلا يشعر الطفل بالاستغراب أو الفجوة بين اللغتين لأنه تعوّد عليهما. وقد يُلاحظ الأهل أن الأشخاص ثنائيي اللغة لديهم نوع من الاختلاف في التعامل بسبب لغتهم، والحقيقة أن هذا الاختلاف يأتي بسبب اتباع عادات وتقاليد مستخدمي اللغة وليس بسبب تعدّد اللغات التي يتقنها، ما يُعطيك دافعاً لتعليم طفلك لغة ثانية مع الحفاظ على قيمه التربوية والمجتمعية. ومن الدراسات ذات النتائج المشجعة، ثبوت أن معظم ثنائيي اللغة، لديهم قدرة على إتقان اللغتين بنفس القوة، أي أن إحداهما لن تؤثر على الأخرى ما دام الطفل يستخدم اللغتين باستمرار. هذه الحقائق تدفع للبحث عن طرق جذابة للطفل لتعلم لغة ثانية وتقويتها، ولعل أبرزها الاختلاط المباشر مع متحدثين باللغة الثانية، ثم استخدام القصص والألعاب، ويمكن اللجوء أيضاً إلى الأفلام التعليمية ومواقع الإنترنت أو تطبيقات الهواتف الذكية. ومن بين أبرز تطبيقات الهواتف المميزة في تعليم الطفل لغة ثانية تطبيق NameO الذي يُحوّل الصورة إلى كلمة وينطقها، عبر التصوير الفوتوغرافي لكل ما يحيط بالطفل، فيعلّمه التطبيق معنى الصورة وينطقها له. وللاستفادة من التطبيق، يحتاج الطفل تدريباً بسيطاً على التقاط صورة مقرّبة لأي شيء حوله، حتى تظهر له مباشرة الكلمة باللغة المختارة وتهجئتها، وبعد التصوير لا يحتفظ التطبيق بالصور ولا يجمع أي بيانات شخصية، كما يتوافق مع سياسة حماية خصوصية الأطفال على الإنترنت. تعرّف إلى التطبيق وقم بتحميله من هنا http://www.nameoapp.com

تعليم الطفل لغة ثانية لا يضرّ باللغة الأم

1

تعليم-الطفل-لغة-ثانية-لا-يضرّ-باللغة-الأم-1e1f455d2ae

2018-02-10

2018-02-10 16:43:36

https://medium.com/s/story/تعليم-الطفل-لغة-ثانية-لا-يضرّ-باللغة-الأم-1e1f455d2ae

false

271

null

Learning

learning

Learning

37,342

NameO app

NameO is a picture-to-words interpreter app for kids. It identifies objects in photos you load or capture using the app. www.nameoapp.com

7bd064afd838

nameoapp18

2

1

20,181,104

null

0

null

0

null

2017-09-12

2017-09-12 10:40:47

2017-09-12

2017-09-12 10:41:04

1

false

en

2017-09-12

2017-09-12 10:41:04

0

1e1ff6c44e42

0.909434

0

On Thursday, September 14th, Spaceti’s co-founder and CTO Ondřej Plevka will share his thoughts on how smart technologies help people…

1

Spaceti at IBM Watson Solution Market On Thursday, September 14th, Spaceti’s co-founder and CTO Ondřej Plevka will share his thoughts on how smart technologies help people understand and manage the buildings and spaces they interact with on a daily basis. The presentation will be held as part of the Watson Solution Market conference in Prague, which aims to present solutions based on IBM services. “PropTech is changing interaction of people and buildings.” Ondřej Plevka, Co-Founder & CTO, Spaceti Spaceti, an indoor location solution combining custom built applications and sensors, with services running on IBM Bluemix, enables digitalization of buildings. The communication between buildings and people is taken to a new level thanks these intelligent technologies disrupting the world of Real Estate today. The emerging PropTech sector is changing how buildings and different types of spaces are being occupied, operated and owned, in order to bring sustainable quality service and attract a new generation of occupants. Come to hear Ondřej present how our solutions are leading PropTech innovation and discuss more about our technology, and the future of workplaces at Spaceti’s booth on September 14th, Parkhotel Praha.

Spaceti at IBM Watson Solution Market

0

spaceti-at-ibm-watson-solution-market-1e1ff6c44e42

2018-01-22

2018-01-22 22:05:52

https://medium.com/s/story/spaceti-at-ibm-watson-solution-market-1e1ff6c44e42

false

188

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Spaceti

null

670ea8826cdb

Spaceti

2

1

20,181,104

null

0

null

0

945c2d13ce26

2018-07-02

2018-07-02 07:32:17

2018-07-02

2018-07-02 07:57:03

1

false

en

2018-07-02

2018-07-02 07:57:03

1

1e20ae1709b0

1.581132

0

I was working on some AI for a few different projects I have going on right now and found myself taking a picture of a house. The algorithm…

1

Artificially Intelligent Cowboys I was working on some AI for a few different projects I have going on right now and found myself taking a picture of a house. The algorithm told me, “I am 68% sure it’s a purse.” A view of the houses in Palestine / Photo by Rob Bye Now, that would be completely normal for anyone starting out in AI. There are many complexities in training the models properly. You can use Google’s generous models like Inception, ResNet, MobileNet, etc. and get pretty decent results. These models took plenty of time and resources to build and they can do a really good job of identifying main objects in an image. So what happened? I live in the Middle East and the algorithm didn’t know a Middle Eastern house as well as it knew a European victorian house or a modern Silicon Valley apartment building. Bias is nothing new. Even in AI circles it is also a lengthy and worthy ongoing discussion. The West has been throwing all of their resources towards “winning” this AI race and naturally they forget the biases held in the selective training of models. The business intelligence and anthropological understanding we can gain from investing in AI is absolutely worth the effort. However, when we train new AI / ML models we must understand the information on which we train the model as well as our blindspots in the selection of our data/streams. Chimamanda Ngozi Adichie expresses this concept well in her Ted Talk, The Danger of a Single Story, where she outlines the narratives she grew up with. At one point she says that when asked to draw a house, or ask, “how’s the weather,” or talking about tea time and how unnatural that was to her culture. She had adopted that speech, those appearances, and those concepts due to the large amount of Western literature she had access to. An African, indigenous to her area, would never ask, “How’s the weather?” because it never changed. So as we are progressing with AI, how can we better average out the AI Cowboys of the West with the rest of the world’s indigenous and unique experience of life? How can AI be trained to produce a truly multicultural neurological network?

Artificially Intelligent Cowboys

0

artificially-intelligent-cowboys-1e20ae1709b0

2018-07-02

2018-07-02 07:57:04

https://medium.com/s/story/artificially-intelligent-cowboys-1e20ae1709b0

false

366

We stand in the space between technology, humanity and creation envisioning and cultivating opportunities that unlock their vast potential.

null

tuluaio

null

Tulua

tribe@tulua.io

tulua

TECHNOLOGY,CREATIVITY,SUSTAINABILITY,AGRICULTURE TECHNOLOGY,AI

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Jonathan Reyes

One day, I’m gonna make the onions cry.

9ba66156626a

albanda

85

148

20,181,104

null

0

null

0

f41f22e4979f

2018-09-11

2018-09-11 07:06:57

2018-09-11

2018-09-11 07:16:47

1

false

en

2018-09-11

2018-09-11 07:16:47

11

1e21747cb95b

2.275472

3

0

The governor of the state of California, Jerry Brown, has signed a law that, as of October 2019, will replace cash bail with an algorithmic…

5

IMAGE: Max Pixel — CC0 Justice and algorithms: California and bail The governor of the state of California, Jerry Brown, has signed a law that, as of October 2019, will replace cash bail with an algorithmic system that will estimate flight risk or the likelihood of a defendant committing a crime while awaiting trial. The algorithm, which counties will be obliged to use, either through a provider or by developing it themselves, will qualify risk as low, medium or high and, depending on the result, a judge will decide to release alleged offenders or keep them in custody. Civil rights organizations have hailed the elimination of bail in the form of monetary payment a success, who argue the practice is discriminatory and creates a two-tier justice system for the rich and the poor, whereby those lacking economic resources spent time in prison while awaiting trial, which, in many cases, became part of the problem. In Brown’s words, “Today, California reforms its bail system so that rich and poor alike are treated fairly” However, it has been shown that algorithms, fed with data obtained from the justice system itself, have in some cases shown racial or socio-economic bias. Algorithms are not inherently neutral: they reach their conclusions from the data they are fed, which increases the risk of perpetuating biases that in many cases already existed. On the other hand, many algorithms are defined as owners that protect the intellectual property of the companies that develop them, the so-called black boxes against which there is virtually no appeal. This issue has been highlighted with the use of COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), an algorithm developed by Equivant to assess the risk of prisoners reoffending and to help judges make decisions on their sentences. A decision made by that algorithm was the subject of an appeal in Loomis v. Wisconsin, in which the defendant alleged that the use of such an algorithm in the calculation of his six-year sentence, violated his rights because it prevented him from contesting the scientific validity and accuracy of evidence against him because his defense team did not know how the algorithm worked and also because it also took into account variables such as the gender and race of the accused. The legislation approved by the State of California establishes a period of review for the algorithmic system until 2023, after four years of operation and data generation, at which time the decisions it has taken will be examined. Algorithms could certainly help reduce courts’ workloads, particularly in the early stages of routine cases, albeit with the appropriate oversight. However, this would require mechanisms to allow all parties involved to know how the algorithms work, the variables used to determine the results, in addition to ensuring that the data used to train these algorithms do not contain factors that create biases. No easy task, but one whose results could help prevent the collapse of the justice system in many countries, prompting campaigns along the lines of justice delayed is justice denied. How will the system approved by the state of California evolve? Will it provide better justice and equality than the previous system? And what discussions will we need to have has more and more algorithms are used by our judicial systems? (En español, aquí)

Justice and algorithms: California and bail

5

justice-and-algorithms-california-and-bail-1e21747cb95b

2018-09-12

2018-09-12 00:11:12

https://medium.com/s/story/justice-and-algorithms-california-and-bail-1e21747cb95b

false

550

On the effects of technology innovation (writing in Spanish at enriquedans.com since 2003)

null

enriquedans

null

Enrique Dans

enrique.dans@ie.edu

enrique-dans

TECHNOLOGY,DISRUPTION,SOCIAL MEDIA,INNOVATION

edans

Criminal Justice

criminal-justice

Criminal Justice

4,204

Enrique Dans

Professor of Innovation at IE Business School and blogger at enriquedans.com

acfc0d77aeef

edans

32,707

236

20,181,104

null

0

null

0

null

2017-09-21

2017-09-21 04:01:23

2017-09-21

2017-09-21 04:03:28

1

false

en

2017-09-21

2017-09-21 04:03:28

2

1e228fb4afcb

1.528302

1

0

He’ll be back. She’ll be back. They’ll be back. And, no one, except for some misguided, desperate, insulated Hollywood bigshots asked for…

5

Why The Next Three Terminator Reboots That No One Wanted Will Fail He’ll be back. She’ll be back. They’ll be back. And, no one, except for some misguided, desperate, insulated Hollywood bigshots asked for three more Terminator movies. As The Hollywood Reporter noted, James Cameron announced the next three installments at a private gathering celebrating the franchise. It’s set to feature Linda Hamilton and Arnold Schwarzenegger as the storyline’s anchors. Why? It’s far from a mystery that Paramount Pictures is struggling. The pressure on new Paramount Chief Jim Gianopoulos to strike gold is mounting by the day. Compound that with James Cameron, whose mega hit Avatar hasn’t really aged well, isn’t even directing these next three Terminator installments. Instead, he’s going to executive produce by managing a team of writers and directors. Sounds like a winning formula, right? But wait, there’s much, much more. As if there isn’t enough cucked content out there already, it seems they want to make this a girl power reboot in the vein of the latest Star Wars movies. More on that from The Hollywood Reporter: But the new movie will also be seen as a passing of the baton to a new generation of characters. “We’re starting a search for an 18-something woman to be the new centerpiece of the new story,” Cameron said. “We still fold time. We will have characters from the future and the present. There will be mostly new characters, but we’ll have Arnold and Linda’s characters to anchor it.” Confused yet? So, Arnold and Linda will hang out as old timey tour guides for yet another Katniss Everdeen heroine even though the entire franchise hangs on mankind’s savior John Connor? Cool story, bros. With so much imagination going to waste at a time when Artifical Intelligence is an existential hot topic, it’s a shame that we’re getting more girl power retreads, rather than bold, visionary sci-fi projects not tainted by Cultural Marxism. New stories are sorely needed to guide us through the next few decades of technological perils. Americans won’t be back for these flops. Originally published at americansowoke.com.

Why The Next Three Terminator Reboots That No One Wanted Will Fail

3

why-the-next-three-terminator-reboots-that-no-one-wanted-will-fail-1e228fb4afcb

2018-01-07

2018-01-07 22:28:14

https://medium.com/s/story/why-the-next-three-terminator-reboots-that-no-one-wanted-will-fail-1e228fb4afcb

false

352

null

Movies

movies

Movies

84,914

American So Woke 🇺🇸

hollywood outcast. screenwriter/author of brand new sci-fi action thriller TRUCKER. site: www.americansowoke.com

319544cf7d7

AmericanSoWoke

479

76

20,181,104

null

0

null

0

null

2018-02-07

2018-02-07 04:42:52

2018-02-08

2018-02-08 16:42:41

1

false

en

2018-02-08

2018-02-08 16:42:41

3

1e2331a464ef

3.569811

0

FRENCH TRANSLATION ALSO AVAILABLE

5

“ONCE UPON A TIME” WHY STORYTELLING IS VITALLY IMPORTANT TO MARKETERS FRENCH TRANSLATION ALSO AVAILABLE Once upon a time, there was a world where people didn’t watch TV any more. There was a world where people used adblockers when they go to the Internet. People wanted brands to talk to them only when they needed it. Marketers couldn’t attract attention by disturbing them anymore. But there was a little fairy full of emotions called ‘Storytelling’ that could help marketers. Here is her story Micro Moments Are Changing The Rules The importance of moments Our life is made up of micro-moments. These micro-moments are changing the relationship between brands and clients. We spend most of these micro-moments through the Internet. According to Google, there are four types of “Micro Moments” (1) I Want To Know Moments When someone is exploring, but is not in purchase mode. (2) I Want To Go Moments When someone is considering buying a product at local store. (3) I Want To Do Moments When someone needs help finishing a task or wants to do something new. (4) I Want To Buy Moments When someone is ready to buy, and may need help. All these micro-moments represent chances for brands to engage with the clients. Brands can make this experience great just by giving the right information at the right time. 30,000 Decisions Every Day Micro-moments, micro-experiences mean micro-decisions. We make more than 30,000 micro-decisions a day. It means we have too much information. Do we have the time to analyze them? No, we don’t! This is why we are looking for tailored content. Personalisation Makes Our Choice Easier We give away a large amount of data to companies. In turn, they should help us to make our choices easier. As an example, 75 percent of user activity on Netflix is driven by tailored advices. Emotion Makes Us Act Giving tailored advice however, is not enough. We need a touch of EMOTION to bring us to buy. As Alan Weiss has noted, “Logic makes people think; emotion makes them act.” Even if we ask factual questions, our emotions finally lead our choices. Microsoft has found that our attention span is now less than that of a goldfish. To face this situation, we can no longer use old marketing strategies. No more marketing campaigns focusing on the promise of a product with limited emotional thought. Procter & Gamble Focused On Empathy & Emotions The best-case study to explain this is taken from Proctor & Gamble’s Always® brand. In 2013, they noticed their message was no longer reaching their young target audience anymore. The reason was because they focused mostly on product features. They decided to focus more on people. They began to understand their clients on a deeper level. They discovered that half the women claimed they experienced a decline in confidence at puberty. In response, Proctor & Gamble developed a social media campaign called #likeagirl It focused on empathy and emotions. It worked. Emotion is the key to convey our message. Storytelling Is An Excellent Emotional Tool Storytelling is one of the great tools that can help us engage through emotions. First, we need to understand the specific problem our clients have. Then we have to find how we can solve it. Clients want to get to the point fast. We need to focus on how we can solve their problem or make their life easier. We need to address them through a story. It will create a sense of fit to their needs. Think Like Filmmakers 1) A story has to start with a familiar situation. Then it has to be broken up by a problem. In this way, we catch the attention. 2) The story explains what can happen if we do nothing. We need to find a solution. 3) Then, we bring a solution to create a comfort zone for the client. They feel better. They will continue to listen to the story in order to get the solution 4) Every story ends with a strong call-to-action!! Since emotions drive our purchase decisions, storytelling represents a powerful tool. It creates a world where brands and clients can share an awesome experience. Personalised Storytelling & Profitability Is Challenging Everyone wants a tailored story. We as marketers need to build this one-to-one connection between our brand and our clients. Among the customer base, there exist multiple micro-niches. Telling stories for each micro-niche while striving to maintain a profitable business is a challenge. Data Driven Culture Can Help Tailored storytelling can only work if our company has a data driven culture. By placing a high value on data, the process of collecting customer data, defining micro-niches and detailed analysis become prioritized. From there, we will be able to create stories talking about customer emotions by leveraging Artificial Intelligence. The algorithms will be able to help increase relevance for each micro-niche. As we continue to feed the algorithm additional data, the end results become more and more relevant. In conclusion, we live in a world constrained by decreasing time and increasing information. Storytelling is one of the solutions to talk to people who make decisions driven by their emotions. This article is also available on http://ow.ly/hWGU30ift6A Thank you for reading me ❤️ Fabrice Briatte

“ONCE UPON A TIME” WHY STORYTELLING IS VITALLY IMPORTANT TO MARKETERS

0

once-upon-a-time-why-storytelling-is-vitally-important-to-marketers-1e2331a464ef

2018-02-08

2018-02-08 16:42:42

https://medium.com/s/story/once-upon-a-time-why-storytelling-is-vitally-important-to-marketers-1e2331a464ef

false

893

null

Marketing

marketing

Marketing

170,910

Fabrice Briatte

Customer Experience (CX) and #Marketing Advisor / #Design Thinking Enthusiast. Let's connect https://ca.linkedin.com/in/fabricebriatt

2502b9a0a84b

Fabrice_Briatte01

830

988

20,181,104

null

0

null

0

null

2018-01-27

2018-01-27 01:03:27

2018-01-26

2018-01-26 21:52:14

4

false

en

2018-01-27

2018-01-27 01:04:15

12

1e241abddf2

4.250943

1

0

“The Fourth Industrial Revolution is still in its nascent state. But with the swift pace of change and disruption to business and society…

4

Blockchain, and Industry 4.0: a closer look “The Fourth Industrial Revolution is still in its nascent state. But with the swift pace of change and disruption to business and society, the time to join in is now.” A question steers clear: is Gary Coleman too eager to witness a global shift towards cyber systems? Or is it a work in progress only in the northern hemisphere? We can sure speculate, even theorize, but facts seem to point toward a different diagnosis. Technology drives change and enables new milestones in human history. Industry 4.0 is not an exception here. Here is a closer look into key technologies that make it possible. Beginning with the Internet of Things. Internet of things IoT is a paramount driver of Industry 4.0 — It fuels the ecosystem, like gas to the combustion engine. It’s the new electricity. Just like a computer have an IP address in the wake of the Internet, everything gets an IP address in the IoT. Machines, sensors, appliances and miscellaneous devices are connected. They interact, report, and operate on a standby mode. “We are entering an age where parts can monitor and evaluate their own performance and even order their own replacement when necessary,” said Gutzmer, deputy CEO and CTO at Schaeffler. This ecosystem is often referred to as the Industrial Internet of Things or IIoT. Blockchain The switch towards industry 4.0 has thrust a technological boom where blockchain seems to shine brightly. Industry 4.0 aims to achieve machines autonomy, and blockchain might hold the dialogue between machines (M2M) and as a mean of a “skeleton” or ledger. Using this technology, cyber-physical systems composing the “smart-factories” can safely and autonomously place an order for their replacement parts, optimize the process, and detect process malfunctions beforehand. Furthermore, the potential of blockchain to allow frictionless and transparent financial transactions, among any number of smart devices, makes it indispensable for the economic shifts that industry 4.0 entails. Cloud Computing It’s is the delivery of computing services — servers, storage, databases, networking, software, analytics and more — over the Internet (i.e. the cloud). Cloud providers typically charge for computing services based on usage. In early 2017, Amazon’s AWS posted a 43% YoY increase in revenue. Using Blockchain, a startup called Storj provides decentralized cloud computing service, based on anonymity, and the lowest of fees paid for in crypto-coins. We have here rapid services at the lowest of fares, with the ability to set up a new computing service in weeks. Hence, Cloud computing is regarded as the shaft to Industry 4.0 engine. Augmented Reality Augmented reality (sometimes also referred to as “mixed reality”) is the technique of adding computer graphics to a user’s view of the physical world. You might have experienced this on your smartphone if you played the game Pokémon GO. Or perhaps while placing furniture in your house using the IKEA Place app or the AR View feature on Amazon’s smartphone app. We’ve seen the hype for Virtual Reality primarily in consumer applications but that’s just the hanging fruit. AR is the disruptor in Industry 4.0. For instance, Researchers at MIT Media Lab have demonstrated how results of a product search can be displayed directly on the supermarket shelf. Things are just getting started for the end appliance in a fully connected world — where information about the factory or equipment is instantly over layered into the technician’s view of the task at hand. 3D Printing 3D printing is paramount to Industry 4.0. It revolutionizes the design/testing steps of production. This leaves enough room to transform manufacturing. And that is already taking shape. GE’s Brilliant Factory looks something like this. Customer data is used to understand inventory levels of parts. Production follows real-time customer data, where AI is the shift supervisor. Using 3D printing technologies, they can design, prototype and test a new part in hours instead of days or weeks. The important point to stress is that you can’t start with 3D printing, you need to have an overarching strategy for Industry 4.0. A fully connected ecosystem Industry 4.0 comprises a fully connected ecosystem, built around smart manufacturing processes. Three major manufacturing related shifts drive this: Unprecedented Data Connectivity, Business Model Shits, and Digital Transformation. The current gradual disruption is invigorated by extensive “data revolution” and, to a lesser extent, hardware upgrades. Hence, sizeable amounts of the investment are oriented towards deployment of IoT without substantial hardware expenditure. Integration is a green field for innovation so to speak. One lead to follow Industry 4.0 will be a vast arena for innovation for the decades to follow. Silicon Valley continues to put forth the promise of a great change, with small startups having the furthest reaching of potentials. But IoT will democratize innovation, and decentralize initiative. It’s only logical to stumble upon resistance and feed on failures and unfulfilled effort. Companies embarking on the journey shall make it a priority to form working groups, and train Industry 4.0 “knights”. Trial and error, experience feedback, and scientific publications will certainly pave the way. But taking risks, and having faith in both the machines and human ingenuity that is crafting them is key. Industry 4.0 will bring about a whole new ecosystem, to a whole new world. A connected one that doesn’t lose time adjusting. The speed is high, the stakes are higher, and so is the promise. Since the potential outcome can alter the very business model of companies, willingness to embrace change and act on these disruptive ideas must be paramount. 3D Printing AR Blockchain cloud computing Industry 4.0 IoT M2M Originally published at en.decentral.news on January 26, 2018.

Blockchain, and Industry 4.0: a closer look

10

blockchain-and-industry-4-0-a-closer-look-1e241abddf2

2018-01-29

2018-01-29 00:00:55

https://medium.com/s/story/blockchain-and-industry-4-0-a-closer-look-1e241abddf2

false

941

null

Blockchain

blockchain

Blockchain

265,164

Decentral Blockchain & Cryptocurrency

Decentral magazine provides information, news and expert commentary on Blockchain and cryptocurrencies.

409837acd632

DecentralEng

39

10

20,181,104

null

0

null

0

null

2017-12-24

2017-12-24 23:24:49

2017-12-24

2017-12-24 23:43:19

9

false

es

2018-02-09

2018-02-09 20:58:10

0

1e241fa84abc

6.762264

4

0

Todo empezó al ver que Colombia e India tienen la mayor cantidad de dias festivos al año en el mundo: 18 dias. Un tweet de la politóloga…

5

Colombia con más dias del Sagrado Corazón y menos dias del trabajo. Todo empezó al ver que Colombia e India tienen la mayor cantidad de dias festivos al año en el mundo: 18 dias. Un tweet de la politóloga @sandraborda sobre este tema tratado por la consultora mercer, me puso a pensar: Lo primero que pensé fue que con todo lo que falta por hacer en este país, 18 son demasiados dias y tanto tiempo libre no aporta al crecimiento del país. En la tabla observamos los cinco países con más festivos al año en tres continentes. En la matriz vemos que a más dias festivos, mayor trabajo informal pero con una correlación no muy fuerte (cor 0,46). Las flechas señalan el eje en donde se ubica la variable. También se observa que a más trabajo informal, más muertes violentas por cada 100.000 habitantes (cor 0,5) y aquí hay una gran variación ya que según worldlifeexpentancy Colombia tiene 38 muertes violentas al año por cada 100.000 habitantes, Venezuela 47 y paises como Finlandia tiene 2, Corea 1.1, Malta 0.7 y Japòn 0.28. Gráficamente aparece una correlación positiva mucho más clara (cuadro con flechas azules), lo que lleva a identificar la variable trabajo informal como una variable más nociva que la variable dias festivos. Por otro lado al ver la correlación entre los dias festivos y la cantidad de muertes violentas, vemos que no hay relación, ya que aunque se genera una correlación negativa (cor — 0.10) este valor más cercano al cero no nos lleva a decir que a más dias festivos menos muertes violentas. Sorprendentemente dos variables que pensé que crucificarían completamente a los dias festivos para poder decir tranquilamente que 18 dias festivos al año no son buenos para el crecimiento del país, no lo hicieron. Resulta que analizando los datos de estos 15 paises, al correlacionar los dias festivos con el pib per cápita aparece una tímida correlación negativa (-0.25) , que no permite decir lo que me encantaría decir y es que a más dias festivos menor pib per cápita. Y si vemos la gráfica (cuadro con flechas rojas) observamos que los puntos no parecen descender muy claramente. Al abordar la variable pib per cápita aparece de forma invisible la variable población y llama la atención que según Ciafactbook, el pib per cápita de Colombia es más del doble que el pib per cápita de India y la diferencia en el tamaño de la población es enorme (48 millones frente a 1.200 millones). Llama la atención porque de estos 15 paises, hay unos con mayor población que Colombia y un mayor pib per cápita. Es el caso de Rusia y Brasil. Esto lleva a ver a India con otros ojos mas allá de su tamaño. Además, esta India con un pib per cápita menor al de Colombia es la misma India que Bloomberg destaca como la próxima potencia superando a China. El desempleo no ayudó mucho porque su correlación con los dias festivos fue de solo -0.44. Y es que con paises como Japón con un desempleo de 3.1% ó Corea del Sur con 3.7% y los dos con 16 dias festivos al año, no hay argumentos para atacar a los dias festivos. Es más, Japòn tenía 15 dias festivos al año y desde el 2.016 empezó a celebrar un nuevo festivo: el “dia de la montaña” para completar los 16 dias. La cantidad de dias festivos al año si puede tener efectos en la economía de un país y en la calidad de vida de sus habitantes pero claramente esta relación no siempre es lineal y no necesariamente porque se trabaje más dias al año, la economía de un paìs será más fuerte. Esto me lleva a pensar en un tema cultural, en la forma de ver la vida. Tal vez por esas diferencias culturales a pesar de tener Japón un gran número de festivos al año, muchos japoneses trabajan muchas horas extras y no se toman los dias libres correspondientes, según indica la consultora mercer y por ese motivo y también para incetivar el consumo, estrenaron un nuevo festivo. Será que los dias festivos aportan a la felicidad de los ciudadanos de estos 15 paises? No. La correlación es del - 0.3 (World Happiness Report). Y antes de ver la relación de competitividad con felicidad y dias festivos, es muy evidente que los puntajes en competitividad de los paises con más dias festivos, son menores en el continente americano que en Asia ó en Europa. Vemos que de estos 15 paises, solo dos tienen un pib per cápita inferior al de Colombia (Venezuela y la misteriosa India) y hay cinco paises con más del doble del pib per cápita de Colombia, lo que indica que se trata de economías mucho más fuertes y después de ver el caso de Japón y sus habitantes que trabajan horas extra y no se toman sus dias libres, se podría pensar que en este y otros paises del grupo, hay otras variables que se correlacionarían positivamente con la felicidad pero no se trata de los dias festivos. Colombia se encuentra en el quinto lugar en este grupo de paises en cuanto a puntaje de felicidad con 6537 puntos. En primer lugar está Finlandia con 7469 puntos y en último lugar en estos 15 paises se encuentra India con 4315 puntos. Japón y Corea también tienen calificaciones inferiores a las de Colombia: 5920 y 5838 respectivamente. En cuanto a competitividad, aunque la gráfica nos muestra una correlación positiva entre la cantidad de dias festivos y el puntaje de competitividad del país según la calificación del Foro Económico mundial, la correlación es solo del 0.17 y como en este top 15 por dias festivos hay paises más desarrollados y de otras culturas distintas a la latinoamericana, me inclino a pensar que esa otra forma de ver el mundo es lo que los puede hacer más competitivos y por eso no es necesario que tengan menos dias libres para que sus paises progresen. Bloomberg menciona a un país con 18 dias festivos al año como próxima potencia mundial y no se trata de Colombia. Y aunque Colombia supera a India en pib per cápita, en tener menos trabajo informal, en el ranking doing business, en expectativa de vida y en felicidad, India tiene menos desempleo que Colombia (3.5% vs 9.2% según cifras del banco mundial en 2016), tiene solo 6 muertes violentas por cada 100.000 habitantes cuando Colombia tiene 38 y tiene un puntaje de competitividad superior a Colombia (4.5 vs 4.3 y por ranking, India es puesto 39 vs puesto 61 de Colombia). Un país con este alto nivel de competitividad es un país enfocado y no de forma gratuita ha logrado tener su gran Silicon Valley en la ciudad de Bangalore ni de forma gratuita hace parte de los BRIC countries ó BRIC economies, término acuñado por Jim O’Neill en 2.001 y que se refiere a países con altos niveles de desarrollo. Originalmente eran BRICS (Brasil, Rusia, India, China y Sur Africa). Ya ha salido Sur Africa y el tiempo nos dirá si la “B” sobrevive. Japón con su incremento en dias festivos me parece un buen ejemplo para descifrar si 18 dias festivos son malos ó buenos para Colombia. Pareciera que cuando el mundo se ve de diferente manera, la cantidad de dias libres al año no afecta negativamente en el crecimiento de un país. Por eso no sorprende noticias como la muerte de un médico japonés de 105 años que tenía actividades en su agenda para los siguientes dos años. Trabajar, producir y crecer pareciera ser parte de la cultura de los paises desarrollados de esta lista. Japón se encuentra entre los paises con mayor edad de retiro: 65 años tanto para hombres como para mujeres. Y un japonés jubilado vive en promedio 19 años después de su jubilación según cifras de Bloomberg. Más que la comida ó el ejercicio, el estar activo parece ser el secreto. El secreto para el bienestar de los ciudadanos y de todo el país. Después de ver los datos y ver que la cantidad de dias festivos no inciden en los resultados negativos de las economías y tampoco en los positivos como en la felicidad, creo que la forma de ver el mundo de estos 15 paises puede justificar sus resultados. Y esa forma de ver el mundo tiene que ver con las diferentes culturas y es probable que también con la religión. Así que como Colombia no ha hecho la tarea y aún no somos lo suficientemente competitivos, ayudaría tener menos tiempo libre. Pero además de reducir los dias festivos(México tiene solo siete), es urgente un cambio de mentalidad para que los colombianos piensen en grande y se esfuercen más, aprendan más y produzcan más todos los dias del año. Sonia Ardila @soniaardila1 datelligence

Colombia con más dias del Sagrado Corazón y menos dias del trabajo.

53

colombia-con-más-dias-del-sagrado-corazón-y-menos-dias-del-trabajo-1e241fa84abc

2018-03-19

2018-03-19 15:29:55

https://medium.com/s/story/colombia-con-más-dias-del-sagrado-corazón-y-menos-dias-del-trabajo-1e241fa84abc

false

1,474

null

Data Science

data-science

Data Science

33,617

sonia ardila

#digitaltransformation in @digitai_ . Data storyteller. @soniaardila1

63decd84af6f

soniaardila1

62

85

20,181,104

null

0

null

0

null

2018-03-24

2018-03-24 10:36:34

2018-03-24

2018-03-24 10:40:02

3

false

en

2018-03-24

2018-03-24 10:53:46

2

1e249315535b

0.746226

0

Everybody has a drug … even Artificial Intelligence…

5

Artificial Intelligence Testosterone — holy-chip #17 Everybody has a drug … even Artificial Intelligence… Have a geek laugh! The holy-chip series is a narrative between 2 Artificial Intelligence characters. They do not have names. They are black and white. The date and place posted on the header are absolutely real.

Artificial Intelligence Testosterone — holy-chip #17

0

artificial-intelligence-testosterone-holy-chip-17-1e249315535b

2018-03-24

2018-03-24 10:53:46

https://medium.com/s/story/artificial-intelligence-testosterone-holy-chip-17-1e249315535b

false

52

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Ricardo Mello

null

6df18de4f6fd

rm_90700

18

19

20,181,104

null

0

null

0

7677c8edc203

2018-04-08

2018-04-08 18:01:03

2018-04-11

2018-04-11 01:36:00

1

false

en

2018-04-12

2018-04-12 12:41:51

4

1e25a580239d

2.732075

3

0

FinTech Studios is always trying to stay ahead of the curve on the latest and greatest technologies. So, we’re excited to have attended…

4

Flink Forward 2018 Greetings from San Francisco FinTech Studios is always trying to stay ahead of the curve on the latest and greatest technologies. So, we’re excited to have attended Flink Forward 2018 in San Francisco, April 9–10 2018. Flink Forward is the top conference for all things Apache Flink, an open source stream processing framework specializing in distributed, high performance data streaming applications.We were glad to have the chance to meet up with other industry leaders and discuss the future of data streaming technologies. Core Takeaways Flink is a very healthy Apache project, and we can foresee a robust development effort and pipeline in the future. There are several projects like Apache SOMOA or Apache Beam that aim to generalize SDK languages and domain specific functionality (e.g. machine learning), so application developers don’t have to worry about the underlying stream processor (Flink would work the same as Apache Apex, Spark Streaming, etc, and the developer wouldn’t know the difference). Extra language support seems to be a common theme. Many people use ElasticSearch as a sink for analytics, and then Kibana to display. SQL CLI in Flink 1.5.0! Nifty Tips The Flink AsyncFunction interface is very powerful (you can integrate external operations inside a Flink workflow). However, there are several factors to keep in mind or you’ll run into issues. For example: AsyncFunction uses complete(), not collect(), so the Flink UI shows no backpressure. This can cause for some interesting debugging scenarios. The SQL/Table API is great for data pipelines, low latency ETL, analytics stream, and powering live dashboards. Kafka MirrorMaker makes it easy to duplicate streams across regions for active-active, or any HA architecture. Bootstrapping Flink does not exist of the box. However, one option is to use stream retention (Kafka, Kinesis, etc) rather than rely on Flink itself. Conference Summary The main theme here, without a doubt, is big data. Specifically, how to build an architecture and system that can efficiently scale to meet the growing needs of production applications. Clearly Apache Flink is a very healthy open source project; it’s one of the top 10 most active Apache mailing lists, and has on average 12k downloads per month. The diverse set of use cases is clearly a factor in its quick uptake: capital risk, fraud detection, real-time business intelligence, and machine learning, to name a few. Stream processing is about building applications. Steam processing changes the database-centric architecture. As a general trend, we’re seeing a shift away from the late 2000’s “data lake” architecture, where the idea is to collect/store everything now, and figure it out later. The data lake trend clearly coincided with the dramatic decrease in disk capacity & memory prices, and the rise of powerful batch “hammers”, such as Hadoop. Instead, we’re seeing a move towards the Kappa architecture (or something closer to that), where stream analytics play a crucial part in the larger architecture. However, with Apache Flink as an example, stream-centric architectures are causing a stronger coupling of the application and database — this in turn creates challenges in the operational aspects of maintaining these types of systems throughout the software development lifecycle. Some examples: How do you carry state forward when updating, changing, or scaling Flink? How can you fix issues in history, aka data reprocessing/replay? These are issues anyone trying to deploy Flink into production would run into. Data Artisans, the company that is running Flink Forward, has a platform (dA Platform) that aims to solve these issues. As the core Flink project is being improved, several open source projects specializing in certain use cases or sections of Flink are popping up. This includes Pravega, which focuses on stream storage, and Apache Beam, which aims to provide multiple language bindings and domain specific tools, generalized to work with underlying “runners” (including Flink). Overall, the general “atmosphere” is that Flink is a great tool for production use today in several use-cases, but it’s not yet a full enterprise streaming platform. For that, it needs things like RBAC, metadata management, failover/failback, etc.

Flink Forward 2018

3

flink-forward-2018-1e25a580239d

2018-06-17

2018-06-17 16:41:56

https://medium.com/s/story/flink-forward-2018-1e25a580239d

false

671

FinTech Studios Engineering & AI

null

fintechstudios

null

FTS Engineering

support@fintechstudios.com

fintech-studios-engineering

ENGINEERING BLOG,ARTIFICIAL INTELLIGENCE,DATA SCIENCE,FINANCIAL TECHNOLOGY,NYC

fintechstudios

Big Data

big-data

Big Data

24,602

Kevin Barresi

CTO @FinTechStudios

4738af33e6be

kevin_73266

15

1

20,181,104

null

0

null

0

863f502aede2

2017-11-23

2017-11-23 20:36:39

2017-11-23

2017-11-23 20:37:08

1

false

en

2017-11-23

2017-11-23 20:37:08

1

1e27858b97b6

0.962264

9

0

Nature will publish the first issue of Machine Intelligence in January 2019. The new journal will cover the “best research from across the…

3

Nature’s Machine Intelligence Journal to Launch in Jan 2019 Nature will publish the first issue of Machine Intelligence in January 2019. The new journal will cover the “best research from across the field of artificial intelligence,” ranging from machine learning, robotics and human-machine interactions to ethical, legal and societal dimensions of AI. Machine Intelligence will be an online-only publication, enabling readers to access content weeks before it appears in print. The journal will amalgamate fundamental and applied research papers, reviews, commentaries, and relevant news, adhering to Nature’s editorial style. The editorial department will function independently of other Nature journals, allowing authors to either start a fresh submission or re-submit rejected manuscripts. Unlike other Nature publications, Machine Intelligence will have no external editors. Nature announced that Liesbeth Venema, who has a PhD in applied physics from Delft University of Technology in Netherlands and 17 years of experience with Nature, will serve as Machine Intelligence’s Chief Editor. Guide2Research has compiled a list of top journals in the cross-disciplinary field of machine intelligence, ranked according to their JCR impact factor, SJR, and Scopus H-index. Interested readers can check here for more information: http://www.guide2research.com/journals/ Journalist: Meghan Han | Editor: Michael Sarazen

Nature’s Machine Intelligence Journal to Launch in Jan 2019

29

natures-machine-intelligence-journal-to-launch-in-jan-2019-1e27858b97b6

2018-05-30

2018-05-30 08:39:01

https://medium.com/s/story/natures-machine-intelligence-journal-to-launch-in-jan-2019-1e27858b97b6

false

202

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

null

SyncedGlobal

null

SyncedReview

global.sns@jiqizhixin.com

syncedreview

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

Synced_Global

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Synced

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

960feca52112

Synced

8,138

15

20,181,104

null

0

null

0

7559b9dce2fb

2018-07-18

2018-07-18 10:02:41

2018-07-20

2018-07-20 06:49:36

6

false

en

2018-07-20

2018-07-20 07:10:10

0

1e27d9608181

3.387736

3

0

Last week, QARAsoft invited over 25 people to participate in its first ever Hackathon event in Jeju Island. In just three days, this…

5

QARAsoft Heads to Jeju Island for Hackathon 2018 Last week, QARAsoft invited over 25 people to participate in its first ever Hackathon event in Jeju Island. In just three days, this invitation-only event brought together in-house marketers and developers to collaborate on Deep Learning related projects. Team Picture Hangout by the beach Winning Criteria Two teams, QARA and KOSHO, have competed against each other for almost 30 hours straight. The list of criteria for winning consisted of: · Algorithmic Complexity — The algorithm has levels of sophistication and quality/intricacy · Creativity — The algorithm is innovative and unique · Functionality — The program is effective and highly applicable for daily use · Organization — Coding has clarity and structure · Business Model — The model is creative, thorough, and realistic · Marketing Strategy — Creative strategies that fit the business model and targets the right audience and investors · Design — The program has overall good UI/UX · Participation — Everyone has contributed equally to the competition Judges included Luke Changwhan Byun and Katie Bomi Son, the CEOs at QARAsoft, as well as Dr. Shim. Working throughout the night Event Outcome Everyone worked around the clock to create programs geared towards helping people in the financial sector. Each team provided a unique technology with a separate business model, marketing strategy, and design. Here is what they’ve created: Team QARA found a company called StocKloud AI. StocKloud AI uses Natural Language Processing (NLP) to summarize company news to help financial professionals with their businesses. Team KOSHO found a company called Glance. Glance uses the latest Deep Learning technologies to understand, streamline, and display information related to the financial markets. The judges voted Glance as the winner for Hackathon 2018. While the scores for both teams were high, what stood out for Glance was its functionality and its algorithmic complexity. Hogun Kee, an AI Scientist at QARAsoft, was voted the MVP for contributing the most to Glance’s technology development. Hogun Kee with the MVP award Looking Ahead The ultimate goal of this year’s Hackathon is to apply the technologies we’ve developed for future use at the company. Each member who participated in the Hackathon received gifts, including Airpods, AI Speaker, Starbucks gift card, three months of Netflix membership, and two days of vacation. QARAsoft invested a lot into the event, so we made sure to make good use of our time here. After the competition was over, we had a bit of time to enjoy the island’s most exotic foods and tour the famous Cheonjiyeon Waterfall. Visited the Cheonjiyeon Waterfall Since our event was so successful, maybe hosting another Hackathon in the near future might not be such a bad idea. This time, we could invite more people to participate. But we’ll cross that bridge when we get there. A Special Announcement At the beginning of Hackathon, Luke and Katie gave a special announcement about our new upcoming Chief Technology Officer (CTO). As part of the opening jeopardy game, a question was asked about who our next CTO would be. Team QARA guessed the Junior Software Developer, Shin Jung Chul and team KOSHO guessed the Head of Back-End Development, Kevin Hyek Lee. Surprisingly, the decision was made that Shin Jung Chul would be our next CTO. Katie said that this aligns with her vision of giving younger generations more chances to step up into leadership roles. Shin Jung Chul joined QARAsoft on April of this year, and he has shown tremendous efforts ever since. He is sharp, intelligent, and extremely passionate about his job. Shin Jung Chul, our new CTO A Special Thanks This event wouldn’t have been possible without the help of our interns, Kenny Lee and Amy Kang. They took care of all the logistics and handled every details involving program objectives and games. Thanks to our new interns, as well as everyone who had participated in the event, our first Hackathon was a huge success!! To see more about the event, please check out our Hackathon video below! QARAsoft Jeju Island Video

QARAsoft Heads to Jeju Island for Hackathon 2018

30

qarasoft-heads-to-jeju-island-for-hackathon-2018-1e27d9608181

2018-07-20

2018-07-20 07:10:10

https://medium.com/s/story/qarasoft-heads-to-jeju-island-for-hackathon-2018-1e27d9608181

false

646

QARAsoft is a fintech startup that provides online asset management service through the AI (artificial intelligence) deep learning technology.

null

qaracoop

null

QARAsoft

jamesihn@qarasoft.com

qara-soft

BITCOIN,DEEP LEARNING,MACHINE LEARNING,FINTECH,SEOUL

QaraSoft

Hackathons

hackathons

Hackathons

7,969

James Ihn

Content Creator

a9e71f4890cf

jamesihn

5

4

20,181,104

null

0

def prepare_training_data(data_folder_path): faces = [] labels = [] for dir_name in dirs: if not dir_name.startswith("s"): continue; Next we extract label number of subject from each dir_name. Since the format of the dir_name is sLabel so removing the s will give us the label: label = int(dir_name.replace("s", "")) subject_dir_path = data_folder_path + "/" + dir_name subject_images_names = os.listdir(subject_dir_path) Go through each image name, read image, detect face and add face to list of faces: for image_name in subject_images_names: ignore system files like .DS_Store if image_name.startswith("."): continue; Next build image path,sample image path = training-data/s1/1.pgm: image_path = subject_dir_path + "/" + image_name image = cv2.imread(image_path) cv2.imshow("Training on image...", cv2.resize(image, (400, 500))) cv2.waitKey(100) face, rect = detect_face(image) image = cv2.imread(image_path) cv2.imshow("Training on image...", cv2.resize(image, (400, 500))) cv2.waitKey(100) face, rect = detect_face(image) gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) face_cascade = cv2.CascadeClassifier('opencv-files/haarcascade_frontalface_alt.xml') faces = face_cascade.detectMultiScale(gray, scaleFactor=1.2, minNeighbors=5); OpenCV Error: Assertion failed (size.width>0 && size.height>0) in cv.resize, file .../opencv/modules/imgproc/src/resize.cpp, line 4044 if (len(faces) == 0): print("No face detected") r = cv2.selectROI(gray) faces = r (x, y, w, h) = faces return gray[y:y + w, x:x + h], faces face_recognizer = cv2.face.LBPHFaceRecognizer_create() for t in range(100000): face_recognizer.train(faces, np.array(labels)) face_recognizer.save('trainer/trainer.yml') test_img = cv2.imread("test-data/image9.jpg")

15

null

2018-09-03

2018-09-03 09:54:37

2018-09-03

2018-09-03 09:41:16

0

false

en

2018-09-03

2018-09-03 12:20:15

2

1e2d5597a601

4.475472

2

0

So what comes to your mind when you come across the term Face-Recognition? You have seen a face many times and kept it in your memory that…

4

Face-Recognition Using OpenCv So what comes to your mind when you come across the term Face-Recognition? You have seen a face many times and kept it in your memory that this face belongs to this person, and then when you come across that known face you recollect from memory and recognize that face! so it is as simple as that!! So how does your computer do that task?? Face/Image Recognition is a part of Computer Vision, and a process to identify and detect objects or attributes in a digital video or image. We consider here Python’s OpenCv for accomplishing this task. OpenCv is python’s library build to improve computer vision problem. OpenCv comes with its prebuilt FaceRecognizer class for face recognition. The technique used here is pretty straightforward : Collecting image of the faces we want to identify. Training the model over the collected images. Testing the model by feeding it with different images of the faces. Here the git link of the entire code https://github.com/ayindrilla-maiti/Face-Reco-openCV. There are two main python files that are involved here. trainer.py and face detection.py. The trainer is for training the model and face-recognition is for recognizing images using the training. Here we will discuss how we are training the model. Training Data Usually, for training, a large number of images are used. A large number of images of a person is used so that the Recognizer can learn from the different looks of the person. The training data is inside the training-data folder. The training data folder or the subject directory is named in the type slabel i.e first folder containing the images as s1, second as s2. Preparing the Training Data Two lists are taken for storing the images of faces and their labels. Since the subject directories start with ‘s’ we ignore the other non-relevant directories if any. Detecting Faces First, the images are needed to be converted in grayscale as sometimes color information doesn’t help in identifying the edges and other important features. Next, the OpenCv haarcascade classifiers is used. OpenCv comes with a set of pre-trained classifiers for face, eyes, smile. Also, you can create a classifier for your own objects. Haar cascades classifiers consist of XML files which contain a lot of features for a specific set of objects. The XML files are stored in –> opencv/data/haarcascade folder. In our case, we are using the haarcascade_frontalface_xml. Next, these XML features are used to detect the face in our image, so that is basically the Region of Image or ROI. here the parameters scaleFactor: It is for creating an image pyramid. An image pyramid is a multi-scale representation of the image so that it is scale-invariant i.e in simple word it could detect both small and large picture. Here the value 1.2 denotes that the image is reduced by 20% in each step. minNeighbors: it denotes the minimum no. of neighboring rectangles. Haar classifier uses the concept of sliding window. In every image that is being used, Haar classifier creates small resizable rectangles within each object that are been detected. Among those objects, some are termed as false positives meaning they are not the actual object we want to detect. Thus to eliminate the false positives and to get the actual object comes to the concept of neighboring rectangles. It is that a rectangle is the neighbor of some other rectangles then that could be the actual object we want to detect. Here is the issue that I encountered, that in some images that is used in training data the Region of Image(ROI) could not be identified using the haarcascade classifier, hence the variable faces can return null values and as a result, it throws an exception like : at runtime. I have analyzed that this could be because it is not getting the ROI. so to identify the ROI cv2.selectROI() can be used. This enables the user to manually select the region of the image using a bounding box. Although this might not look like an optimized solution, I am still looking at how to do it better. Training Recognizer OpenCv has its built-in Recognizer class which can be easily used in the for face-recognizing. OpenCv has three algorithms for face recognition: EigenFaces — cv2.face.createEigenFaceRecognizer() FisherFaces — cv2.face.createFisherFaceRecognizer() Local Binary Patterns Histogram(LBPH) — cv2.face.createLBPHFaceRecognizer() All three methods perform the recognition by comparing the face to be recognized with some training set of known faces. In the training set, we supply the algorithm faces and tell it to which person they belong. When the algorithm is asked to recognize some unknown face, it uses the training set to make the recognition. Each of the three aforementioned methods uses the training set a bit differently. Eigenfaces and FisherFaces find a mathematical description of the most dominant features of the training set as a whole. LBPH analyzes each face in the training set separately and independently. Here LBPH recognizer is being used. Next, we train the recognizer for around 1 million steps and in each step store it in a trainer.yml file, which will be used for recognizing faces. Recognize Images Now after the training is completed run the face-recognizer.py. Remember to feed it the input image you want it to recognize. And again for some images, it fails to detect the ROI and for that, we are selecting it manually. Although am trying to do some better method for this, I will update again if I find any. Okay, finally we are almost at the end of our mission where this face-recognizer correctly recognizes the image. Now you might get an incorrect result, in that case, increasing the training data and again training the model will help, also the system is not 100% accurate I am working in it still and would always welcome feedback and contributions! Thanks! Originally published at ayindrilla.wordpress.com on September 3, 2018.

Face-Recognition Using OpenCv

52

face-recognition-using-opencv-1e2d5597a601

2018-09-03

2018-09-03 12:20:15

https://medium.com/s/story/face-recognition-using-opencv-1e2d5597a601

false

1,186

null

Machine Learning

machine-learning

Machine Learning

51,320

Ayindrilla Maiti

null

85b5cab7ae27

ayindrilla.maiti

2

1

20,181,104

null

0

# Importing tensorflow import tensorflow as tf # importing the data from tensorflow.examples.tutorials.mnist import input_data # Importing some more libraries import matplotlib.pyplot as plt from numpy import loadtxt import numpy as np from pylab import rcParams # reading the data mnist = input_data.read_data_sets("/tmp/data/", one_hot=True) X_train = mnist.train.images X_test = mnist.test.images # deciding how big we want our print out to be rcParams['figure.figsize'] = 20,20 # looping through the first 10 test images and printing them out for i in range(10): plt.subplot(1, 10, i+1) plt.imshow(X_test[i].reshape(28,28), cmap='Greys') plt.axis('off') plt.show() # printing out the noisy images for i in range(10): plt.subplot(1, 10, i+1) plt.imshow(X_test_noisy[i].reshape(28,28), cmap='Greys') plt.axis('off') plt.show() # creating the noise matrix n_rows = X_test.shape[0] n_cols = X_test.shape[1] mean = 0.5 stddev = 0.3 noise = np.random.normal(mean, stddev, (n_rows, n_cols)) # creating the noisy test data by adding X_test with noise X_test_noisy = X_test + noise # Deciding how many nodes each layer should have n_nodes_inpl = 784 #encoder n_nodes_hl1 = 32 #encoder n_nodes_hl2 = 32 #decoder n_nodes_outl = 784 #decoder # first hidden layer has 784*32 weights and 32 biases hidden_1_layer_vals = { 'weights':tf.Variable(tf.random_normal([n_nodes_inpl,n_nodes_hl1])), 'biases':tf.Variable(tf.random_normal([n_nodes_hl1])) } # second hidden layer has 32*32 weights and 32 biases hidden_2_layer_vals = { 'weights':tf.Variable(tf.random_normal([n_nodes_hl1, n_nodes_hl2])), 'biases':tf.Variable(tf.random_normal([n_nodes_hl2])) } # second hidden layer has 32*784 weights and 784 biases output_layer_vals = { 'weights':tf.Variable(tf.random_normal([n_nodes_hl2,n_nodes_outl])), 'biases':tf.Variable(tf.random_normal([n_nodes_outl])) } # image with shape 784 goes in input_layer = tf.placeholder('float', [None, 784]) # multiply output of input_layer wth a weight matrix and add biases layer_1 = tf.nn.sigmoid( tf.add(tf.matmul(input_layer,hidden_1_layer_vals['weights']), hidden_1_layer_vals['biases'])) # multiply output of layer_1 wth a weight matrix and add biases layer_2 = tf.nn.sigmoid( tf.add(tf.matmul(layer_1,hidden_2_layer_vals['weights']), hidden_2_layer_vals['biases'])) # multiply output of layer_2 wth a weight matrix and add biases output_layer = tf.matmul(layer_2,output_layer_vals['weights']) + output_layer_vals['biases'] # output_true shall have the original image for error calculations output_true = tf.placeholder('float', [None, 784]) # define our cost function meansq = tf.reduce_mean(tf.square(output_layer - output_true)) # define our optimizer learn_rate = 0.1 # how fast the model should learn optimizer = tf.train.AdagradOptimizer(learn_rate).minimize(meansq) # initialising stuff and starting the session init = tf.global_variables_initializer() sess = tf.Session() sess.run(init) # defining batch size, number of epochs and learning rate batch_size = 100 # how many images to use together for training hm_epochs =1000 # how many times to go through the entire dataset tot_images = X_train.shape[0] # total number of images # running the model for a 10000 epochs taking 100 images in batches # total improvement is printed out after each epoch for epoch in range(hm_epochs): epoch_loss = 0 # initializing error as 0 for i in range(int(tot_images/batch_size)): epoch_x = X_train[ i*batch_size : (i+1)*batch_size ] _, c = sess.run([optimizer, meansq],\ feed_dict={input_layer: epoch_x, \ output_true: epoch_x}) epoch_loss += c print('Epoch', epoch, '/', hm_epochs, 'loss:',epoch_loss) # pick any image any_image = X_test_noisy[234] # run it though the autoencoder output_any_image = sess.run(output_layer,\ feed_dict={input_layer:[any_image]})

49

null

2018-03-12

2018-03-12 20:14:29

2018-03-15

2018-03-15 21:20:20

1

false

en

2018-03-17

2018-03-17 19:37:04

1

1e2e62932837

3.381132

7

1

0

Previously I had written sort of a tutorial on building a simple autoencoder in tensorflow. In that tutorial I had used the autoencoder for…

1

Denoising MNIST images using an Autoencoder and Tensorflow in python Previously I had written sort of a tutorial on building a simple autoencoder in tensorflow. In that tutorial I had used the autoencoder for dimensionality reduction. Check it out if you want to. It has a much more detailed explanation on how to build the autoencoder itself. Here, I’ll use the exact same model to show another use of autoencoders — denoising images. So let’s get started. I’ll use the famous MNIST handwriting data here. I’ll import both the train and test set. I’ll train our model using the train set first. Then i’ll put in noisy test data and see if cleaner images come out. Now mnist doesn’t give us noisy data — no worries, we’ll just make some ourselves. This sweet peice of code below prints out ten test images alogwith their noisy version. The reason we need to reshape the images before printing is because we’re supplied with an array of 784 for each observation, which is just 28*28 pixels. Now i’ll add some noise to the test images. I’ll just create an array as same as the test dataset and just add that to the original data. Now i’ll build my autoencoder. I’m not going into details about the model. Im using the same model as I’ve used here. If you’re interested please have a look at that tutorial where I’ve explained the model somewhat. Its a lot of code but all that it is, is a neural network with a 784 node input layer, two 32 node hidden layers and again a 784 nodes output layer. An image goes in, an image come out. Now I’ll train this model with the training images. I’m doing a couple thousand epochs here which means a couple thousand passes through the whole training set, with a batch size of hundred, that is training the net with a hundred images at a time. My computers too slow, but you’re probably gonna have to run way more epochs to see some good results. Done! Its a good idea to do some validation every once in a while using the test set. Now for the see if the autoencoder actually works as a denoiser. I’ll put in a couple of noisy images and see what comes out.

Denoising MNIST images using an Autoencoder and Tensorflow in python

8

denoising-images-using-an-autoencoder-using-tensorflow-in-python-1e2e62932837

2018-03-23

2018-03-23 16:18:46

https://medium.com/s/story/denoising-images-using-an-autoencoder-using-tensorflow-in-python-1e2e62932837

false

843

null

Machine Learning

machine-learning

Machine Learning

51,320

Soumya Ghosh

null

258f683365d7

connectwithghosh

47

4

20,181,104

null

0

null

0

92dfe0f73116

2018-03-13

2018-03-13 07:34:05

2018-03-13

2018-03-13 07:36:47

1

false

en

2018-03-15

2018-03-15 09:28:14

3

1e2faf3a6479

3.4

2

0

The recruitment landscape is no longer what it was 10 years ago. It has changed drastically from the time companies would put out job…

5

Param.ai launches stand-alone hiring platform The recruitment landscape is no longer what it was 10 years ago. It has changed drastically from the time companies would put out job advertisements, waiting for candidates to apply. In the current scenario, a company’s success in meeting their business growth objectives is dependent on its ability to find, attract and hire talent. The war for talent is real: Till recently, recruitment was reactive, with recruiters working on requisitions, responding to a present need. However, we are seeing recruitment becoming more and more strategic and proactive, with recruiters increasingly anticipating future needs and working towards it to build the workforce. Photo by freestocks.org on Unsplash The fact that the current job-market is candidate-driven means that candidates now have the luxury of being selective in their job search. At any given point in time, top prospects for a given job are being approached by multiple recruiters at the same time. Hence the age-old practice of sending bulk ‘cold emails’ or posting a job whenever a new position is open will not help the recruiters of today. Recruiters will also have to start working on building and nurturing talent ‘pipelines’ in anticipation of future needs. Various bottlenecks in the recruitment process: Sourcing: Sourcing is one of the most important aspects of the recruitment cycle, especially for hard-to-fill roles or those with in-demand skills. While sourcing is fundamental for successful hiring for these roles, with the number of priorities and tasks vying for the recruiters’ time, recruiters aren’t able to devote enough time to this task. Increased cost of hiring: Since sourcing efforts do not give the returns expected, recruiters turn to channels such as external staffing firms to fill open positions. This results in increased costs of hiring. Manual applicant screening: Recruiters today are dealing with large volumes of applications for their open positions. A report states that on average 75–80% of the applications coming in, in response to a job advertisement are not relevant. This means that recruiters spend a substantial amount of time screening through applications (most of which are not relevant) to find the relevant ones. Losing candidates for lack of prioritization methods: Also, since recruiters do not have methods to prioritize candidates from the sourcing bucket, by the time they get to the best candidate in the lot, the candidate would have already gone off the job market (On average, top candidates stay on the market for about 10 days) Inadequate candidate engagement: The recruiter is the face of the company for the applicant. And in the war for talent, the first-mover advantage is substantial, if not the deciding factor. And given how top candidates tend to go off the job market in about 10 days’ time, it is important that they are engaged by recruiters, and well in time. However, given the operational load because of tasks like manual screening of resumes, recruiters are not left with a lot of time to engage their candidates. Leverage AI technology in recruitment The Param platform helps recruiters with various aspects of the recruiting process. The Platform: -Stack-ranks the candidates/prospects from all the sources for a given job, giving the recruiter a stack-ranked list of candidates to start engaging with, helping them increase their conversions. -Has the AI sourcing feature which continuously mines the database and recommends candidates to multiple potential job matches, helping optimize the utilization of candidate data within the platform. -Enables engaging applicants and prospects through the Param platform instead of individual inboxes and various networking sites. This way, team members can track all correspondence history with the candidate in one location, and can work in collaboration. -Enables users to create customizable career pages which can be used by recruiters to post the jobs on social media, and collects all the applications received. All applicants from job boards (which form bulk of the applicant volume) can be routed to Param, where the top prospects are surfaced for the recruiter to start engaging with. Optimizing the usage of ATS data: Param can be used by companies to track all their sourcing efforts in one location. Companies can also choose to optimize the usage of their ATS data. This can be done by migrating the ATS data into the Param platform, post which the candidates can be recommended for various open jobs. With the large volume of resumes coming in from various sources (like job boards and career site) added into Param, recruiters no longer need to screen each resume manually for each open position. Param helps cut through the noise and ranks resumes based on the fitment for the job selected. Not only that, Param also auto-recommends other jobs for which the candidate is better suited, helping recruiters align candidates to the best-suited jobs. This way, the applicant data is utilized optimally. With the intelligence added to the recruitment process by Param, recruiters can hire better and faster. If you’d like to know how your company’s hiring can benefit from Param, drop us a note at hello@param.ai. We’d love to talk to you.

Param.ai launches stand-alone hiring platform

100

param-ai-launches-stand-alone-hiring-platform-1e2faf3a6479

2018-05-29

2018-05-29 10:06:44

https://medium.com/s/story/param-ai-launches-stand-alone-hiring-platform-1e2faf3a6479

false

848

Param.ai is an intelligent platform built for recruiters to discover new candidates and engage with them using smart algorithms.

null

hirewithparam

null

hirewithparam

hello@param.ai

hirewithparam

RECRUITING,RECRUITERS,JOBS,NLP,DATA SCIENCE

hirewithparam

Recruiting

recruiting

Recruiting

15,454

Param.ai

An intelligent recruiting platform to help companies Discover, Match, Engage and Retarget potential candidates at a scale. #SuperchargeyourATS

19cd8a6a8c71

hirewithparam

28

14

20,181,104

null

0

null

0

9d4567475f36

2018-03-10

2018-03-10 05:06:09

2018-03-10

2018-03-10 20:31:10

10

false

en

2018-03-18

2018-03-18 15:48:35

2

1e30fbc91252

2.317925

2

0

“Parenclitic Network” What is it ? How does it help in Outlier Detection?

4

Understanding — Credit card transaction fraud feature extraction using “Parenclitic Networks” Paper Source:- https://arxiv.org/pdf/1706.01953.pdf “Parenclitic Network” What is it ? How does it help in Outlier Detection? The fundamental idea is to use graph network measures as features in addition to the normal transnational features. Also, there is one difference between the traditional graph networks and parenclitic networks. Nodes are associated with the features, but the links or edges are weighted according to the deviation of the transaction value from the values of two features. Below picture explains it a little bit more. Sourrce https://arxiv.org/pdf/1304.1896.pdf This representation identifies which parts and relationships have deviated by the transaction. Data is projected on to every pair of features and a line is fit using linear regression. When an unlabeled data comes in you calculate the deviation and use it as to create the edge. Let us dive into the paper I agree, ordinary features are only useful when the fraud transaction standout on their own. There is always a motivation to figure out new features. [Pg.1] Parenclitic networks are borrowed from the medical field. They are used to extract new features and augment them with the transaction features. [Pg.1] Explanation of the parenclitic network representation in the paper. [Pg.2] A list of the network features [Pg.3] A list of features from the credit card transactions [Pg.3] A little bit about the data set. [Pg.3] A fully connected layers with sigmoid activation activation is used. [Pg.4] Rest of it the usual stuff normally followed. Results show the ROC with AUC along with various comparisons. Details in the original paper https://arxiv.org/pdf/1706.01953.pdf At the end I had few questions which I think were not clear from the paper for me. Here they are

Understanding — Credit card transaction fraud feature extraction using “Parenclitic Networks”

21

credit-card-transaction-fraud-feature-extraction-using-parenclitic-networks-1e30fbc91252

2018-03-18

2018-03-18 17:39:14

https://medium.com/s/story/credit-card-transaction-fraud-feature-extraction-using-parenclitic-networks-1e30fbc91252

false

283

This is a collection of all my stories which focus on feature engineering in machine learning.

null

It’s all about feature engineering

anil.kemisetti@neocortex.com

its-all-about-feature-engineering

MACHINE LEARNING,FEATURE ENGINEERING,DEEP LEARNING,DATA SCIENCE,ENTERPRISE APPLICATION

anilkemisetti

Machine Learning

machine-learning

Machine Learning

51,320

Anil Kemisetti

AI & Deep Learning Enthusiast; Life Long Learner

e77eb74a42e

anilkemisetti

11

8

20,181,104

null

0

# From tensorflow/models/research/ protoc object_detection/protos/*.proto --python_out=.

1

e89b5902490b

2017-11-04

2017-11-04 00:23:02

2017-11-04

2017-11-04 00:52:42

1

false

pt

2017-11-04

2017-11-04 02:15:23

20

1e311f3e85c6

4.807547

29

1

0

Artigo traduzido do original “Building a Real-Time Object Recognition App with Tensorflow and OpenCV” de autoria do Dat Tran.

4

Construindo uma App para Reconhecimento de Objetos em Tempo Real com Tensorflow e OpenCV Artigo traduzido do original “Building a Real-Time Object Recognition App with Tensorflow and OpenCV” de autoria do Dat Tran. Tradução: Simony Cantanhede Nesta matéria demonstrarei, de forma simples, o passo-a-passo para a construção de uma aplicação capaz de reconhecer objetos em tempo real utilizando a nova API de detecção de objetos da Tensorflow e o OpenCV em Python 3 (mais especificamente, com a versão 3.5 do Python). Me concentrarei na descrição dos desafios que tive de superar quando estava tentando construir a minha própria aplicação, para que você, leitor, fique esperto quando estiver colocando a mão na massa. Você pode encontrar o código completo aqui no meu repositório. E aqui está uma demonstração da app em execução: Motivação Há pouco tempo a Google lançou sua nova API de detecção de objetos TensorFlow. Esta primeira versão contém: Alguns modelos pré-treinados (é preciso ressaltar que houve um cuidado especial em fornecer modelos leves, que podem rodar sem problemas em dispositivos móveis); Um exemplo de como montar um “Jupyter Notebook”[1] de um destes modelos disponíveis; E alguns scripts bem práticos que podem ser utilizados para re-treinar os modelos disponíveis usando um dataset de sua própria autoria, caso você deseje fazê-lo. Tendo em vista todas essas possibilidades, eu não via a hora de colocar minhas mãos em todo esse material incrível e dedicar alguma parte de meu tempo para construir algum tipo de aplicação capaz de reconhecer objetos em tempo real. Aplicação Primeiro, eu baixei o repositório de modelos do TensorFlow e então dei uma boa olhada no “jupyter notebook” que o acompanha: basicamente, ele descreve todas as etapas para o uso de um modelo pré-treinado. Neste exemplo, em específico, eles usaram o modelo “SSD com Mobilenet”, mas há muitos outros modelos disponíveis para download em um repositório que eles chamaram de “Tensorflow detection model zoo”. Os modelos deste repositório são treinados com o dataset COCO e variam dependendo da velocidade do modelo (lento, médio e rápido) e de sua performance (mAP — baseada na precisão média). Em seguida, eu executei o exemplo. Vale ressaltar o quão bem documentado é o exemplo disponibilizado, tanto que, é bem fácil entender logo de cara suas principais funções: 1. Importar os pacotes necessários como o TensorFlow, PIL, etc. ; 2. Definir algumas variáveis, por exemplo: número de classe, nome do modelo, etc. ; 3. Baixar o modelo pré-computado (extensão .pb — “protobuf” [2]) e carregá-lo na memória ; 4. Carregar códigos auxiliares, por exemplo um índice para rotular o tradutor ; 5. O próprio código de detecção já contendo duas imagens para teste. OBS: Antes de executar o exemplo, dê uma olhada nas observações de configuração, especialmente, naquelas que se referem à compilação do “protobuf”. Se você não rodar este comando aqui embaixo, o exemplo não vai funcionar: Depois de ter executado o exemplo e começar a ter alguma de noção de como ele funcionava, eu modifiquei o código original da seguinte maneira: Excluí a seção de download do modelo; Excluí também a dependência da biblioteca PIL, pois os fluxos de vídeo no OpenCV já estão no formato de arrays numpy. Além disso, PIL também é uma sobrecarga muito grande, especificamente ao usá-lo para leitura de imagens nos streams de vídeo; Nenhuma declaração “com” para a sessão TensorFlow, pois esta é uma sobrecarga enorme, principalmente quando é necessário reiniciar a sessão depois de cada stream. Daí eu utilizei o OpenCV para fazer a conexão entre o exemplo modificado e minha webcam. Há vários tutoriais disponíveis em vários cantos da internet que explicam direitinho como fazer essa conexão, além da documentação oficial no próprio site do OpenCV. Geralmente, a implementação básica de exemplos OpenCV não é otimizada, por exemplo, algumas funções em OpenCV estão fortemente vinculadas à entrada e saída de dados. Por causa disso, eu tive que buscar algumas soluções para contornar este problema: A leitura de frames através da webcam provoca um intenso fluxo de entrada e saída de dados. Minha ideia foi transferir essa responsabilidade para um novo processo Python que utilizasse uma biblioteca de multiprocessamento. Mas, infelizmente, isto não funcionou. Havia algumas explicações no Stackoverflow sobre o porque uma solução como esta não funcionaria, mas não consegui me aprofundar muito no assunto. Felizmente, encontrei um exemplo muito bom do Adrian Rosebrock em seu site “pyimagesearch” usando threading em vez de multiprocessamento, o que melhorou bastante o meu fps. Por sinal, se você quiser saber a diferença entre multiprocessamento e threading, no Stackoverflow você vai encontrar ótimas explicações. O carregamento do modelo pré-computado na memória a cada vez que a aplicação inicializa provoca uma enorme sobrecarga. Primeiro eu tentei solucionar utilizando uma sessão TensorFlow para cada execução, mas o carregamento continuou muito lento. Então o que eu fiz para solucionar de vez o problema? A solução é bastante simples. Neste caso, usei a biblioteca de multiprocessamento para dividir a pesada carga de trabalho da detecção de objetos em múltiplos processos. O gatilho inicial da aplicação continuará lento, pois cada um desses processos precisa carregar o modelo na memória e iniciar a sessão do TensorFlow, mas, depois disso, o paralelismo nos ajudará. Reduzir a largura e a altura dos frames no stream de vídeo também ajudou a melhorar bastante o fps. OBS: Se você estiver em um Mac OSX como eu e estiver usando o OpenCV 3.1, pode haver uma chance de o VideoCapture da OpenCV falhar depois de um tempo. Voltar ao OpenCV 3.0 resolverá o problema. Conclusão e considerações futuras Me dê um 👏 se você gostou desta matéria :) Faça o download do código e experimente você mesmo. E definitivamente dê uma olhada na API Tensorflow Object Detection. É muito prático e simples desde o primeiro contato. A próxima coisa que eu quero tentar é treinar meu próprio conjunto de dados com a API e também usar os modelos pré-treinados para outras aplicações que tenho em mente. Eu também não estou totalmente satisfeito com o desempenho desta aplicação ainda. A taxa de fps ainda não está otimizada. Há muitos gargalos no OpenCV que ainda não consegui contornar, mas há alternativas que eu posso experimentar, tais como usar o WebRTC. Entretanto, esta alteração precisaria ser baseada em web. Além disso, estou pensando em usar chamadas de método assíncronas (async) para melhorar minha taxa de fps. Fique ligado nas possíveis atualizações! [1] “Jupyter Notebook” é uma aplicação web open-source integrante do Projeto Jupyter cujo objetivo é permitir a criação e o compartilhamento de documentos que contenham a um só tempo código-fonte, equações, visualizações e textos narrativos. Um “jupyter notebook” suporta mais de 40 linguagens de programação; pode ser compartilhado por meio de e-mail, dropbox e GitHub; pode gerar saídas interativas em diversos formatos, desde HTML até imagens e vídeos; e também suporta ferramentas de big data como o Apache Spark, R e Scala, além do TensorFlow. Para saber mais sobre o projeto clique aqui. [2] Um “protobuffer” (ou “protocol buffer”) é um mecanismo extensível criado pela Google capaz de executar em várias plataformas e em várias linguagens com o objetivo de serializar dados estruturados — algo como o XML, porém mais rápido, leve e simples. Saiba mais em: https://developers.google.com/protocol-buffers/ .

Construindo uma App para Reconhecimento de Objetos em Tempo Real com Tensorflow e OpenCV

118

construindo-um-app-para-reconhecimento-de-objetos-em-tempo-real-com-tensorflow-e-opencv-1e311f3e85c6

2018-06-14

2018-06-14 19:44:27

https://medium.com/s/story/construindo-um-app-para-reconhecimento-de-objetos-em-tempo-real-com-tensorflow-e-opencv-1e311f3e85c6

false

1,221

Inteligencia artificial, aprendizagem de máquina e análise de dados

null

Machina Sapiens

null

machina-sapiens

ARTIFICIAL INTELLIGENCE,MACHINE LEARNING,INTELIGENCIA ARTIFICIAL,APRENDIZAEM DE MÁQUINA

null

Inteligencia Artificial

inteligencia-artificial

Inteligencia Artificial

1,614

Simony Cantanhede

null

88dba9664c84

simonycanunes

14

10

20,181,104

null

0

null

0

33272dbbd858

2017-12-07

2017-12-07 05:14:02

2017-12-09

2017-12-09 02:20:17

9

false

en

2018-07-06

2018-07-06 20:30:03

19

1e317ae4587d

7.09434

959

20

3

Simplifying a complex algorithm

3

Gradient Boosting from scratch Simplifying a complex algorithm Motivation Although most of the Kaggle competition winners use stack/ensemble of various models, one particular model that is part of most of the ensembles is some variant of Gradient Boosting (GBM) algorithm. Take for an example the winner of latest Kaggle competition: Michael Jahrer’s solution with representation learning in Safe Driver Prediction. His solution was a blend of 6 models. 1 LightGBM (a variant of GBM) and 5 Neural Nets. Although his success is attributed to the semi-supervised learning that he used for the structured data, but gradient boosting model has done the useful part too. Even though GBM is being used widely, many practitioners still treat it as complex black-box algorithm and just run the models using pre-built libraries. The purpose of this post is to simplify a supposedly complex algorithm and to help the reader to understand the algorithm intuitively. I am going to explain the pure vanilla version of the gradient boosting algorithm and will share links for its different variants at the end. I have taken base DecisionTree code from fast.ai library (fastai/courses/ml1/lesson3-rf_foundations.ipynb) and on top of that, I have built my own simple version of basic gradient boosting model. Brief description for Ensemble, Bagging and Boosting When we try to predict the target variable using any machine learning technique, the main causes of difference in actual and predicted values are noise, variance, and bias. Ensemble helps to reduce these factors (except noise, which is irreducible error) An ensemble is just a collection of predictors which come together (e.g. mean of all predictions) to give a final prediction. The reason we use ensembles is that many different predictors trying to predict same target variable will perform a better job than any single predictor alone. Ensembling techniques are further classified into Bagging and Boosting. Bagging is a simple ensembling technique in which we build many independent predictors/models/learners and combine them using some model averaging techniques. (e.g. weighted average, majority vote or normal average) We typically take random sub-sample/bootstrap of data for each model, so that all the models are little different from each other. Each observation is chosen with replacement to be used as input for each of the model. So, each model will have different observations based on the bootstrap process. Because this technique takes many uncorrelated learners to make a final model, it reduces error by reducing variance. Example of bagging ensemble is Random Forest models. Boosting is an ensemble technique in which the predictors are not made independently, but sequentially. This technique employs the logic in which the subsequent predictors learn from the mistakes of the previous predictors. Therefore, the observations have an unequal probability of appearing in subsequent models and ones with the highest error appear most. (So the observations are not chosen based on the bootstrap process, but based on the error). The predictors can be chosen from a range of models like decision trees, regressors, classifiers etc. Because new predictors are learning from mistakes committed by previous predictors, it takes less time/iterations to reach close to actual predictions. But we have to choose the stopping criteria carefully or it could lead to overfitting on training data. Gradient Boosting is an example of boosting algorithm. Fig 1. Ensembling Fig 2. Bagging (independent models) & Boosting (sequential models). Reference: https://quantdare.com/what-is-the-difference-between-bagging-and-boosting/ Gradient Boosting algorithm Gradient boosting is a machine learning technique for regression and classification problems, which produces a prediction model in the form of an ensemble of weak prediction models, typically decision trees. (Wikipedia definition) The objective of any supervised learning algorithm is to define a loss function and minimize it. Let’s see how maths work out for Gradient Boosting algorithm. Say we have mean squared error (MSE) as loss defined as: We want our predictions, such that our loss function (MSE) is minimum. By using gradient descent and updating our predictions based on a learning rate, we can find the values where MSE is minimum. So, we are basically updating the predictions such that the sum of our residuals is close to 0 (or minimum) and predicted values are sufficiently close to actual values. Intuition behind Gradient Boosting The logic behind gradient boosting is simple, (can be understood intuitively, without using mathematical notation). I expect that whoever is reading this post might be familiar with simple linear regression modeling. A basic assumption of linear regression is that sum of its residuals is 0, i.e. the residuals should be spread randomly around zero. Fig 3. Sample random normally distributed residuals with mean around 0 Now think of these residuals as mistakes committed by our predictor model. Although, tree-based models (considering decision tree as base models for our gradient boosting here) are not based on such assumptions, but if we think logically (not statistically) about this assumption, we might argue that, if we are able to see some pattern of residuals around 0, we can leverage that pattern to fit a model. So, the intuition behind gradient boosting algorithm is to repetitively leverage the patterns in residuals and strengthen a model with weak predictions and make it better. Once we reach a stage that residuals do not have any pattern that could be modeled, we can stop modeling residuals (otherwise it might lead to overfitting). Algorithmically, we are minimizing our loss function, such that test loss reach its minima. In summary, • We first model data with simple models and analyze data for errors. • These errors signify data points that are difficult to fit by a simple model. • Then for later models, we particularly focus on those hard to fit data to get them right. • In the end, we combine all the predictors by giving some weights to each predictor. A more technical quotation of the same logic is written in Probably Approximately Correct: Nature’s Algorithms for Learning and Prospering in a Complex World, “The idea is to use the weak learning method several times to get a succession of hypotheses, each one refocused on the examples that the previous ones found difficult and misclassified. … Note, however, it is not obvious at all how this can be done” Steps to fit a Gradient Boosting model Let’s consider simulated data as shown in scatter plot below with 1 input (x) and 1 output (y) variables. Fig 4. Simulated data (x: input, y: output) Data for above shown plot is generated using below python code: Code Chunk 1. Data simulation 1. Fit a simple linear regressor or decision tree on data (I have chosen decision tree in my code) [call x as input and y as output] Code Chunk 2. (Step 1)Using decision tree to find best split (here depth of our tree is 1) 2. Calculate error residuals. Actual target value, minus predicted target value [e1= y - y_predicted1 ] 3. Fit a new model on error residuals as target variable with same input variables [call it e1_predicted] 4. Add the predicted residuals to the previous predictions [y_predicted2 = y_predicted1 + e1_predicted] 5. Fit another model on residuals that is still left. i.e. [e2 = y - y_predicted2] and repeat steps 2 to 5 until it starts overfitting or the sum of residuals become constant. Overfitting can be controlled by consistently checking accuracy on validation data. Code Chunk 3. (Steps 2 to 5) Calculate residuals and update new target variable and new predictions To aid the understanding of the underlying concepts, here is the link with complete implementation of a simple gradient boosting model from scratch. [Link: Gradient Boosting from scratch] Shared code is a non-optimized vanilla implementation of gradient boosting. Most of the gradient boosting models available in libraries are well optimized and have many hyper-parameters. Visualization of working Gradient Boosting Tree Blue dots (left) plots are input (x) vs. output (y) • Red line (left) shows values predicted by decision tree • Green dots (right) shows residuals vs. input (x) for ith iteration • Iteration represent sequential order of fitting gradient boosting tree Fig 5. Visualization of gradient boosting predictions (First 4 iterations) Fig 6. Visualization of gradient boosting predictions (18th to 20th iterations) We observe that after 20th iteration , residuals are randomly distributed (I am not saying random normal here) around 0 and our predictions are very close to true values. (iterations are called n_estimators in sklearn implementation). This would be a good point to stop or our model will start overfitting. Let’s see how our model look like for 50th iteration. Fig 7. Visualization of gradient boosting prediction (iteration 50th) We see that even after 50th iteration, residuals vs. x plot look similar to what we see at 20th iteration. But the model is becoming more complex and predictions are overfitting on the training data and are trying to learn each training data. So, it would have been better to stop at 20th iteration. Python code snippet used for plotting all the above figures. Code Chunk 4. Plotting predictions and residuals (fed in 1st code chunk’s loop) I hope that this blog helped you to get basic intuition behind how gradient boosting works. To understand gradient boosting for regression in detail, I would strongly recommend you to read this amazing article by the faculty at University of San Francisco, Terence Parr (Creator of the ANTLR parser generator) and Jeremy Howard (Founding researcher at fast.ai) : How to explain gradient boosting. More useful resources My github repo and kaggle kernel link for GBM from scratch: https://www.kaggle.com/grroverpr/gradient-boosting-simplified/ https://nbviewer.jupyter.org/github/groverpr/Machine-Learning/blob/master/notebooks/01_Gradient_Boosting_Scratch.ipynb A detailed and intuitive explanation of gradient boosting: How to explain gradient boosting by Terence Parr and Jeremy Howard Fast.ai github repo link for DecisionTree from scratch (Massive ML/DL related resources): https://github.com/fastai/fastai Video by Alexander Ihler. This video really helped me to build my understanding. 4. A Kaggle Master Explains Gradient Boosting: Ben Gorman A Kaggle Master Explains Gradient Boosting If linear regression was a Toyota Camry, then gradient boosting would be a UH-60 Blackhawk Helicopter. A particular…blog.kaggle.com 5. Widely used GBM algorithms: XGBoost || Lightgbm || Catboost || sklearn.ensemble.GradientBoostingClassifier

Gradient Boosting from scratch

5,236

gradient-boosting-from-scratch-1e317ae4587d

2018-07-06

2018-07-06 20:30:03

https://medium.com/s/story/gradient-boosting-from-scratch-1e317ae4587d

false

1,562

Highlights from Machine Learning Research, Projects and Learning Materials. From and For ML Scientists, Engineers an Enthusiasts.

null

ML Review

medium@mlreview.com

mlreview

MACHINE LEARNING,DATA SCIENCE,COMPUTER VISION,ARTIFICIAL INTELLIGENCE,DEEP LEARNING

ML_review

Machine Learning

machine-learning

Machine Learning

51,320

Prince Grover

Machine Learning Engineer @Manifold.ai, USF-MSDS and IIT-Roorkee Alumnus (Twitter: @groverpr4)

2f052b054cee

pgrover3

1,780

90

20,181,104

null

0

null

0

null

2018-07-23

2018-07-23 01:20:24

2018-07-23

2018-07-23 19:47:14

0

false

en

2018-08-01

2018-08-01 19:11:10

5

1e3184b1450e

0.935849

0

I did a talk on why Javascript is going to be important to the future of machine learning at Mesosphere last month. Thanks again to Chris…

3

tensorflow.js: model training and predicting at the edge I did a talk on why Javascript is going to be important to the future of machine learning at Mesosphere last month. Thanks again to Chris Fregly for setting things up! TL/DR: for data science, most people prefer python wrappers to simplify dealing with more complicated tools. Python has historically been a great wrapper language for dealing with Tensorflow and Javascript has been looked down upon as a toy language. Basically, I agree that Javascript has a lot of faults, but on the other hand has a massive number of developers working on it daily. What this means in practice is that now that the team at Google has done the hard work of allowing us to write machine learning scripts in Javascript, all that remains is for people to start converting their workflows to the new language. Under the hood, all these scripts generate machine code in some form or another. Time will tell, but things like keras.js show that intrinsically, there’s nothing that you can do in other programming languages that can’t be done in Javascript as well. I’m not jumping programming languages quite yet, but long term it is my belief that the large number of developers that using Tensorflow.js will allow to start experimenting with machine learning will ultimately mean that soon Javascript could very well be the dominant language in this space. It’s worth your time to click around the tutorials at the very least!

tensorflow.js: model training and predicting at the edge

0

tensorflow-js-model-training-and-predicting-at-the-edge-1e3184b1450e

2018-08-01

2018-08-01 19:11:10

https://medium.com/s/story/tensorflow-js-model-training-and-predicting-at-the-edge-1e3184b1450e

false

248

null

Machine Learning

machine-learning

Machine Learning

51,320

brett koonce

cto, quarkworks

ba03f25005e5

brett_koonce

9

22

20,181,104

null

0

null

0

null

2018-07-23

2018-07-23 13:11:41

2018-07-24

2018-07-24 10:10:17

0

false

en

2018-07-24

2018-07-24 10:29:25

0

1e31967ef23c

2.116981

2

0

The struggle to help others grow

3

The 3 People That Have Helped Me At findworka Academy The struggle to help others grow… My journey at findworka academy started one beautiful night when a wonderful friend Sarafa Olalekan sent me a link for the academy application. I immediately filled the form and picked the data science track due to some reasons reserved for another post. As hyperactive and confident as I have always been I made friends at the interview but sadly, most of them were in the back-end track and not data science. I think it is quite impossible to pick three people from the data science track that have helped me since we began the journey, reason been that the data science class is more of a family than a class. We have worked together to help each member of the family in LLG-ing (learning, laughing and growing). So for me, I would rather reluctantly pick the top three persons that have helped me so far. Please note that this list is subjected to change in the coming weeks as it is our thing in the family (data science track, findworka academy) to battle to unblock colleagues and also share useful resources. That been said, let’s begin! The first person on the leaderboard is Princess Egbuna, a software developer with over 5 years programming experience using the python language and other technologies. This Dev Princess came to my rescue ( ironic right?) in the first class, UNIX commands for Data Science. I have never used the Bash before and didn’t know where to start when we were told to run some commands and I hope you know how frustrating it can be when it’s your first day in school and you don’t seem to have all the necessary materials…., she pointed out that I could use the Git Bash on my laptop (which I had downloaded due to the VS Code prompt )and further showed me how to move files from a particular directory to another. Boy! that was a relief, it made my learning experience for that day fantastic and I was willing myself to help others. The next person on the list is Yusuf ogunbiyi, the Agricultural economist who seems to know and have done more data analytics than any other person in the family (data science track). He is always there when you need the extra explanation to understand a particular concept with lucid and sometimes weird illustrations (Lol). He has shared valuable links to resources and opportunities relating to data science opportunities and has a very good sense of humour ( I’ll still beat him to it any day though). The third person on the leaderboard is the ever-passionate Bolu Olufade. A very energetic python programmer, very willing to push herself to learn and develop. the most fascinating attribute about this wonderful lady is the energy to give guys a run for their money when it comes to solving challenges and getting things done! She keeps me on my toes ready to learn and beat her to solving challenges. Also, she is my favorite candidate for my savage exercise (tongue out). That has been a part of my wonderful experience so far at findworka Academy . See you soon, when I get to paint the other wonderful personalities of the family to you. Hope it was what your time.

The 3 People That Have Helped Me At findworka Academy

2

the-3-people-that-have-helped-me-at-findworka-academy-1e31967ef23c

2018-07-24

2018-07-24 10:29:25

https://medium.com/s/story/the-3-people-that-have-helped-me-at-findworka-academy-1e31967ef23c

false

561

null

Data Science

data-science

Data Science

33,617

Adelanwa A. Adedeji

null

f231daea5884

saubandexter

3

17

20,181,104

null

0

null

0

null

2018-07-15

2018-07-15 03:58:35

2018-07-15

2018-07-15 04:05:57

0

false

en

2018-07-15

2018-07-15 04:06:45

0

1e341bd50a70

4.366038

0

To our shareowners:

1

2010 Amazon Letter to Shareholders To our shareowners: Random forests, naïve Bayesian estimators, RESTful services, gossip protocols, eventual consistency, data sharding, anti-entropy, Byzantine quorum, erasure coding, vector clocks … walk into certain Amazon meetings, and you may momentarily think you’ve stumbled into a computer science lecture. Look inside a current textbook on software architecture, and you’ll find few patterns that we don’t apply at Amazon. We use high-performance transactions systems, complex rendering and object caching, workflow and queuing systems, business intelligence and data analytics, machine learning and pattern recognition, neural networks and probabilistic decision making, and a wide variety of other techniques. And while many of our systems are based on the latest in computer science research, this often hasn’t been sufficient: our architects and engineers have had to advance research in directions that no academic had yet taken. Many of the problems we face have no textbook solutions, and so we — happily — invent new approaches. Our technologies are almost exclusively implemented as services: bits of logic that encapsulate the data they operate on and provide hardened interfaces as the only way to access their functionality. This approach reduces side effects and allows services to evolve at their own pace without impacting the other components of the overall system. Service-oriented architecture — or SOA — is the fundamental building abstraction for Amazon technologies. Thanks to a thoughtful and far-sighted team of engineers and architects, this approach was applied at Amazon long before SOA became a buzzword in the industry. Our e-commerce platform is composed of a federation of hundreds of software services that work in concert to deliver functionality ranging from recommendations to order fulfillment to inventory tracking. For example, to construct a product detail page for a customer visiting Amazon.com, our software calls on between 200 and 300 services to present a highly personalized experience for that customer. State management is the heart of any system that needs to grow to very large size. Many years ago, Amazon’s requirements reached a point where many of our systems could no longer be served by any commercial solution: our key data services store many petabytes of data and handle millions of requests per second. To meet these demanding and unusual requirements, we’ve developed several alternative, purpose-built persistence solutions, including our own key-value store and single table store. To do so, we’ve leaned heavily on the core principles from the distributed systems and database research communities and invented from there. The storage systems we’ve pioneered demonstrate extreme scalability while maintaining tight control over performance, availability, and cost. To achieve their ultra-scale properties these systems take a novel approach to data update management: by relaxing the synchronization requirements of updates that need to be disseminated to large numbers of replicas, these systems are able to survive under the harshest performance and availability conditions. These implementations are based on the concept of eventual consistency. The advances in data management developed by Amazon engineers have been the starting point for the architectures underneath the cloud storage and data management services offered by Amazon Web Services (AWS). For example, our Simple Storage Service, Elastic Block Store, and SimpleDB all derive their basic architecture from unique Amazon technologies. Other areas of Amazon’s business face similarly complex data processing and decision problems, such as product data ingestion and categorization, demand forecasting, inventory allocation, and fraud detection. Rulebased systems can be used successfully, but they can be hard to maintain and can become brittle over time. In many cases, advanced machine learning techniques provide more accurate classification and can self-heal to adapt to changing conditions. For example, our search engine employs data mining and machine learning algorithms that run in the background to build topic models, and we apply information extraction algorithms to identify attributes and extract entities from unstructured descriptions, allowing customers to narrow their searches and quickly find the desired product. We consider a large number of factors in search relevance to predict the probability of a customer’s interest and optimize the ranking of results. The diversity of products demands that we employ modern regression techniques like trained random forests of decision trees to flexibly incorporate thousands of product attributes at rank time. The end result of all this behind-the-scenes software? Fast, accurate search results that help you find what you want. All the effort we put into technology might not matter that much if we kept technology off to the side in some sort of R&D department, but we don’t take that approach. Technology infuses all of our teams, all of our processes, our decision-making, and our approach to innovation in each of our businesses. It is deeply integrated into everything we do. One example is Whispersync, our Kindle service designed to ensure that everywhere you go, no matter what devices you have with you, you can access your reading library and all of your highlights, notes, and bookmarks, all in sync across your Kindle devices and mobile apps. The technical challenge is making this a reality for millions of Kindle owners, with hundreds of millions of books, and hundreds of device types, living in over 100 countries around the world — at 24x7 reliability. At the heart of Whispersync is an eventually consistent replicated data store, with application defined conflict resolution that must and can deal with device isolation lasting weeks or longer. As a Kindle customer, of course, we hide all this technology from you. So when you open your Kindle, it’s in sync and on the right page. To paraphrase Arthur C. Clarke, like any sufficiently advanced technology, it’s indistinguishable from magic. Now, if the eyes of some shareowners dutifully reading this letter are by this point glazing over, I will awaken you by pointing out that, in my opinion, these techniques are not idly pursued — they lead directly to free cash flow. We live in an era of extraordinary increases in available bandwidth, disk space, and processing power, all of which continue to get cheap fast. We have on our team some of the most sophisticated technologists in the world — helping to solve challenges that are right on the edge of what’s possible today. As I’ve discussed many times before, we have unshakeable conviction that the long-term interests of shareowners are perfectly aligned with the interests of customers. And we like it that way. Invention is in our DNA and technology is the fundamental tool we wield to evolve and improve every aspect of the experience we provide our customers. We still have a lot to learn, and I expect and hope we’ll continue to have so much fun learning it. I take great pride in being part of this team. As always, I attach a copy of our original 1997 letter. Our approach remains the same, and it’s still Day 1. Jeffrey P. Bezos Founder and Chief Executive Officer Amazon.com, Inc.

2010 Amazon Letter to Shareholders

0

2010-amazon-shareholders-letter-1e341bd50a70

2018-07-15

2018-07-15 04:06:45

https://medium.com/s/story/2010-amazon-shareholders-letter-1e341bd50a70

false

1,157

null

Machine Learning

machine-learning

Machine Learning

51,320

Janus Bonaparte

null

187a1ead2b3e

david.m.kagan

0

1

20,181,104

null

0

null

0

null

2018-04-09

2018-04-09 17:32:42

2018-04-09

2018-04-09 17:35:57

1

false

en

2018-04-09

2018-04-09 17:35:57

4

1e34e074383d

3.173585

1

0

Artificial intelligence is already making our devices more personal — from simplifying daily tasks to increasing productivity. Emotion AI…

5

Emotion AI: Why your refrigerator could soon understand your moods Artificial intelligence is already making our devices more personal — from simplifying daily tasks to increasing productivity. Emotion AI (also called affective computing) will take this to new heights by helping our devices understand our moods. That means we can expect smart refrigerators that interpret how we feel (based on what we say, how we slam the door) and then suggest foods to match those feelings. Our cars could even know when we’re angry, based on our driving habits. Humans use non-verbal cues, such as facial expressions, gestures, and tone of voice, to communicate a range of feelings. Emotion AI goes beyond natural language processing by using computer vision and voice analysis to detect those moods and emotions. Voice of the customer (VoC) programs will leverage emotion AI technology to perform granular and individual sentiment analysis at scale. The result: Our devices will be in tune with us. Conversational services Digital giants — including Google, Amazon, Apple, Facebook, Microsoft, Baidu, and Tencent — have been investing in AI techniques that enhance their platforms and ecosystems. We are still at “Level 1” when it comes to conversational services such as Apple’s Siri, Microsoft’s Cortana, and Google Assistant. However, the market is set to reach new levels in the next one to two years. Nearly 40 percent of smartphone users employ conversational systems on a daily basis, according to a 2017 Gartner survey of online adults in the United States. These services will not only become more intelligent and sophisticated in terms of processing verbal commands and questions, they will also grow to understand emotional states and contexts. Today, there are a handful of available smartphone apps and connected home devices that can capture a user’s emotions. Additional prototypes and commercial products exist — for example, Emoshape’s connected home hub, Beyond Verbal‘s voice recognition app, and the connected home VPA Hubble. Large technology vendors such as IBM, Google, and Microsoft are investing in this emerging area, as are ambitious startups. At this stage, one of the most significant shortcomings of such systems is a lack of contextual information. Adding emotional context by analyzing data points from facial expressions, voice intonation, and behavioral patterns will significantly enhance the user experience. Wearables and connected cars In the second wave of development for emotion AI, we will see value brought to many more areas, including educational software, video games, diagnostic software, athletic and health performance, and autonomous cars. Developments are underway in all of these fields, but 2018 will see many products realized and an increased number of new projects. Beyond smartphones and connected-home devices, wearables and the connected car will collect, analyze, and process users’ emotional data via computer vision, audio, or sensors. The captured behavioral data will allow these devices to adapt or respond to a user’s needs. Technology vendors, including Affectiva, Eyeris, and Audeering, are working with the automotive OEMs to develop new experiences inside the car that monitor users’ behavior in order to offer assistance, monitor safe-driving behavior, and enhance their ride. There is also an opportunity for more specialized devices, such as medical wristbands that can anticipate a seizure a few minutes before the actual event, facilitating early response. Special apps developed for diagnostics and therapy may be able to recognize conditions such as depression or help children with autism. Another important area is the development of anthropomorphic qualities in AI systems — such as personal assistant robots (PARs) that can adapt to different emotional contexts or individuals. A PAR will develop a “personality” as it has more interactions with a specific person, allowing it to better meet the user’s needs. Vendors such as IBM, as well as startups like Emoshape, are developing techniques to lend such anthropomorphic qualities to robotic systems. VoC will help brands understand their consumers Beyond enhancing robotics and personal devices, emotion AI can be applied in customer experience initiatives, such as VoC programs. A fleet of vendors already offer sentiment analysis by mining billions of data points on social media platforms and user forums. Some of these programs are limited to distinguishing between positive and negative sentiments while others are more advanced, capable of attributing nuanced emotional states — but so far, only in the aggregate. We are still at an early stages when it comes to enhancing VoC programs with emotion AI. Technology providers will have to take a consultative approach with their clients — most of whom will be new to the concept of emotion AI. While there are only a few isolated use cases for emotion AI at the moment, we can expect it to eventually offer tools that transform virtually every aspect of our daily lives.

Emotion AI: Why your refrigerator could soon understand your moods

1

emotion-ai-why-your-refrigerator-could-soon-understand-your-moods-1e34e074383d

2018-04-10

2018-04-10 11:55:58

https://medium.com/s/story/emotion-ai-why-your-refrigerator-could-soon-understand-your-moods-1e34e074383d

false

788

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

lakshya

open sources (tech geek)

fe2b4cd6d2d6

xtaraim

48

201

20,181,104

null

0

null

0

null

2018-02-26

2018-02-26 13:47:13

2018-02-26

2018-02-26 13:51:12

0

false

en

2018-03-25

2018-03-25 13:21:09

63

1e363a3fd9f9

2.369811

0

February 26, 2018

5

Week 8: Reading list & news to keep up-to-date with Javascript, Python, DevOps, Cloud, Serverless, Blockchain, AI, Big Data, Personal Growth, Startup. February 26, 2018 Javascript & Python https://medium.com/webpack/webpack-4-released-today-6cdb994702d4 https://medium.com/webpack/webpack-4-code-splitting-chunk-graph-and-the-splitchunks-optimization-be739a861366 https://medium.com/webpack/webpack-4-mode-and-optimization-5423a6bc597a https://hackernoon.com/o-api-an-alternative-to-rest-apis-e9a2ed53b93c https://medium.freecodecamp.org/how-to-create-and-publish-a-chrome-extension-in-20-minutes-6dc8395d7153 https://medium.freecodecamp.org/elegant-patterns-in-modern-javascript-roro-be01e7669cbd https://blog.jupyter.org/jupyterlab-is-ready-for-users-5a6f039b8906 https://medium.freecodecamp.org/want-to-learn-css-variables-heres-my-free-8-part-course-f2ff452e5140 DevOps https://medium.com/@hiflan/continuous-deployment-with-aws-codepipeline-1882128571ed https://blog.devopspro.co.uk/how-i-helped-my-company-ship-features-10-times-faster-and-made-dev-and-ops-win-a758a83b530c https://hackernoon.com/the-best-architecture-with-docker-and-kubernetes-myth-or-reality-77b4f8f3804d Cloud & Serverless https://koukia.ca/a-microservices-implementation-journey-part-4-9c19a16385e9 https://hackernoon.com/building-a-highly-scalable-imgur-clone-with-lambda-and-s3-aaf9da422c3e https://medium.freecodecamp.org/express-js-and-aws-lambda-a-serverless-love-story-7c77ba0eaa35 https://read.acloud.guru/building-an-imgur-clone-part-2-image-rekognition-and-a-dynamodb-backend-abc9af300123 https://hackernoon.com/running-a-scalable-reliable-graphql-endpoint-with-serverless-24c3bb5acb43 https://read.acloud.guru/continuous-deployment-with-serverless-and-circleci-772f990820ee AI https://www.cnbc.com/2018/02/21/elon-musk-is-leaving-the-board-of-openai.html https://eng.uber.com/uber-ai-residency/ http://deeplearning.cs.cmu.edu/ http://konukoii.com/blog/2018/02/19/twitter-sentiment-analysis-using-combined-lstm-cnn-models/ http://www.ipam.ucla.edu/programs/workshops/new-deep-learning-techniques/?tab=schedule https://blog.openai.com/preparing-for-malicious-uses-of-ai/ https://research.googleblog.com/2018/02/a-summary-of-google-zurich-algorithms.html https://research.googleblog.com/2018/02/assessing-cardiovascular-risk-factors.html https://www.quora.com/How-did-SpaceX-land-the-Falcon-Heavy-boosters-so-accurately http://money.cnn.com/2018/02/22/technology/self-driving-car-wash/index.html https://towardsdatascience.com/improving-vanilla-gradient-descent-f9d91031ab1d Big Data https://towardsdatascience.com/a-beginners-guide-to-data-engineering-part-ii-47c4e7cbda71 https://medium.com/@rchang/a-beginners-guide-to-data-engineering-part-i-4227c5c457d7 https://tech.flipkart.com/overview-of-flipkart-data-platform-20c6d3e9a196 https://towardsdatascience.com/the-best-data-science-learning-resources-out-there-and-my-journey-into-data-science-81c5a6ec67bc Blockchain https://medium.com/@cryptonomii/endor-ico-analysis-the-google-of-predictive-analytics-586040b0649a https://medium.com/datastreamx/introducing-the-dsx-network-203c5d3e9b2a https://medium.com/helloiconworld/launching-of-iconex-b88c998442cf https://hackernoon.com/3-cryptocurrencies-to-earn-you-money-while-you-sleep-part-1-3ef340fa4c70 https://medium.com/@apompliano/the-official-guide-to-tokenized-securities-44e8342bb24f https://medium.com/theblock1/popping-the-bubble-blockchain-and-cryptocurrency-7130156f91b2 https://hackernoon.com/cryptocurrency-platforms-the-key-to-disruptive-innovation-ddb5bcf8c5c7 https://hackernoon.com/who-is-the-next-microsoft-in-blockchain-d81771539ccc https://medium.com/icodrops/the-power-of-nuclear-fusion-for-the-era-of-cryptofinance-50a5d14cc8ad https://medium.com/@obirum/telegram-name-logo-obirum-ico-airdrop-bounty-d635ec9ab76d https://medium.com/@wandererli/crypto-trading-tips-how-to-stop-losing-money-like-a-noob-d07c0ee3f21 https://hackernoon.com/stablecoins-designing-a-price-stable-cryptocurrency-6bf24e2689e5 Personal Growth https://artplusmarketing.com/5-refreshing-things-you-should-do-every-sunday-morning-no-excuses-5f2cda4b2d2c https://medium.com/personal-growth/focus-on-learning-and-creating-rather-than-entertainment-and-distraction-4cbc469ed4ab https://medium.com/swlh/deep-thinking-in-the-age-of-distraction-f7cf765b2762 https://medium.com/personal-growth/youll-never-make-up-for-your-past-c59421d944ab https://medium.com/personal-growth/10-things-you-can-do-in-your-daily-life-to-improve-your-personal-development-55e6ffa5c1c2 https://medium.com/personal-growth/you-dont-have-to-live-the-life-everyone-else-is-living-7faa2c26c507 https://medium.com/swlh/12-crucial-things-that-everyone-in-their-20s-need-to-do-in-order-to-be-successful-348082c4eb3c https://medium.com/personal-growth/the-mindset-shift-all-great-leaders-make-to-turn-conflict-into-cooperation-ff2dfcaf6129 https://medium.com/personal-growth/9-things-people-who-fail-do-on-a-regular-basis-dont-do-these-a50059820ac8 https://medium.com/personal-growth/7-brutal-life-lessons-everyone-has-to-learn-multiple-times-6e0eb0c7c0f4 https://medium.com/the-mission/consistency-beats-talent-luck-goodintentions-and-even-quality-46b094c34240 https://medium.com/swlh/how-to-be-productive-10-ways-to-actually-work-smarter-4eece37a5c8e https://medium.com/@benjaminhardy/how-to-know-if-youve-leveled-up-as-a-person-3ff0611e16fc Startups https://medium.com/@posttweetism/lets-have-no-managers-instead-of-managers-with-no-engineering-experience-e8b7cd29d398 https://medium.com/swlh/why-customer-service-is-the-new-marketing-a45fcbd4c962 https://medium.com/the-mission/when-to-quit-and-when-to-double-down-4b38f4cce144 Misc https://hackernoon.com/life-after-google-a-comparison-to-startup-life-831e74d5c6f8 https://medium.freecodecamp.org/what-i-learned-from-an-old-github-project-that-won-3-000-stars-in-a-week-628349a5ee14 https://medium.freecodecamp.org/how-i-applied-lessons-learned-from-a-failed-technical-interview-to-get-5-job-offers-656fcf58034d

Week 8: Reading list & news to keep up-to-date with Javascript, Python, DevOps, Cloud, Serverless…

0

week-8-reading-list-news-to-keep-up-to-date-with-javascript-python-devops-cloud-serverless-1e363a3fd9f9

2018-03-25

2018-03-25 13:21:10

https://medium.com/s/story/week-8-reading-list-news-to-keep-up-to-date-with-javascript-python-devops-cloud-serverless-1e363a3fd9f9

false

628

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

trannotes

keep links already read so I may revisit http://www.paulgraham.com/know.html

ce02e684c6eb

trannotes

98

39

20,181,104

null

0

null

0

null

2018-09-13

2018-09-13 18:17:11

2018-09-13

2018-09-13 18:20:21

6

false

en

2018-09-13

2018-09-13 18:20:21

22

1e365fe0ba41

8.576415

1

0

Mycroft is a global company. One of the key advantages of having a distributed world-wide workforce is our ability to engage with…

4

Digital AI Summit Melbourne — Session Notes Mycroft is a global company. One of the key advantages of having a distributed world-wide workforce is our ability to engage with entrepreneurial and technology ecosystems in cities across the world. As our Australian contingent, I recently attended the Digital AI Summit in Melbourne. In our spirit of openness and sharing, I have provided my notes for our entire community to benefit from. Tweets from the day are available at the hashtag #digitalaisummit Session 1 — Blair Bryant, Global Digital Advisor, Microsoft Blair outlined how important it was to get the CEO on board with digital transformation programs. He underscored how digital transformation needs to solve specific problems for the C-suite — don’t sell digital transformation itself, sell the problems that it solves. He outlined a four-point plan; Inspire the CEO — the CEO needs to be inspired to adopt digital transformation Align the senior leadership team Build a digital transformation strategy Understand how the execution will proceed Business should immediately start to gather their data; understanding what data they collect — as data is the “fuel” for machine learning. Then, business should start to understand how hyper-personalization will influence their offerings, services, and products. Key takeaway: Get your house in order, build a plan and execute it. Blair Bryant Session 2 — Professor Phil Cohen, Monash University Laboratory for Dialogue Research — Conversational Technology, Present, and Future Prof Cohen has recently joined Monash University to head up the Laboratory for Dialogue Research. He previously headed up a startup called VoiceBox, which was recently acquired by Nuance. This was the standout presentation of the day. Prof Cohen provided an overview of the evolution of voice technologies. Although much recent progress has been made, largely in part due to advances in machine learning and computational speed, dialogue and interaction are still stunted by several challenges, including; Paraphrasing — there are many ways of saying the same thing Ambiguity — one phrase can have multiple meanings depending on context Meaning and semantics — meaning can be different to spoken words — “Do you have the time” means “tell me the time” but this is not the way it is worded Pragmatics — this is the word Cohen used to describe “context” — the history of a dialogue and the meaning that is imbued in that history. Chatbots Cohen outlined the basic premise behind Chatbots — they’re essentially “stimulus-response” engines. Where an Intent is matched to a stimulus and a response is provided. However, they only provide the illusion of having a dialogue — they currently don’t handle context well, nor do they handle follow up questions or interactions that fall outside the ‘dialogue tree’. He provided an excellent walkthrough of failures of Chatbots — how they don’t handle diplomatic nuances well — ie “What’s your grandmother’s name” => “My grandmother is dead” — nor do they handle follow up questions well because their ability to determine context is severely limited. Spoken natural language retrieval Natural language (NL) retrieval allows a system to respond to a query using an FAQ type system, or route a query to the right department using keyword matching. Again, the ability to hold a dialogue is very limited. Voice assistants Voice assistants are able to handle broad topics, but there are limited ways to phrase a query — it’s not really “natural language”. Semantic parsing Semantic parsing is a technology currently in research labs that allows users to have better “expressivity” and better “understanding” of natural language. Semantic parsing has more capability than the current generation of voice assistants, and can handle things like superlatives — ie “find the best pizza in downtown closest to the space needle, but not McDonald’s”. https://en.wikipedia.org/wiki/Semantic_parsing Side note: I had a quick look at open source Semantic Parsing libraries and this one called SLING from Google looked like the most commonly used one — if anyone has thoughts, feel free to post in the Forum! https://github.com/google/sling Prof Cohen gave a live demo of Voicebox using web-based speech to text, which was then converted in real time to Intents, and it came back with a pretty slick answer. https://www.voicebox.com/ — the live demo website isn’t available. Comparison voice assistants He then presented a chart (not available online) of what Voicebox can do compared to Siri, Google Home, and Alexa, positioning Voicebox in a positive light. Transactional dialogue The next big frontier for voice assistants, as we’re starting to see with projects like Duplex from Google is transactional dialogue — being able to have a natural sounding conversation with a voice assistant that allows the user to complete a task — booking tickets for a movie, ordering a pizza and so on. Most transactional dialogues are ‘slot-filling’ systems where the aim of the dialogue is to ensure the slots are filled so that an API can be used to complete the transaction. That is, the voice assistant will prompt for the missing ‘slot values’. There are not many systems like this on the market today. Semantic parsing and collaboration The next horizon in speech recognition research is semantic parsing coupled with collaboration. Under this model, the voice assistant will be able to: Analyse meaning Infer the intent of the user from the utterance — that is what is the user really trying to do? Debug the plan — be able to explain how it arrived at a conclusion about meaning Offer the user solutions to the Intent that has been inferred The Laboratory for Dialogue Research intends to collaborate with industry via a membership model. Key takeaways: Investigate semantic parsing libraries as part of emerging tech roadmap Professor Phil Cohen Session 3 — Whole of government and whole of community approach to AI — Cheryl George, Kathy Coultas, and Martine Letts Kathy Coultas — Director, Technology Innovation and Investment, Department of Economic Development, Victoria Cheryl George — Government and Stakeholder Relations, Data61 CSIRO Clive Dwyer and Martine Letts — Committee for Melbourne The AI space is evolving rapidly, and there are opportunities that Australia is well positioned to take advantage of — especially using our domain knowledge. To really take advantage of the opportunities requires a broad cross-sector collaborative approach. Coordination will be critical. The government needs an “enabling environment” and “enabling infrastructure” to harness the opportunities, and this needs to be done quickly. Kathy Coultas recognized the need for ALL citizens to engage with digital, and explained that this was part of the reason behind the Digital Innovation Festival. Martine Letts noted that the cross-partisan Parliamentary AI group had been established at a state level to help politicians really grapple with the opportunities and challenges of AI. She would like to see this initiative expanded into a national working party. The only other country in the world that has a similar cross-partisan working party is the United Kingdom. Kathy Coultas noted the issues surrounding data privacy, security, and ethics and how we need to tackle these as part of coming to terms with artificial intelligence. Part of this will be the necessity for legislative amendments to harness emerging technology effectively while protecting citizens from foreseeable harm. She was firm that political point-scoring won’t be effective; regulatory reform for emerging technologies requires a multi-partisan approach. Martine Letts noted that Australia is well behind investment in the AI space; China and the USA are really leading both investment and technology development in this space. AI is seen not as an integral part of emerging technology strategy — and a pillar of strategy — but as a “bolt on” to existing measures. This view is anachronistic and will not service organizations well. Kathy Coultas followed up by outlining that most technology investment now is poured into ABC — AI, blockchain and crypto. Australia is significantly behind in terms of technology investment, and that needs to seriously change in order to realize the vision of Melbourne becoming the technology capital of Australia. We have the capability, but it’s fragmented across multiple sectors. Kathy Coultas also noted the investment needed in workforce capability and training to be able to have the skills needed to harness AI — right now there is a significant skills shortable in this space, and universities are only now moving to catch up and address this shortfall. Cheryl George highlighted that CSIRO/Data 61 had been tasked with developing a national AI technologies roadmap to lay out the opportunities, threats and possible approaches. A draft of this is due by the end of the year 2018. Key takeaways: Whole of government approaches to AI require significant coordination, whole of sector and whole of pipeline approaches. Dr Cheryl George, Kathy Coultas and Martine Letts Session 4 — Dr Nathan Faggian, Google — Machine Learning Infrastructure Nathan attempted to provide an overview of machine learning in 30 minutes — much respect to him. This was the second-best presentation of the day. He opened by explaining that what used to be fantasy is now reality. There have been massive improvements in artificial intelligence and machine learning. Google uses a huge amount of machine learning and AI — and he quoted the statistic that over 20% of Gmail responses are now the “automated” pre-canned responses that are available. He provided a demonstration of Google Duplex, and explained how it was using semantic parsing based on massive machine learning efforts to be able to have natural interaction style. In terms of Professor Cohen’s earlier presentation, Duplex is certainly at the “cutting edge” of where speech recognition technology is at the moment. He went on to show how machine learning can be valuable in industrial contexts — citing both the case of where illegal fishing was identified in Vanuatu due to fishing boat movements, and the case of soy crop forecasting — which was accurate to within 2% with a 5 month lead time — an incredible level of accuracy. He underscored the importance of data and the huge volumes of data that are required for effective machine learning. He also outlined that machine learning does not exist in isolation in an organization, and that there is a large supporting infrastructure that sits around it. This was essentially a shadow-pitch for the Google Cloud Platform, but did make the important point that there are many parts to an effective machine learning capability within an organization. They all have to be considered as a whole in order to be able to effectively scale an organization’s machine learning efforts. Side note: None of the existing governance frameworks like COBIT, ITIL or SFIA have recognized machine learning as a key organizational capability yet. Key takeaways: Google uses a lot of AI, they’re the leaders in the field. Make sure you have the infrastructure to scale your ML efforts. It’s not just about ML and algorithms, the infrastructure is a key component of your machine learning capability. ML can solve some pretty hoary problems for business and industry. If you’re not considering it yet you should be. Dr. Nathan Faggian Session 5 — Health Technology Panel Annette Hicks, IBM Alice Sidhu Dr. Priscilla Rogers Mike Wang, NVIDIA This was a really interesting panel that looked at the application of machine learning to the health sector. The key points were; Getting access to patient records in a way that is consistent is very difficult because data formats differ between providers; data standards matter. Machine learning has a large role to play in effective pharmacology dosing. For instance, the industrial strength antibiotic vancomycin is nephrotoxic, so being able to get the right dose by personalizing it for the patient based on their unique characteristics means less kidney damage. IBM have faced the challenge of trying to convince clinicians that tools like Watson are there as complementary tools. They’re not trying to replace the clinician — the goal is to work in a complementary way. Key takeaways: Data, data, data. Getting the right data in a machine-readable format is absolutely essential for machine learning. Annette Hicks and Dr. Priscilla Rogers Orginally Posted on Mycroft Blog here!

Digital AI Summit Melbourne — Session Notes

50

digital-ai-summit-melbourne-session-notes-1e365fe0ba41

2018-09-13

2018-09-13 18:20:21

https://medium.com/s/story/digital-ai-summit-melbourne-session-notes-1e365fe0ba41

false

2,021

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Mycroft AI

The Open Source Privacy Focused Voice Company

384b053af525

mycroftai

7

2

20,181,104

null

0

null

0

null

2018-08-15

2018-08-15 06:44:14

2018-08-15

2018-08-15 06:54:29

0

false

en

2018-08-15

2018-08-15 06:54:29

4

1e376708348b

1.362264

0

While joining Forsk bootcamps, participants has lot of queries. We have compiled all those queries and listing them in public domain.

5

Forsk Bootcamps: Frequently Asked Questions While joining Forsk bootcamps, participants has lot of queries. We have compiled all those queries and listing them in public domain. 1. What is Forsk Technologies? ANS: Forsk Technologies is based out of Jaipur and works with universities to prepare students for skill based hiring using data and technologies. Currently focused on IoT, Machine Learning, Deep Learning, and AI. Forsk’s ESTP (Engineering Specialist Training Program) helps students fill industry gap by improving skills in emerging technologies, better connect with industry and secure quality career in the industry. Forsk’s Python course is part of IIIT Kota curriculum. Currently Forsk is developing a Paytm like wallet for Bombay based company that includes web development, mobile app development, cloud and integrationg ML/AI features. We are working with Manipal for last three years and close to 300 students have graduated from our previous bootcamps. 2. What tracks you are covering this semester and what is the format of execution? ANS: We are offering three tracks this time: a. Deep Learning and AI b. Machine Learning with Python c. IoT The bootcamps are 70% practical (project based learning) and 30% includes demo and discussion. We follow no PPT slides strategy. 3. Do these bootcamps help me in getting internships and jobs? ANS: Yes, many students from past have got internship offers through Forsk connections in industry. 4. What is the profile of mentors from Forsk? ANS: The mentors are from industry having more than 12+ years. Forsk founders are Ex-Qualcomm. To know more about founder, please visit their LinkedIn profile. Yogendra: www.linkedin.com/in/yogendrasinsinwar Sylvester: https://www.linkedin.com/in/dr-sylvester-fernandes-b4a43a6/ 5. Can I see reviews from past participants? ANS: You can find the reviews about Forsk on Quora. https://www.quora.com/What-is-your-review-of-Forsk-Technologies-Jaipur 6. I have no coding experience or have fear with coding? ANS: As long as you know basics of C programming and interested to learn new things, no experience with coding will stop you learning. 7. How can I register for the bootcamps? ANS: You can register by filling a Google form: http://bit.ly/2NjDrnb We will keep updating this list as we come across new queries from participants.

Forsk Bootcamps: Frequently Asked Questions

0

forsk-bootcamps-frequently-asked-questions-1e376708348b

2018-08-15

2018-08-15 06:54:30

https://medium.com/s/story/forsk-bootcamps-frequently-asked-questions-1e376708348b

false

361

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Forsk Technologies

Forsk Technologies, Jaipur prepares students for skill based hiring using data and technologies in areas of C, Java, Python, Machine Learning, DL, AI and IoT.

646364405e1a

yogendrasinsinwar

1

2

20,181,104

null

0

null

0

null

2018-06-11

2018-06-11 16:16:44

2018-06-11

2018-06-11 22:25:43

2

false

en

2018-06-11

2018-06-11 22:25:43

1

1e37d48f5f01

2.262579

3

0

Navigating the digital news cycle is like a high school prom: socially uncomfortable, economically dated, and no one’s having any fun. The…

5

CrossCheck News Navigating the digital news cycle is like a high school prom: socially uncomfortable, economically dated, and no one’s having any fun. The internet has set fire to our little sweaty soiree, with the punch bowl of common decency being the first casualty. Next to perish was objectivity, and then attention-span. While we don’t have a flux capacitor to go back and change things, we do have something better: Machine Learning! With CrossCheck’s help, we can all be the Marty McFly of the dance, change our future, and have fun while doing it. Usable on leading social platforms. To stop our collective tilting at windmills, or Twitter tweets, CrossCheck offers a real-time approach to content validation through our algorithm with mission of deincentivising users from engaging with unverifiable digital content. Manually verifying this info has become demanding, for individuals and organizations alike. As evidenced by the recent presidential election cycle, there is a genuine danger in the capacity of unverified information being passed as legitimate news to mislead the public. Current fact-checking mechanisms employed by large companies such as Facebook and Google are slow, inefficient, and expensive. At CrossCheck, our servers work in real-time to gather articles from across the web and identify which ones are verifiable accurate, all the while offering users the ability to verify and contextualize their news diet for the better. Plus we’re free to use. Other startups attempting to tackle fake news have also emerged but have yet to launch to market. These HAL 9000s are not targeting news consumers, but are looking at an industry level product, relying on crude domain-level blacklists. Here at CrossCheck we prefer to be more grounded and personal, like R2D2. We still use advanced technology to address fraudulent content, A big difference for CrossCheck is that we work article by article, post by post, rather than whole labeling websites as fake or not. You can use us on any text-based content whether its a tweet or a New York Times article. Using machine learning our algorithm collects millions of articles from across the web in real time, extracts their content, and sorts all of them into clusters. Then these clusters are sorted into good or bad piles. When users highlight a Facebook post or an article our product matches said post/article to these clusters to produce a CrossCheck score. This score represents the percentage of how much the highlighted content matches the “good” clusters. Kevin Kelly, co-founder of Wired Magazine addressed the issue in a BBC-Future interview saying that, “Truth is no longer dictated by authorities, but is networked by peers. For every fact, there is a contract. All those contracts and facts look identical online, which is confusing to most people (Gray).” CrossCheck provides the capability of differentiating amongst the facts and contracts, so users can engage with online news content with the knowledge that what they’re reading is validated. CrossCheck currently offers a browser extension for Google Chrome, downloadable for free.

CrossCheck News

52

crosscheck-news-1e37d48f5f01

2018-06-14

2018-06-14 01:32:19

https://medium.com/s/story/crosscheck-news-1e37d48f5f01

false

498

null

Journalism

journalism

Journalism

39,588

Franklin Young

null

870e3f093ee6

franklin_48594

0

1

20,181,104

null

436

null

0

null

2018-03-20

2018-03-20 16:55:23

2018-03-20

2018-03-20 20:10:59

1

true

en

2018-03-20

2018-03-20 20:12:48

5

1e37ef0649da

4.818868

599

32

1

Your Facebook data was being used by countless organizations for years, including Obama’s 2012 campaign

5

The Media’s Double Standard on Privacy and Cambridge Analytica Your Facebook data was being used by countless organizations for years, including Obama’s 2012 campaign Getty Listening to most of the analysis of Cambridge Analytica’s use of Facebook data, one would think that our deepest, darkest secrets were pilfered from Facebook’s servers and hand-delivered to Trump Tower and the Kremlin, which skillfully used them to exploit our fears and manipulate our emotions. One could be forgiven for thinking this when Cambridge waged a brilliant marketing campaign to convince you that their mastery of the dark arts — honed through waging information war in the third world — was the secret ingredient behind Donald Trump’s shocking victory. Further revelations that Cambridge was not completely on the straight and narrow in how it handled data has added fuel to the conflagration over social media, fake news, and the Russian influence campaign. This firestorm has put political data collection practices in the crosshairs of Robert Mueller’s investigation, a British parliamentary inquiry, and government regulators in the US and the European Union. As with many stories associated with the trigger words “Trump,” “Russia,” and “fake news”, the story is so bound up in the political passions of the moment that it’s difficult to discern what’s real and what isn’t. So what exactly was Cambridge Analytica trying to do? By collecting data through an academic front, Cambridge Analytica violated Facebook rules. This has been described in terms of a “data breach” — but it wasn’t in the traditional sense. Countless developers accessed the same data under the rules — which were changed in 2014 and sunset in 2015 — including the Obama 2012 campaign. The data itself was vast in scale, but its contents were hardly earth shattering, consisting of a list of names, hometowns, and the pages a user had liked, which most people on Facebook are public about. The origins of this controversy can be traced back to Facebook’s launch of an application programming interface (or API) for app developers in 2008. With this platform, Facebook aimed to “lock in” users by organizing its users social lives not just on Facebook, but across a range of popular apps and services. To do this, Facebook granted developers liberal access to its user data so they could build social features in their own apps. Rather than creating an account and then requiring users to painstakingly re-friend everyone on Spotify, or Pinterest, or Instagram, these services sped adoption by letting you quickly import your existing social graph from Facebook. If your friends weren’t yet on these sites, Facebook gave developers your friend list so you could invite them. Along with your friend list, the API would return other information about your friends — specifically, the other pages they liked. This was done to help you find things you liked and friends with similar interests. In one API call, you could get an astonishing amount of data if you wanted it — your friend list and all the pages they liked, which amounted to potentially tens of thousands of data points per app authorization. Popular apps were able to recreate Facebook’s “social graph” — showing who was friends with whom — as well as the so-called “interest graph”, showing who liked what, and how these likes related to one another. This might let you see how one’s choice of liquor correlated with their choice of political candidate, for instance. Analytics companies which had spent millions of dollars to know exactly this — through massive surveys and purchasing consumer files from data brokers like Acxiom — now had free access to a comparable source of data. And for those who knew how to use it, Facebook data was potentially superior to the offline data. Pages that you had publicly liked were a high quality signal of evolving preferences and behaviors, while data about offline purchases contained in consumer databases were so sanitized as to be virtually useless. It was this opportunity — tied to its ideas about psychographic targeting — that Cambridge was looking to capitalize on. It wasn’t just technology companies getting in on the Facebook data game, but political campaigns. When Barack Obama announced for re-election in April of 2011, he did so with an unobtrusive Facebook app that asked simply “Are you in?” Authorizing the app allowed you to register yourself with the campaign with one click of a button. In doing so, the campaign also received your friend list. It wasn’t clear what the campaign would do with that list until late in the campaign season when it unveiled the latest iteration of their Facebook app, known as targeted sharing. The app matched your Facebook friend list with voter files in battleground states. One day, out of the blue, Obama supporters were receiving emails with the names and faces of their Facebook friends — asking them to tell their friends to vote. More than five million people were contacted through the app, but the Obama campaign likely had a list of Facebook users numbering in the tens, if not hundreds, of millions. The Facebook app was the subject of one of the few embargoed post-election exclusives the Obama campaign gave detailing its technological and analytical achievements, with one official calling it “the most significant piece of technology developed for this campaign.” In the fawning media coverage of the Obama campaign’s technological prowess, it did not occur to observers at the time to call this a startling invasion of privacy. And it wasn’t, or at a very minimum, the privacy risks were arguably outweighed by the benefits. A tool like this could be the future of politics: door-to-door canvassing for the digital age, and a welcome antidote to impersonal broadcast TV ads or a welcome upgrade from getting a phone call from a stranger telling you to vote. The conversation we had about analytics data following the 2012 campaign — recognizing the privacy tradeoffs but also the potential advances that might come from a more surgical and personalized approach to campaigning — is a far cry from the hysteria that reigns today. Today’s conversation reflects an anxiety that populist forces, specifically Donald Trump, have grown better at harnessing technology and social media. That exposes the purveyors of this technology, Facebook chief among them, to scrutiny and regulatory risk that didn’t exist when when the tools were in the hands of people in line with the sensibilities of the media and political establishment. The idea that technology companies might let a candidate like Obama but not Trump get away with borderline privacy violations isn’t hyperbole. It was essentially confirmed by the Obama campaign’s product manager on the targeted sharing app. Aspects of Cambridge’s use of the Facebook data — not to mention the growing revelations about the rest of its business — are troubling. It’s unclear exactly how the data was used, but we know two things: the Trump campaign was not among its users, and the end product Cambridge was using the dataset to build, personality-based targeting, has been universally and spectacularly panned by a range of ex-Cambridge clients. This could mean that while Facebook’s data might be able to tell us what car you’ll buy or which candidate you’ll vote for, it still can’t divine your personality or tell your secrets.

The Media’s Double Standard on Privacy and Cambridge Analytica

4,345

the-medias-double-standard-on-privacy-and-cambridge-analytica-1e37ef0649da

2018-08-25

2018-08-25 02:03:36

https://medium.com/s/story/the-medias-double-standard-on-privacy-and-cambridge-analytica-1e37ef0649da

false

1,224

null

Media

media

Media

51,760

Patrick Ruffini

Polling/analytics. Digital ex. Co-Founder @EchelonInsights.

5484304e4a61

PatrickRuffini

7,824

1,233

20,181,104

null

0

null

0

null

2018-06-28

2018-06-28 15:33:01

2018-06-28

2018-06-28 16:17:42

1

false

en

2018-06-28

2018-06-28 16:21:29

0

1e3a14f8b0bb

3.928302

1

0

Invisible data is the primary threat many organizations are currently struggling to address. Though they may not be fully conscious of it…

4

Invisible Data Invisible data is the primary threat many organizations are currently struggling to address. Though they may not be fully conscious of it, or even willing to admit it if they are aware of the problem, the pain and consequences are still very real. Come along with me on this journey and I’ll explain. What is “invisible data” you ask? Let’s begin… Within any commercial organization, there are MANY sources of data. There are systems to manage customers, systems to manage products, systems to manage online and offline sales, systems to manage shipping, individual contributor reports, 3rd party demographic data, and many, many, many, many others. These systems and sources are separated by boundaries both technical and organizational. Looked at in isolation, you’ll often hear these storehouses of data individually referred to as a “data silo”. Data silos aren’t necessarily a bad thing. Each system itself is purpose built to do one thing and do it well. A silo tends to do its job very effectively for its domain of expertise and its specific department of the business. The issue is that those who run the business need to ask questions of data that is stored across all these various sources and systems. Think of an organization as similar to a beach ball. A data silo, which may contain multiple data sources, represents a single stripe of the business. In isolation, you can query any single silo and easily get an idea of the condition of any individual stripe. But the business needs to understand the condition of the entire beach ball. To do this, you have to be able to connect the silos to see what the stripes look like together. Analysis is required to understand if the stripes continue to fit properly together to create a healthy and functional beach ball. A stripe may be faded. Some stripes may be larger or smaller than others. One stripe’s edges may be frayed. The ball may be lopsided or deflated or leaking air somewhere. The ball could be completely jacked up. You get the picture. An unhealthy beach ball means an unhealthy business. Since data is not stored and managed in a single, central location to be able to easily ask the questions the business wants to ask, an extraordinary amount of time and effort is spent to bring disparate sources together into some form of unified view. Data marts have been implemented to try and create a unified source of data from the silos, but the source systems and data being introduced to organizations are rapidly evolving and coming in faster than data marts and IT can adapt to. So, to provide the answers so desperately sought, businesses turn to their favorite data integration tool of all time: Excel! A BI tool such as Tableau or MicroStrategy or Qlikview or PowerBI or whatever may be used for this task as well. Many of these function like Excel on steroids. Access databases are popular too. People tend to turn to something they have easy access to so they can get their job done. Using applications of this type, analysis commences by performing “Swivel Chair Integrations”. A person queries one system, cuts and pastes the results in Excel, swivels their chair, queries the next system, and repeats, until all the data they think they need is stored in a single workbook. A variation of this Excel integration art is known as the “Stare and Compare”. This is where someone stares at one source of data, stares at another, compares the two, and then captures some notes or details from the comparison into an “integrated” spreadsheet. Having consolidated the data, an individual will next go about the laborious task of munging and manipulating the data to normalize and rationalize the overlaps and inconsistencies from across systems. For example, a transaction date may be labeled TXDATE in one system and PURCHASEDATE in another. The systems may store the date values in different date formats as well. Munging will be done within the spreadsheet, and mappings and rationalizations will be completed in a fashion totally dependent on the decisions of the person performing the data wrangling to normalize the data. ( Bob from Accounting decides this field should be called TDATE and persisted as an Excel date type. Good one Bob! ) No joke. This occurs regularly. An individual may perform this type of data integration, but I’ve also witnessed orgs that literally have floors of people collaborating on a single spreadsheet to deliver the analysis a business requires. Of course, there is a multitude and myriad of concerns with this approach. People are working with older, stale snapshots of data, mapping of columns and values is dependent on individuals and not documented anywhere, spreadsheets aren’t versioned and are passed around via email. All this manual effort is VERY error prone, and it can take days and even WEEKS to complete a single analysis. To add insult to injury, by the time analysis is “complete”, it’s irrelevant to where a business is today and more relevant to a time that existed at some point in the past. And what’s even more crazy, is at the end of all this effort, once a quarter, somehow, these numbers that “answer” very important questions on the state of the business, actually make it into the boardroom. (True story. And in a PowerPoint of course.) Now, if any executive in the meeting were to request: “Please rerun this report so we have the latest numbers.” IT CAN’T BE DONE. Why? The data that was used to create the report is floating around in copies of Excel, in Access DBs, in emails and in people’s heads. Most of the data and “rules” applied to unify it are gone! None of this is persisted or documented anywhere. There is no central, unfiied source to refer back to… It’s INVISIBLE DATA! And the cycle continues. More to come…

Invisible Data

7

invisible-data-1e3a14f8b0bb

2018-06-28

2018-06-28 16:21:29

https://medium.com/s/story/invisible-data-1e3a14f8b0bb

false

988

null

Fintech

fintech

Fintech

38,568

Pete Aven

Connecting people, information, and systems. Fights for the users.

eeb0710c8f54

wpaven

16

8

20,181,104

null

0

def hybrid_descent(foo, initial_point, iterations = 5, damping = 0.25): point = initial_point dim = len(initial_point) # gradient of foo dfoo = nd.Gradient(foo) iterL = [] for iter in range(iterations): iterD = {} function_value = foo(point) # gradient of foo at the current iteration point gradient = tf.reshape(dfoo(point), shape = [dim,1]) # practical gradient change in the variables, note the default damping = 0.25 delta_grad = - gradient * damping # Hessian (matrix of mixed partial derivatives) of foo hessian = tf.constant(nd.Hessian(f)(point)) # inverse of the Hessian inv_hessian = tf.matrix_inverse(hessian, adjoint=False) # optimal vector change in the variables, note the '-' optimal_delta = -tf.matmul(inv_hessian, gradient) # practical Hessian vector change in the variables, note the Hessian damping = 2 * damping for gradient delta_hess = optimal_delta * damping * 2 # Hessian "norm" of (negative) gradient = optimal_delta * ( - gradient), ?> 0 # = Hessian **(-1) * gradient * gradient hess_grad_sq = - tf.matmul(tf.reshape(optimal_delta, shape = [1, dim]), gradient) # reshape gradient = tf.reshape(gradient, shape = [dim,]) delta_grad = tf.reshape(delta_grad, shape = [dim,]) delta_hess = tf.reshape(delta_hess, shape = [dim,]) hess_grad_sq = tf.reshape(hess_grad_sq, shape = [1,])[0] with tf.Session() as sess: # sess.run(tf.global_variables_initializer()) gradient = sess.run(gradient) hess_grad_sq = sess.run(hess_grad_sq) if hess_grad_sq > 0: delta = sess.run(delta_hess) method = 'Hessian' else: delta = sess.run(delta_grad) method = 'Gradient' iterD = {"iteration": iter, "function_value": function_value, "point": map(lambda x: round(x, 3), point), "gradient": map(lambda x: round(x, 3), gradient), "method": method, "delta": map(lambda x: round(x, 3), delta), "Hessian_Gradient_Sq": hess_grad_sq} iterL.append(iterD) point = map(lambda i: point[i] + delta[i], range(dim)) hybrid_descentDF = pd.DataFrame(iterL) return hybrid_descentDF

5

null

2018-06-05

2018-06-05 20:54:12

2018-06-05

2018-06-05 21:02:41

2

false

en

2018-06-07

2018-06-07 02:03:10

2

1e3a44594171

1.673899

4

1

Newton = Gradient + Hessian

3

Newton’s Method vs. Gradient descent Newton = Gradient + Hessian See https://medium.com/@ranjeettate/optimal-learning-rate-from-the-hessian-examples-e89f8d1af977 for some background. See https://en.wikipedia.org/wiki/Newton%27s_method_in_optimization for a complete derivation of Newton’s Method and how it uses the second derivative (the Hessian) to improve Gradient descent. Post in progress, but meanwhile, here are the pictures: Hybrid and gradient descent paths on the contour plot Function values approaching the minimum Key part of the code for hybrid descent, note that it includes Gradient descent as well as Newton’s Method (or what I was calling Hessian descent), code for which I haven’t included separately.

Newton’s Method vs. Gradient descent

15

hybrid-descent-vs-gradient-descent-1e3a44594171

2018-06-16

2018-06-16 18:17:21

https://medium.com/s/story/hybrid-descent-vs-gradient-descent-1e3a44594171

false

342

null

Machine Learning

machine-learning

Machine Learning

51,320

Ranjeet Tate

I stop to miau to cats.

7c224ff58a97

ranjeettate

148

158

20,181,104

null

0

null

0

null

2018-07-19

2018-07-19 08:07:25

2018-07-23

2018-07-23 03:06:01

7

false

en

2018-07-23

2018-07-23 03:06:01

8

1e3bace175cd

2.800943

6

0

Elon Musk has long been a proponent of AI safety and ethics so it’s no surprise to see him taking a pledge to never create deadly weapons…

5

Build your own Computer Vision model, Plotly 3.0.0 Released, Google’s Free Pose Estimation AI, and more Machine Learning Developments! Elon Musk has long been a proponent of AI safety and ethics so it’s no surprise to see him taking a pledge to never create deadly weapons with AI. Three co-founders from DeepMind and over 2000 other AI leaders and researchers have taken the pledge with him. Great news to start the week! Other highlights from the past week: Plotly 3.0.0 launched with some amazing new interactive features, NVIDIA’s use of machine learning to convert standard videos into incredibly beautiful slow motion, and more updates below! You can get these AVBytes articles delivered straight to your inbox on a daily basis! All you need to do is subscribe here. We’ll do everything else. :) Click on the headline to read the full article. Google’s Latest AI Experiment is Fun, Free and Combines TensorFlow.js and Pose Estimation: Combining the power of TensorFlow.js and PoseNet, Google’s latest experiment matches your pose with it’s catalog of 80,000 images! It’s entirely based in your browser, free and a whole lot of fun. Give it a try! Evolutionary Algorithm: The Surprising and Incredibly Useful Alternative to Neural Networks: Neural networks are at the core of most of the breakthroughs in machine learning these days. What comes after neural networks though? A new type of algorithm, called Evolutionary Algorithm, has been developed that could significantly change the way we build and design deep learning models. This article looks at how it works! Elon Musk, DeepMind Co-Founders and other AI Leaders Pledge Against Developing Autonomous Weapons: With the rapid adoption of deep learning and AI, ethics and safety are a growing concern. So top AI leaders, including Elon Musk and 3 DeepMind co-founders, have take a pledge to never develop deadly AI systems. Plotly.py 3.0.0 Launched with Major Update — a Must-Download for all Python & Visualization Users!: Do you perform data visualization? Great news — Plotly.py 3.0.0 has been released and it includes some AWESOME features! It’s a must-download for anyone who does data visualization, especially in Python. NVIDIA’s Machine Learning Model Converts a Standard Video into Stunning Slow Motion: Here’s a must-read for all data scientists! NVIDIA’s model, using computer vision and neural networks, can convert a standard video into incredibly detailed and high-quality slow motion. Check out the video and research paper inside! Build your own Computer Vision Model with the Latest TensorFlow Object Detection API Update: The latest TensorFlow Object Detection API is out! This update makes it even more easier to build your computer vision model and also includes TPU support, and several pretrained models to get you started. The above AVBytes were published from 16th to 22nd July, 2018.

Build your own Computer Vision model, Plotly 3.0.0

31

elon-musk-deepmind-ai-computer-vision-nvidia-avbytes-1e3bace175cd

2018-07-23

2018-07-23 03:06:01

https://medium.com/s/story/elon-musk-deepmind-ai-computer-vision-nvidia-avbytes-1e3bace175cd

false

464

null

Machine Learning

machine-learning

Machine Learning

51,320

Team AV

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

c7c686fcd4b

analytics

1,907

124

20,181,104

null

0

null

0

null

2017-11-23

2017-11-23 15:58:50

2017-11-23

2017-11-23 00:00:00

0

false

en

2017-11-23

2017-11-23 15:59:19

14

1e3d94adf9a1

1.984906

0

The self driving technology has been exploring by almost every tech giant and Apple is no exception. We have already got hints in past that…

5

Apple Used Machine Learning To Power Up LiDAR In Self-Driving Car The self driving technology has been exploring by almost every tech giant and Apple is no exception. We have already got hints in past that the iPhone maker is secretly working on its autonomous vehicle technology. According to the latest paper published in Cornell’s arXiv open directory of scientific research submitted by Apple shows the company’s new breakthrough in the self-driving project. The paper revealed that Apple has successfully utilizes the machine learning to translates the raw data collected by LiDAR into 3D objects including the cars, cyclist, and pedestrian and eliminates the need for any other sensor. Apple Unveils its new Series ‘The Planet of Apps’, First Episode Released LiDAR is the technology used to identify the object and at what distance it is present by mapping out the whole information. It uses the laser pulses to scan the 3D objects and gathers all the needed information to tell the control system about the detected object. The technology is used by the all the companies exploring the self-driving technology. However, Apple might have cracked the way to reduce the sensors load from the technology with the help of machine learning. The company is working on the autonomous vehicle under the hood and don’t want to reveal much unlike Uber who has just signed a deal with Volvo for self-driving car productions. But, Apple can’t deny the fact that is has requested the self-driving test permit from the from the California Department of Motor Vehicles. Moreover, the company’s future self-driving car has been spotted many times for testing purpose. Twitter Video of Apple’s Secret Self-Driving Car is ‘The Thing’ to know Now, Apple is using a different approach to complete its project by sharing the machine learning idea in its research papers with another community. The main aim behind the step is to get in contact with more intelligent minds who are looking to work with a bigger platform. By revealing its machine learning integration to the LiDAR for optimum results the company wants to invite more people to the team for speeding up the project. However, the true intention of Apple is still unknown behind publishing the research paper. The paper authors Yin Zhou and Oncel Tuzel, who most probably working on the Apple self-driving car shared their creation called VoxelNet. The device which is used to extrapolate the objects information collected by the LiDAR. Google is Giving Billions To Apple To Stay Default Search Option in Apple Devices Keeping the company’s motive aside for publishing the paper, the implementation of machine learning makes LiDAR more effective in scanning the objects. The LiDAR is paired with other motion sensors to design a map showing all details to help the self-driving system to make its way. This information will also help the other companies working on the technology to explore new possibilities of enhancing their self-driving vehicle performance. For the Latest Mobile App Trends and Mobile App Reviews, follow MobileAppDaily on Twitter, Facebook, LinkedIn, Instagram and Flipboard. Originally published at www.mobileappdaily.com on November 23, 2017.

Apple Used Machine Learning To Power Up LiDAR In Self-Driving Car

0

apple-used-machine-learning-to-power-up-lidar-in-self-driving-car-1e3d94adf9a1

2017-11-23

2017-11-23 15:59:20

https://medium.com/s/story/apple-used-machine-learning-to-power-up-lidar-in-self-driving-car-1e3d94adf9a1

false

526

null

Self Driving Cars

self-driving-cars

Self Driving Cars

13,349

Mobileappdaily

MAD provides information on several mobile apps such as travel apps, police apps and game apps. visit: https://www.mobileappdaily.com

a9125211f16f

mobappdaily

96

343

20,181,104

null

0

null

0

661955b6291e

2018-05-03

2018-05-03 06:50:53

2018-05-25

2018-05-25 04:29:54

1

false

en

2018-05-31

2018-05-31 00:37:23

1

1e4039d5b62a

5.520755

2

0

(Note: This is a work-in-draft blog post on my work so far on this model. Feel free to comment on sections that you think I can explain in…

3

Using machine learning / AI to recommend which stocks to buy (Note: This is a work-in-draft blog post on my work so far on this model. Feel free to comment on sections that you think I can explain in more detail.) I’m an avid investor, but I always have trouble deciding which stocks to buy. I decided to build a simple Machine Learning model to recommend stocks. To start simply, I wanted to see if ML models can predict which stock will most likely increase by 1% within 5 days. In general, ML models are good at identifying patterns in data. I am wondering if ML models can identify patterns within stock price movements, much like a trader who does technical analysis. For this task, I built a Long-Short Term Memory model, which is a type of Recurrent Neural Network. The LSTM model will identify the probability that a certain stock will increase by 1% in 5 days. After making these predictions for multiple stocks, we can choose to invest in the stock with the highest likliehood of increasing. A big cautionary note: I built this model for myself, so don’t blame me if this model doesn’t work out for your own trading! If you’re you’d like to first understand the basics of how neural networks work, please refer to my easy intro post on the subject. The Hypothesis A Machine Learning model can predict if a stock will rise 1% within 5 days better than random guessing can. What our neural network looks like I’m going to jump right in and show you a diagram of the neural network I will use today: If you haven’t seen this type of diagram before, let me explain how to read it. This diagram is showing the way our inputs (the X variable) will flow through the neural network to arrive at our output (the Y variable). Through the model, the data is transformed a number of different ways, as indicated by the types of transformations (LSTM, ReLU, Dropout, etc.). What makes neural networks different than a run-of-the-mill “complicated” formula is the weights that are stored within each of the green dots. These are the weights that are updated on each pass of training to slowly match closer to the data — allowing the machine to “learn” from the data that is fed to it. I’ll explain each of the three sections — The Inputs, The Model, and The Output — in more detail. The Inputs We will feed the model a 90-day history of price movements, with 5 different price movements for each day. My belief is that the more information we provide the model, the better it’s predictions will be. The 5 different price movements (for each day) include: Open to High price % change (H/O) Open to Low price % change (L/O) Open to Close price % change (C/O) (Previous day) Close to (present day) Open price % change (O/C) (Previous day) Close to (present day) Close price % change (C/C) In other words, we’re feeding 450 data points to create a single prediction. Let me provide a concrete example for how this prediction model will work. Imagine today is 2 Mar 2018. I want to predict if the stock AAPL will increase by 1% within the next 5 days. I’ll feed the model with price movements from last 90 trading days, which happens to be 27 Oct 2017 to 1 Mar 2018. In practice, preparing the inputs took the most time to build this model. I talk about preparing The Inputs in more detail in another post. The Model Here comes the fun part. I’ll walk through what each of the steps do in this model: First LSTM layer: The data is first fed through a sequential Long-Short Term Memory layer. This layer has our usual neural network weights in them. But they also have additional “gates” that allow information to “remembered” more easily. In general, recurrent neural networks have a tendency to “forget” older information. These memory gates in the LSTM model help solve this issue by putting in additional weights that carry forward older information. This will be important in case there are some price movement patterns early in the history that help inform the final prediction. There’s another great thing about these gates — which is that they also have “learnable” weights! Meaning, we don’t have to manually dictate which memory node should equal 1 (to pass the information forward) or 0 (data is inconsequential to future prediction). The model will automatically update these gates based on the data fed to train it. However, it does mean that there are more weights to learn, which in turn means we require more data to train it. Rectified Linear Unit (“ReLU”): This is a function that returns a positive number if the input is positive. Otherwise, it returns 0. This function has been proven to speed up the model training process. I’ll paste the source here as soon as I find it. Dropout (20%): This model will “dropout” 20% of our nodes everytime we feed a data sample through the model. This helps reduce the overfitting of our model by simulating “missing” information that often comes with our noisy data. Second LSTM layer: Similar to the first LSTM model. However, notice that the final output of this layer is a single output at the end of day 90 to our next ReLu transformation. This is because we only need a single prediction after 90 days, not 90 different predictions after every day. ReLU: Another ReLU transformation like above. Dropout (20%): Another dropout transformation like above. Fully-connected network: This is our typical (non-sequential) fully-connected network. We use this network to reduce our multiple nodes to a single node, because our output only has a single number. Sigmoid: With the single number, we want to transform it to a range that can output a probability to be close to 0 or 1, our true predictions. This sigmoid function does that for us. This is considered a “2-layer” LSTM model because of the 2 rounds of LSTM transformations that we put in. A 2-layer model should be able to identify more complicated patterns than a 1-layer model can, but it has a higher tendency to “overfit” the data and may require more data to train. The outputs of our test data might suffer from this. So, did it work? Here are the training results from our model. Our model’s accuracy increases as it’s being trained on the training data. That’s good! Simultaneously, our model loss is decreasing (also good!) As we fed our model with our training data over over again, the model’s accuracy increased and the “loss” function decreased. Basically this means our predictions are getting better on the training data. Great! But not so fast… Our Y_test prediction (the model’s prediction using the test data) is lower than the average # of times the stock will sell. When we feed our test data through the model, the accuracy is worse than random guessing! This is most likely due to overfitting. Our model learned the patterns in our training data too well, but it could not generalize to data outside of the training set well. Can we improve this model? Definitely, yes. This is a very simplistic model, though it still took a long time to pre-process the data, as always! There’s more complex features we can build into the model that will hopefully inform the predictions better. My immediate next step is to feed non-sequential data into the model, such as the stock name, so that it can distinguish unique patterns with each stock. Right now, we’ve erased the stock identity from each X variable, so the machine thinks the trading pattern of GE is similar to the trading pattern for FB. This is most likely not the case. We’ll need to incorporate the stock details into our X variables for the machine to better differentiate. Other possible improvements include: Adding in other fundamental analysis data, such as Earnings per Share, revenue or net income growth, etc. Adding in other non-standard signals, such as aggregate analyst ratings Adding in stock sentiment analysis (ie. from Twitter, etc.) Anything else? Hope you enjoyed this article and learned a thing or two about LSTM models, or machine learning models in general. Please feel free to give me feedback about this post — what would you like to see more of? Did some parts not make sense? Any parts that I should delve deeper into? Let me know! P.S. If you’re interested, here’s the code to the model above.

Using machine learning / AI to recommend which stocks to buy

2

recommending-which-stocks-to-buy-with-machine-learning-ai-1e4039d5b62a

2018-05-31

2018-05-31 00:37:24

https://medium.com/s/story/recommending-which-stocks-to-buy-with-machine-learning-ai-1e4039d5b62a

false

1,410

My exploration in the tools provided by AI and machine learning

null

Young and AI

young.chingjui@gmail.com

young-and-ai

AI,MACHINE LEARNING,TUTORIAL,DATA SCIENCE

null

Machine Learning

machine-learning

Machine Learning

51,320

Ching Jui Young

AI and data science practitioner

a4ec12caa4a4

youngchingjui

96

143

20,181,104

null

0

null

0

null

2017-10-23

2017-10-23 05:01:46

2017-10-23

2017-10-23 13:45:23

1

false

en

2017-10-29

2017-10-29 00:13:24

7

1e40d5866411

15.120755

4

0

1

This tutorial aims to be the simplest machine learning tutorial on the Internet that covers the most common points of misunderstanding. I…

5

Core Understanding: Machine Learning Math I This tutorial aims to be the simplest machine learning tutorial on the Internet that covers the most common points of misunderstanding. I dare you to find a simpler one that covers more — ok maybe not dare, maybe something less contractual like ‘wonder if you could’ or ‘idly posit’. That reminds me - time for a disclaimer. Disclaimer This series is not intended to take you through all the details or possibilities of machine-learning, but to build core understanding that you can then confidently apply to decode other tutorials. My goal by the end of this series is for you to get to a place where machine-learning terms, math, graphs, tools, & code don’t look foreign to you anymore. If you’re already there, congrats — this post is for newbies (like I was until recently!) At the end of the series, you’ll also find a list of links to good resources that have my stamp of approval for beginners to machine-learning, which will introduce you to other examples and more advanced concepts. Machine-learning will not only enrich your career with an abundance of job security, smart coworkers, & interesting projects — it also happens to be insanely cool & interesting. Welcome to the metallic jungle. Now’s a good time to grab a snack or refill your drink & dive into your favorite chair or pillow fort— you’re about to get 100x smarter, so buckle up & prepare yourself for your brand new bragging rights. I usually start my posts off with definitions, but this time I’m going to try something different. Secret Time I’ve decided to tell you a secret — but only if you promise not to tell anyone else. Promise? Ok, here goes: You’ve probably already done the ‘complex’ types of logic and math involved with machine-learning. If you’ve ever come up with a way to hide your diary, keys, or passwords from your friends who’ve had a ‘hilarious’ prank idea, you’ve already solved an information security problem. If you’ve ever identified an inconvenient process and thought of a way to get other people to do it, you’ve solved a programming problem. If you’ve ever estimated the likelihood of an event occurring or noticed a pattern, you’ve done statistics. If you’ve ever explored the impact something has on something else, such as the impact of your movie habits on your budget, you’ve done algebra. If you’ve ever come up with a simpler rule than the complex rules other people taught you that has the same impact, you’ve done linear algebra. If you’ve estimated how fast something is changing, you’ve solved a calculus problem. If you’ve dodged a ball thrown at you, you’ve calculated the solution to a physics problem. If you’ve ever compared different combinations of things, like playlists, you’ve done combinatorics. And you’ve probably also done the math directly involved with machine-learning too: If you’ve analyzed the features of something that differentiate it from other objects and come up with a rule to group them into categories, you’ve built a model. If you’ve thought about other problems that relate to your current problem, you’ve identified a problem type. If you’ve chosen a strategy when playing a game, you’ve already selected an algorithm that’s appropriate for a problem type. If you’ve ever predicted your friend’s behavior based on your memories of their behavior within your social group, you’ve implemented a neural network. Machine learning isn’t doing some magical type of math no one understands. The math behind machine-learning (and other kinds of math) are not too difficult for you — you’ve been doing these kinds of math for much of your life without even showing off about it. Competitive people might try to scare you with fancy terms or symbols to see how tough your brain is — but as you get good at translating those fancy things into the simple concepts and rules driving them, you’ll start recognizing the advanced math behind your own decisions, and realize you’ve been doing this all along. Definitions Now that we’ve broken down the math a little, let’s define machine-learning, then the terms necessary to support that definition. If you already know all this stuff & want to fast-forward to a quick intro on the kind of thinking people do when applying machine-learning, that section is called ‘Decision Rules’. Machine-learning: delegating math to a program in order to extract insights from data. True AI would be able to generate the rules governing a system without data (using insight patterns & category descriptions to generate insights about a system, where the insights are the system rules and categories are the system nodes), but we’re not there yet in our tool ecosystem. Currently our AI programs can still do amazing things, even though they rely on data. The math necessary to build machine-learning tools includes linear algebra, calculus, and statistics. Note: you won’t need to understand most of this math to use machine-learning tools for basic tasks. Machine-learning libraries do this math behind the scenes, and with the most popular libraries, all you have to do is configure the model, point it to your data, and they do the advanced math for you. If you want to skip to a tutorial for a basic example with hardly any math, this is a good place to start: https://pythonprogramming.net/linear-svc-example-scikit-learn-svm-python/ Linear algebra: used to describe relationships between forces (vectors) and operations on these forces, usually with the aim of approximating solutions to functions Example: you build a matrix to track the operations you use on a set of forces, and then use this matrix to find a simplified version of that set of forces. Calculus: used to calculate relationships between functions. Related functions may produce descriptive metrics of the original function, like rate of change or area. Example: you might calculate the derivative of a function to find its rate of change, and the integral of a function to calculate the area it represents. Statistics: used to fit patterns to data to help make predictions You might use statistics to find the formula for a line that fits a set of points the best (which is called Regression), then examine the line formula, and try to determine which real-life variables go in which positions of the formula, so you can figure out if this formula will be a good predictor of future data points. Moments: the key points or limits that determine the identity of a function or force. I think of moments as the minimum amount of information that you would need to reconstruct an object. For people, our moments may include our memories & our DNA. What do you consider to be your other defining characteristics? You might call those your moments — they’re essential to your identity and differentiate you from other people. In a similar way, we refer to moments of a function, usually to reference function metadata used to determine its identity — metadata like rate of change, pivot points, limits, function type, input variables, etc. Model: this is one of those words that’s so frequently overused that it’s misunderstood solely because someone heard this word used for a totally different thing that seems related. Let’s focus on the relevant machine-learning & math definition: a model is a theorized relationship (or rule) evaluated by the algorithm you chose. For a simple intro to models, skip down to the section below called Examples. Features: explanatory variables represented in your data; if you wanted to identify which animals a person likes the most, and you only have data on people’s personalities, the features you choose might include numerical scores of personal qualities like ‘friendliness’, ‘patience’, ‘aversion to noise’, etc. Labels: categories in your data that you want to identify; if you wanted to identify which images in a set of pictures show a cat or dog, your labels would be ‘Cat’ and ‘Dog’. Target: the correct label category that you’re trying to group data into; for example, the ‘Dog’ label would be the target attribute in a supervised model, if you wanted to identify the dogs in a set of images. Supervised vs. unsupervised vs. reinforcement learning: you’ll only assign your own target attributes for supervised learning, which is where you know what categories you want to group data into. For unsupervised learning, the program will try to identify groups of data that you might want to use as categories. Reinforcement learning is where you give positive or negative feedback to results in real time, and the program adjusts its categorization based on each new round of your feedback. Regression is a form of supervised learning, because when you teach a program how to fit a formula to data, you know what you’re looking for in advance. Unsupervised learning is like a double-blind study, where neither the participants or the scientists knows the right answer, and therefore can’t insert their bias into the experiment; supervised learning is where you know the answers but your code does not. Neural Network: a set of ordered node groups (called layers) where nodes have a state value (0 or 1) and weight to represent the influence its value should have in the next layer of processing, and where the links between nodes represent functions executed on the input values in each layer. Picture of a neural network found here: https://ml.berkeley.edu/blog/2017/02/04/tutorial-3/ In the above diagram courtesy of the Berkeley site, each node represents a function, and if there’s a line between a node in one layer and a node in the next layer, that means the next node’s function is executed on the value sent from the original node (if the activation function of the next node returns a ‘yes’ value). In a network with layers that are executed from left to right, the first left-most layers might be lower-level processing, like identifying a pattern in data, and the final right-most layers might be higher-level processing like identifying a category that pattern belongs to. You can see how this would be applied to multiple problem types: creating a machine-learning model that can generate a web app could identify patterns in relevant data models (articles, accounts, history) at the first layer, and patterns in code (UI features, styles, event-handling, and APIs) in the final layer. For natural language processing AI, this might evaluate lower-level processing in the first layer (identifying parts-of-speech like verbs and sentence subjects), mid-level processing in the interim layers (identifying intent or tone), and high-level processing in the final layer (identifying anecdotes, jokes, opinions, and insights). Activation Function: this is a function associated with each node in a neural network that helps decide if that node’s operation should be executed on the input value or not. The softmax activation function tends to be the most popular activation function in neural networks, so you’ll see that term a lot. A discussion of various activation functions is here: https://medium.com/the-theory-of-everything/understanding-activation-functions-in-neural-networks-9491262884e0 Decision Rules: from Data to Insight Now that we have background info on what these terms mean, let’s assemble them into decision rules for applying machine-learning, so we can extract insights from our data. Decision 1: Choosing the target insight type Let’s assume you already have data that’s been formatted, sanitized, standardized, and selected for significance, based on what features you expect will be useful in making predictions. A good intro to why you would want to process your data is right here: https://pythonprogramming.net/regression-introduction-machine-learning-tutorial/ Your first big decision may be choosing which type of insight you’re trying to identify. Insight Types Strange data (called Anomaly Detection) Identifying unknown categories (Unsupervised Learning, which uses cluster-identification to suggest categories) Grouping data into known categories (Supervised Learning, with either 2 classes or multiple classes) Predicting future data (Regression) AWS gives some good examples of problems you’d solve with different models: Types of ML Models - Amazon Machine Learning Amazon ML supports three types of ML models: binary classification, multiclass classification, and regression. The type…docs.aws.amazon.com Decision 2: Choosing Parameters You’ll need to choose parameters for your model based on: characteristics of your data number of data points, number of variables, number of categories whether you’re predicting a quantity, category, or identifying categories obvious data clusters & whether categories are known or unknown whether the order of categories matters what level of approximation is acceptable whether you want to predict one value or a set of values whether the difference between categories should be particularly clear — known as ‘explainable class boundaries’ whether features are numerical 2. the insight type you’re aiming for: predicting future data, categorizing data, identifying categories, or spotting strange data. 3. the priorities of your model performance (speed) accuracy/precision specificity sensitivity bias/variance trade-off degree of under/over-fitting Decision 3: Choosing an Algorithm There are many study guides & tutorials out there that provide good decision flow charts to help you choose an algorithm, like this one from Microsoft: Machine learning algorithm cheat sheet The Microsoft Azure Machine Learning Algorithm Cheat Sheet helps you choose the right algorithm for a predictive…docs.microsoft.com As your understanding of these algorithms evolves, you may end up choosing your own functions rather than the default functions recommended to you by a tutorial or library: data-processing functions activation function for nodes in a neural network cost function for nodes in a neural network while other decisions have a clear optimal function for certain data types & target insight types. While those resources are helpful for people who are ready to start building their model, for everyone else, let’s go over the math behind choosing a model with a simple example. Examples Why would you choose one model over another? Let’s look at the difference between discrete vs. continuous variables for a quick insight into why some models are better than others for different problem types. Math newbies might not see the reasoning behind why a certain feature of a problem space points to a certain algorithm. But there’s a clear answer on why you would use a discrete model (with limited values - typically integers) to classify objects, and a continuous model (with unlimited values - typically any value within a range) for values that show up in degrees on a spectrum. What would happen if you didn’t use the clear right answer? If, for instance, you used a continuous model to classify objects, which is clearly better represented with a discrete model? Let’s think about a sample problem, and then try to solve that problem using the ‘wrong’ type of model for that problem type. Example 1: Cats vs. Dogs Say you need to find out who’s a cat and who’s a dog, and you want to use a continuous model to assess ‘degrees of catness’ and ‘degrees of dogness’ — even though you’ve never seen any half-cat, half-dog creatures outside of cartoons (except for that one who clearly had sympathies with their enemy forces). Let’s evaluate this continuous model based on degrees of catness & dogness: would this actually help you group animals into the cat & dog categories? What if one animal has an equal degree of cat & dog qualities, or if they’re right under the qualifying degree of dogness, and people always mistake that animal for a dog — do you kick them out of the dog category even though everyone’s convinced they’re a dog & they’re right underneath the qualifying dog score? If not, then where do you draw the line between cats & dogs? It’s difficult to decide on a fair separating line on a continuous model. Example 2: Good vs. Evil Or maybe you don’t care about the cat vs. dog wars but you do care about the war between good and evil, so you decide to calculate who’s in the ‘good’ category and who’s in the ‘evil’ category. Let’s evaluate this discrete 2-category model for a match with reality: does categorizing people into two groups match up with reality? Can you think of anyone who is all good or all evil? In real life, we find, there are clearly degrees of good & evil, and two categories isn’t enough to truthfully represent reality. So you could use the ‘wrong’ model — it’s not really wrong, in the sense of being unethical, unless you consider misrepresenting reality to be unethical, and then it totally is wrong. But evaluating the problem using the wrong model actually gives us insight into why some problems still exist in society. There really are people who try to group others into overly simplified ‘good’ and ‘evil’ groups, even though most people fall somewhere in the middle, and this problem type is clearly better reflected by a continuous model. And similarly, there really are people who might believe in the mythical Cat-Dog creature, despite reality telling us that they are two different species that split from a common ancestor and have had difficulty getting along ever since. So there’s no reason to shy away from using the ‘wrong’ model for a problem type, if your goal is to gain creative insights into how variables might be related in other people’s brains — but if your goal is to extract insights that help you find the true relationship between variables, using the right model will help speed up that process. Example 3: Static vs. Dynamic Another example of a model applied to the ‘wrong’ problem space: using static values (what programmers call hard-coded values, or constants) when you should use dynamic values (what programmers call variables). For example, if someone says something I don’t like, and I look in my brain’s folder of static priorities, & spit out an overused argument that I didn’t adjust for this situation instead of checking my priorities or analyzing what the person said to see if it’s true before making an argument — that’s likelier to lead to a misrepresentation of truth, instead of learning something new and pivoting to a perspective that is cohesive with this new fact. But maybe you’re the smartest person in the universe and you’ve researched machine-learning to the extent that you have a brain folder with all the models in the world — in which you could rely on that static list instead of coming up with a new list for every new situation. However, most of us are not like you, so for the rest of us mere mortals, it’s better if we apply dynamic analysis to each situation, rather than relying on static values & rules. So the model ‘use static values instead of dynamic values’ doesn’t fit for most situations. Bias in Machine Learning This static vs. dynamic choice is actually relevant to machine-learning — if a data set is biased (in favor of our evil cat overlords), then the program will fail to accurately predict reality. Our brains can do a lot, but our culture & experiences tend to mis-wire the ideal relationships between variables & problem types in our heads. We can fix these mis-wirings with dynamic critical thinking through statistical analysis & machine-learning, rather than relying on static brain default settings. But there are points where this analysis can be open to bias attacks, and the data is one of those points. There are a lot of ‘wrong’ models out there in common problem types that people try to solve using models that don’t adequately represent reality. We’ve all learned to evaluate the source of information; now that we’re all grown up, we do more advanced calculations, like evaluating if a model fits a problem type. But maybe you’ve misidentified the problem type, you might be wondering. Or maybe the problem is too ambiguous to have a clear right answer when choosing an algorithm. How do we know for sure that a model fits or if it’s biased, either from oversimplification, incorrect problem type identification, biased data, or other scientific sins? It’s true that misidentifying the problem types & other aspects of the problem space are problems unto themselves. But it’ll be clear whether a certain model fits based on its ability to predict future data. Simplicity vs. Complexity All the experts out there didn’t understand these concepts at one point either, until they read and re-read and looked at graphs and worked through examples until it clicked. It takes time for your brain to process these things, but if you do yourself a favor and use simple terms & examples when starting out, you’ll speed up your learning process over time. That’s not to say there aren’t legitimately hard problems out there. The difference in complexity between the simplest & hardest problems is just as enormous as the difference between the set of adjectives you commonly use to describe yourself and your true identity, with all of your memories, habits, strategy evolution, decision heuristics, belief systems, perspectives, opinions, culture, social groups, priorities, biases, favorite thought processes, insights, mannerisms, personality, potential, and the relationships between all these factors. So, quite a big difference between the simplest examples and the most mind-boggling problems that remain unsolved. But as you move up the ranks of complexity, step by step, you’ll find yourself becoming a master at learning — almost like a learning machine. Please go easy on yourself as you ascend that ladder. Pick the simplest possible examples when you’re learning a new concept. Our brains tend to like complexity about as much as our arteries like plaque, so if you want to optimize your initial grasp of a concept, you’ll use a simple example. At the other extreme, if you over-prioritize simplicity to the point where you systematically ignore complexity, people may start wondering if you’re a chat bot driven by biased data come to signal the apocalypse. But that doesn’t mean simplicity isn’t useful in the context of introducing your brain to new concepts. Simplicity is where we begin, but the infinite richness of complexity & diversity is where we might end up, if we’re lucky. Choosing Tools For every machine-learning project, you’ll need data, a way to define & tune your model, a way to visualize your data so it’s useful to other people, and an IDE (a tool that helps you write code) to write & test all of this code. Two primary languages can do all of the things you’ll need for machine-learning projects — R and Python. Other languages are catching up but these two languages have the most diverse & capable toolboxes. If you choose Python, you’ll probably want to use a IPython Notebook program, which lets you run & test pieces of code, like web developers run & test snippets of Javascript code in the browser console or in a JS fiddle. Commonly used notebook programs include Jupyter, as well as the various cloud products that allow you to read & run these notebooks in a testing environment, such as GCloud, AWS, and Microsoft Azure. If you want to compare R & Python code for applying various models before deciding on a tool, I found this guide useful: Essentials of Machine Learning Algorithms (with Python and R Codes) Note: This article was originally published on Aug 10, 2015 and updated on Sept 9th, 2017 Introduction Google's self…www.analyticsvidhya.com Onward and Upward In the next post on this series, we’ll go over the process of choosing different algorithms in more detail, as well as how to optimize your model. But for now, I shall stop asking you to read even more words, and wish you a good day.

Core Understanding: Machine Learning Math I

7

core-understanding-machine-learning-i-1e40d5866411

2017-11-13

2017-11-13 08:45:04

https://medium.com/s/story/core-understanding-machine-learning-i-1e40d5866411

false

3,954

null

Machine Learning

machine-learning

Machine Learning

51,320

JJ

Idea Machine, Monster System Fighter, Problem Space-Contortionist, Unknown Systematizer, Meme Advocate, Imaginary American

54bccbf256dd

insighterator

90

16

20,181,104

null

0

null

0

null

2018-08-22

2018-08-22 17:14:21

2018-08-22

2018-08-22 17:16:47

3

false

en

2018-08-22

2018-08-22 17:16:47

3

1e4170c69c89

2.15

0

Machine Learning is one of the trending topics in today’s digital world. It is a type of Artificial Intelligence (AI) in which the software…

5

Difficulties in Adapting Machine Learning Machine Learning is one of the trending topics in today’s digital world. It is a type of Artificial Intelligence (AI) in which the software gives you the ability to learn based on data. For instance, you must have noticed how YouTube, Amazon and many other online platforms recommend similar products or videos you should check out. Those recommendations are based on your previous actions. Nowadays, everyone is looking to incorporate machine learning into their business. The services generated through Artificial Intelligence or machine learning may seem magical, but reaching to that point involves a lot of work and numerous challenges. 1. Requires Quality Data: The machine learning completely depends upon the input provided to it. Data plays the crucial part of machine learning, the machine will learn what has given to it. The data decides how the machine will evolve and what value can be added to your product/project by utilizing machine learning. For instance, on images it scans has small cars, you will not get the accurate results. To achieve success, you need quality images. The machine’s ability to grasp directly depends on the quality of the data it encounters. 2. The Quantity of Data: Not only quality data, but to achieve the expected outcome you need data in high quantities. If you are providing a machine with limited data for hundreds of users per month, the machine won’t have enough information to deliver the same service to thousands of users per months. Its sample might be too small to be accurate. 3. Requires Continuous Testing: Having quality data in sufficient quantity is not enough. Instead, you need to keep testing the data at regular intervals. The machine might learn techniques on its own, but the learning is based on how the machine was designed and the data it’s being fed. 4. It’s Expensive: Machine learning can’t be cheap. So while designing a machine for any of your product/project, keep your budget in your mind. Machine learning experts are high is demand and charge high for their work as there is a lot of effort involved to figure out the best designing and creating the models, offering training to meet expectation, testing them on, etc. The machine learning is not a magic, it needs lots of efforts and dedication to design a perfect model that will suit your requirements. Keep all the points in mind before initializing your AI model. Also Consider Reading: Be the Boss of Robots with Your Updated Skills Future of Project Management with PMP

Difficulties in Adapting Machine Learning

0

difficulties-in-adapting-machine-learning-1e4170c69c89

2018-08-22

2018-08-22 17:16:48

https://medium.com/s/story/difficulties-in-adapting-machine-learning-1e4170c69c89

false

424

null

Machine Learning

machine-learning

Machine Learning

51,320

Daniel Irwin

null

b154b8ccea41

marketing_99371

0

1

20,181,104

null

0

null

0

32881626c9c9

2018-09-08

2018-09-08 05:04:37

2018-09-08

2018-09-08 05:58:13

1

false

en

2018-09-08

2018-09-08 05:58:13

3

1e41f0353e38

4.033962

1

0

Image credit www.pexels.com

3

AI will one day dethrone human intelligence — An Orwellian myth Image credit www.pexels.com In a Ted Talk on November 30, 2017, Sebastian Thrun said that as an AI person, he had not noticed any real progress on creativity and out-of-box thinking by AI. The hotbed was really at the repetitive end of the AI spectrum of activities — the processing of legal documents, contract drafting, screening of x-rays, the works. Such tasks were being mastered by AI at an incredible speed. However, they remained single-domain masteries achieved by processing humongous amounts of data extracted from historical snapshots. Thrun went on to say that we must realise that AI is after all technology. True. AI is technology, not biology. It can progress at an unimaginable speed, and in the not-too-distant future, it may surpass human capabilities in several domains, but can it evolve and mature like a human brain has, over millions of years? AI will make a child’s play out of translating War and Peace in ten languages, but will it even care to know who Tolstoy was? Again, a deep learning-enabled cancer detection app may spot a malignant mole with a superficial body check, without a biopsy. But can that app achieve the full-scale capabilities of an oncologist? One is technology — metallic and hard-wired, algorithmic, aeons away from originality, creativity, and compassion. The other is biology — organic, intelligent, emotional, and growing from cell to cell, achieving out-of-box proficiencies along the way. So, will the hard-wired world of algorithms, digitisation, and data-isation take over the mysterious universe of electronic impulses and chemical reactions? Man vs machine — Human intelligence’s unfair advantages over AI It is a lost battle for AI, really. The odds are greatly against it. AI does not possess the 3Es — Experience, Expression, and Emotion “I don’t think we will ever see a robot telling us a really great joke,” Ken Goldberg remarked in a WEF talk in 2015. And really, we haven’t come across any, since then. Simply because cracking great jokes needs spontaneous wit, and wit needs creativity, and creativity needs originality, and that is where AI proves to be a problem child. It is dumb. It can replicate, but it cannot originate ideas. It can learn, not invent or innovate. And it can observe, not emotianlise. Human intelligence will forever remain ahead of AI on the 3Es — Experience, Expression, and Emotion. Self-driven cars, a flagship AI initiative, will be a road-reality by 2020. Great! These cars will epitomise safe driving and reduce road accidents and casualty significantly. But will one such car stop if a stranded backpacker waves frantically at it on a remote highway? It won’t. Its environment-sensing capability will only identify an unknown behaviour by a human figure. Probably, its laser sensors will miscue the distress of the facial expressions and the urgency of frantic hand-waving as a potential threat, and press the accelerator! Sensors and radars and GPRS and odometry apart, a more important differentiator in this hypothesis is that the machine is devoid of empathy to stop and help a distressed fellow-human. Human intelligence is binary, AI is not Spearman’s two-factor theory of intelligence consists of the G factor for general intelligence, and the S factor for specific intelligence. In this binary structure, the G factor resides at the core of our system of cognitive abilities, around which the peripheral abilities of the S factor orbit. Each of us has a central cognitive ability, depending on what kind of an intellectual we are — kinesthetic, musical, logical, and so on. However, we all perform scores of other types of intellectual tasks equally well. A person with kinesthetic intelligence also composes a wonderful symphony. Ashton Kutcher is a brilliant biochemical engineer. He now works as a product engineer for Lenovo. Can we transplant Spearman’s theory in AI? The answer is an emphatic no, and it comes from the proverbial horse’s mouth. Sebastian Thrun points out that successful AI is all about super-specialised, singular functions — “the Google car can never do anything else, it can’t even control a motorcycle.” That’s the second inherent advantage that human intelligence has over AI. Our brain is the greatest marvel of all times. It juggles and multi-tasks with magical efficiency. In comparison, AI is severely handicapped in being a single-domain expert. If anyone dares to challenge this, they just have to put a logistics robot in a cancer research laboratory. AI cannot achieve the human scale of consciousness For anything to come close to human intelligence, its proportionate consciousness is a prerequisite. Consciousness is self-awareness. One stands in front of a mirror and knows that it is she or he. From a spiritual and philosophical context, one goes into deep meditation and transcends to an alternate reality, creating a higher level of self-awareness. Consciousness per se requires the 2-way flow of interactive patterns between the brain and the surrounding ecosystem to create reality. The Copenhagen Interpretation says that the physical world and consciousness are the two sides of the same coin of reality. At the roots of consciousness lies biology. And AI is technology. It is equivalent to stones and rocks in the sentient context. It is not self-aware. It is devoid of consciousness. It does not know the consequences of its actions. How can such a superficial entity cobbled up in the last sixty years compete with the profoundness of human intelligence achieved over millions of years of evolution? The only way that AI can even start thinking of beating human intelligence is by evolving the brain-way, through cellular growth, electronic impulses, and chemical reactions. Possible? That brings us back to where we began — AI is irreversibly technology-oriented, not biological or cognitive. It is born of algorithms, it feeds and grows on data, out of which patterns are discerned, out of which machine learning and deep learning are derived. That is why AI cannot dethrone human intelligence.

AI will one day dethrone human intelligence — An Orwellian myth

45

ai-will-one-day-dethrone-human-intelligence-an-orwellian-myth-1e41f0353e38

2018-09-11

2018-09-11 10:53:11

https://medium.com/s/story/ai-will-one-day-dethrone-human-intelligence-an-orwellian-myth-1e41f0353e38

false

1,016

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

null

datadriveninvestor

null

Data Driven Investor

info@datadriveninvestor.com

datadriveninvestor

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

dd_invest

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

Abhay Goghari

An experienced writer, copywriter, and published author. Writes for digital and mainline, new technologies, and mobile experiences. Loves to write long copy.

37e25e18f46f

agoghari

2

3

20,181,104

null

0

null

0

null

2017-09-04

2017-09-04 10:03:39

2017-09-04

2017-09-04 10:05:21

1

false

en

2017-09-04

2017-09-04 10:05:21

1

1e426b87eab5

1.822642

0

The invasion of chatbots in HR has triggered a lot of debates about the foreseeable future of the HR work. Will HR work completely die…

5

Will HR work die by Chatbots coming The invasion of chatbots in HR has triggered a lot of debates about the foreseeable future of the HR work. Will HR work completely die? Will the future of Human Resources be completely taken over by the Chatbots? And lots more… Chatbots Well, in simple terms, a Chatbot is a computer program that is designed to simulate conversation with human users, especially over the Internet. They come in various forms and could designed to be as simple or as sophisticated as required. In whatever form they are used, they are becoming wildly popular by taking up the repetitive, mundane tasks that take up a lot of time and efforts. Chatbots can make a difference in the way things work in the HR department. They can help employees be constantly connected through their mobile devices and have the advantage of being connected 24/7. They will definitely make communication faster and easier. · Chatbots can provide quick and customized answers to potential applicants about the company and can provide quick and customized answers to the questions that any visitor could have. · They could provide organizations with a quick, customized and automated way to conduct performance reviews and prove to be great time savers. They can accumulate precise data from employees and ensure that the process is both efficient and effective. · HR teams usually send out a lot of information to their employees. Sometimes they forget to reply to important questions. An HR bot can automate this and ensure that no queries from employees are missed. · Chatbots have more advanced functions than simply answering questions. They can be used to gather employee information and other data that can help an organization take informed decisions. They can collect real time data to analyze what kind of questions do employees ask more frequently and can help determine the problems easily. This will reduce the frequently asked questions and increase employee satisfaction rates. The Verdict Although chatbots have many sophisticated features to complete a lot of mundane HR tasks, they aren’t as intelligent so as to take over the entire HR function. So, although these chatbots are invading and slowly encroaching the HR functionalities, they are not yet designed to completely take over. Nonetheless, the use of chatbots to aid, assist and replace a few HR related tasks definitely makes things easy and saves a lot of time too. If the artificial intelligence is designed well, they have a very promising future in the HR world! For more information please visit: http://blog.vervesys.com/will-hr-work-die-chatbots-coming/

Will HR work die by Chatbots coming

0

will-hr-work-die-by-chatbots-coming-1e426b87eab5

2018-04-27

2018-04-27 18:38:27

https://medium.com/s/story/will-hr-work-die-by-chatbots-coming-1e426b87eab5

false

430

null

Chatbots

chatbots

Chatbots

15,820

Verve Systems

#USA based #IT firm offering #Cloud #IoT #Bigdata #MobileAppliction and more

b02fc25388f6

vervesys

6

8

20,181,104

null

0

null

0

3a054cf431f7

2018-08-17

2018-08-17 10:14:02

2018-08-17

2018-08-17 10:35:33

3

false

en

2018-08-17

2018-08-17 10:35:33

0

1e4323fd3a05

1.168868

2

0

Bottos DPOS Node Winners

5

Bottos DPOS Node Election Results Bottos DPOS Node Winners We are very pleased to announce the final winners of Bottos super nodes! There were 49 teams elected. BVP team, the first prize Andromeda, the second prize BTO Fans Group, the third prize Please review the following list for your reference, 1.BVP志愿者团队（176,633 votes） 2.石墨烯应用中心（171,356 votes） 3.BTO护卫队（154,063 votes） 4.铂链粉丝队（101,518 votes） 5.链世界（97,337 votes） 6.善济 — 河南社区（79,853 votes） 7.Bay Max（76,529 votes） 8.铂链七兄弟（55,928 votes） 9.铂链基金会节点1（50,000 votes） 10.铂链基金会节点2（50,000 votes） 11.铂链基金会节点3（50,000 votes） 12.铂链基金会节点4（49,999 votes） 13.铂链基金会节点5（49,999 votes） 14.铂链基金会节点6（49,999 votes） 15.蜘蛛矿池（45,683 votes） 16.US Bottos Superfan Group（41,126 votes） 17.金山云战队（40,000 votes） 18.天鹅（39,035 votes） 19.守望未来（37,257 votes） 20铂爱今生（23,335 votes） 21.Discovery Community Node（32,207 votes） 22.中原区块链（32,096 votes） 23.DT社区（30,902 votes） 24.币市炒手（29,747 votes） 25.Lbank（25,432 votes） 26节点资本Node Captial（25,000 votes） 27了得生态基金（24,900 votes） 28.Lowergroundfloor（19,255 vote） 29.B-ONE（17,595 votes） 30.旺仔（15,132 votes） 31.聚力（14,967 votes） 32.铂客之家（14,490 votes） 33.Galante Crypto（12,000 votes） 34.B-Engine（10,945 votes） 35.约瑟（9,477 votes） 36.酒匠（8,315 votes） 37.米粒量化（7,833 votes） 38.BTO Hippo（7,713 votes） 39.币知道（6,550 votes） 40.区块链山西社区联盟（5,557 votes） 41.孙中亚（4,976 votes） 42.ENIGMA（3,606 votes） 43.CyberMiles（3,402 votes） 44.ijver（3,162 votes） 45.铂达天下（2,665 votes） 46.MAY6（2,298 votes） 47.Bottos France（2,248 votes） 48.German Bottos Community（2,011 votes） 49.BottosKorea（1,993 votes）

Bottos DPOS Node Election Results

51

bottos-dpos-node-election-results-1e4323fd3a05

2018-08-17

2018-08-17 10:35:33

https://medium.com/s/story/bottos-dpos-node-election-results-1e4323fd3a05

false

164

Official Bottos blog

null

bottos.org

null

Bottos

marketing@bottos.org

bottos

BLOCKCHAIN TECHNOLOGY,ARTIFICIAL INTELLIGENCE,BIG DATA

bottos_ai

Bottos

botto

Bottos

0

Bottos AI

Bottos - A Decentralized AI Data Sharing Network

58b32476bec4

bottos_ai

374

8

20,181,104

null

0

null

0

null

2018-04-06

2018-04-06 13:24:41

2018-04-13

2018-04-13 01:39:03

2

false

en

2018-04-13

2018-04-13 01:39:03

4

1e43ee28a5d6

3.628616

7

0

When it comes to machine learning, I have no trouble finding articles about the latest and greatest in machine learning / statistical…

5

Feature Engineering Made Easy When it comes to machine learning, I have no trouble finding articles about the latest and greatest in machine learning / statistical models. What is harder to find are articles pertaining to a part of the machine learning pipeline that data scientists claim to spend up to 60% of their time doing and who also say is the least enjoyable part of the process, Feature Engineering. For this reason, I thought I would dedicate my first ever Medium post to the topic. https://whatsthebigdata.com/2016/05/01/data-scientists-spend-most-of-their-time-cleaning-data/ What is Feature Engineering? Feature Engineering is the process of transforming data into features that best represent the underlying problem, resulting in improved machine learning performance. How we transform and use our data is imperative to the performance of our machine learning pipeline as much as, if not more than, the models that we select to do the learning themselves. I would say there are a few general types of feature engineering that you should know. Getting to know your Features Generally, when you first start playing around with your data, you need to employ some feature understanding in order to identify what you’re working with. For example, are you working primarily with quantitative data (numerical) or qualitative (categorical)? Depending on this, you might need to apply some transformations to your data to make it even usable. You also need to understand the scale of your data and whether or not you are missing any values from the dataset. Once you have a good idea of what you’re working with, you can work on some feature improvement to fill in missing values, transform categorical features into quantitative ones (if your models require it, which they probably do), scale your features, and a lot more. Scaling (z-score or otherwise) is particularly important as it levels the playing field on your features so that single features who happen to be on a larger scale over-affect the machine learning pipeline. For example, if you are trying to predict the price of a used car using the year the car was built and the number of doors the car has, you will have one column with a much larger average value than the other, and that is usually not a good thing for many machine learning models. It is worth noting that scaling doesn’t affect the performance of all models, but it particularly affects models that rely on the Euclidean distance, like K-Nearest Neighbors. After you’ve applied these improvements, you can deploy some feature selection techniques in order to eliminate any noise from your dataset that is interfering with the results of your machine learning pipeline, usually by means of statistical tests such as Chi-squared tests. For example if you are trying to predict the stock price of $AAPL and one of your features is the price of $AAPL 6 months ago, we might not select that as a feature as it probably does not hold much predictive value anymore. If you can’t get a good enough performance out of your machine learning pipeline by just selecting the best features from the original dataset, we can use feature construction to make brand new features. For example, we could construct many features from a text feature like sentiment, length, number of words, etc. Natural Language Processing (NLP) pipelines use feature construction a great deal in order to convert textual data into numerical representations. The Deep Stuff — Feature Extractions The last two types of feature engineering are known as feature extractions. This is where we get a bit math-y. The goal of feature extraction is to obtain a new set of features that are derived from the original dataset through a series of mathematical formulas. These new features are usually not very interpretable by humans but are generally “mathematically superior” in that they enhance the performance of our machine learning pipelines. Feature transformations are a subset of feature extractions and are the process of using (relatively) simple linear algebra to transform our matrix of data into another matrix of data. An example of this would be principal component analysis. The problem with feature transformations is that they generally have a parametric assumption, meaning that the data has to satisfy certain requirements in order for this method to yield “good” results. Feature Learning takes a step further than feature transformations and can use state-of-the-art methods such as deep learning to learn new feature representations without having to conform to any requirements. Examples of these can be found in Word2Vec, Boltzmann Machines, etc. These processes are much more difficult to get off of the ground and will not always produce better results than the aforementioned algorithms just because they are newer. Feature engineering is a step that machine learning engineers tend to dread, but I find the process very enjoyable! It allows us to learn more about our data and the idea of transforming datasets into something new and better is extremely exciting. Learn more about Feature Engineering - now on Amazon! If you want to learn more about Feature Engineering, check out my new book, Feature Engineering Made Easy, co-authored with Divya Susarla, a Data Scientist at my company, Kylie.ai.

Feature Engineering Made Easy

9

feature-engineering-made-easy-1e43ee28a5d6

2018-05-29

2018-05-29 14:59:06

https://medium.com/s/story/feature-engineering-made-easy-1e43ee28a5d6

false

860

null

Machine Learning

machine-learning

Machine Learning

51,320

Sinan Ozdemir

Founder @aikylie / Author on Data Science / Speaker on AI / Fellow @YCombinator

c8b7dc289260

sinan.u.ozdemir

251

518

20,181,104

null

0

null

0

eccc7aae2241

2018-04-26

2018-04-26 02:52:03

2018-04-26

2018-04-26 03:29:12

8

false

en

2018-04-26

2018-04-26 03:30:08

4

1e443553fb51

2.63522

0

Amadeus Code is an artificial intelligence powered songwriting assistant. Create inspiration, song, and melodies instantly and infinitely…

5

Learn how to make a trap track with Ableton Live and Amadeus Code Amadeus Code is an artificial intelligence powered songwriting assistant. Create inspiration, song, and melodies instantly and infinitely. In today’s demo, we import MIDI files exported from the iOS app to create a quick trap track in Ableton Live. Related: Make a quick trap track with Logic Pro and the Amadeus Code app With Amadeus Code, the creation process starts within the app (until we release our VST/AU plugin) but can easily continue in Ableton Live by exporting MIDI files from the app. Let’s get from this: to this: in just a few steps! First, Open Exported MIDI Files Here’s what you’ll see in the exported folder. Harmony MIDI Melody MIDI .m4a Audio To import the MIDI files to Ableton Live, create a new song and drag and drop the 2 MIDI files onto the song. Tip: To import the tempo setting, just check the tempo in the file name. Import MIDI to Ableton Live Looking good so far! Next, Assign Instruments MIDI files from the Amadeus Code app contain program change information. If this conflicts with your synth, we recommend you delete it. Fire up your favorite synth or keyboard and assign it to the melody and chord tracks. We chose a lead synth from Massive and an electric piano for the chords. Fire up your favorite synth or keyboard and assign it to the melody and chord tracks. We chose a lead synth from Massive and an electric piano for the chords. Editing Harmony MIDI Since you have the chord MIDI file, you can easily edit notes to create a pattern. In our case, we edited the chord into an arpeggio and tuned it an octave higher. Next, add some bass Your bass track also can be made with the harmony MIDI file just by creating a new track from the lowest note of the chord. Since the bass notes are rather long (sustained) in this track, we chose a sound with some character. Tip: You’ll find these types of bass sounds in categories named, sequence, rhythmic, arpeggiator, gatedepending on the synth you use. Lastly, add a drum loop. We selected a hip-hop beat and repeated it for the length of the song. And that’s it! What do you think? Leave a comment, ask a question and follow us on twitter and facebook.

Learn how to make a trap track with Ableton Live and Amadeus Code

0

slearn-how-to-make-a-trap-track-with-ableton-live-and-amadeus-code-1e443553fb51

2018-04-26

2018-04-26 03:30:10

https://medium.com/s/story/slearn-how-to-make-a-trap-track-with-ableton-live-and-amadeus-code-1e443553fb51

false

398

Amadeus Code is an artificial intelligence powered songwriting assistant

null

AmadeusCode

null

AmadeusCode

contact@amadeuscode.com

amadeuscode

ARTIFICIAL INTELLIGENCE,STARTUP,SONGWRITING,MUSIC PRODUCTION,MUSIC

AmadeusCode

Music

music

Music

174,961

Amadeus Code

Artificial intelligence powered songwriting assistant. www.amadeuscode.com

cf323832c33e

AmadeusCode

4

1

20,181,104

null

0

null

0

c4db2b652f12

2018-09-17

2018-09-17 03:19:07

2018-09-17

2018-09-17 08:58:12

3

false

ja

2018-09-17

2018-09-17 23:46:51

10

1e44d7d23d76

6.114

0

アマゾン、グーグルと監視型資本主義の台頭（2018年9月15日発行）

5

あなたの検索はすべて我々のものだアマゾン、グーグルと監視型資本主義の台頭（2018年9月15日発行） Mark Stephen Meadows著 HAL9000 art courtesy of Cryteria. Illustration by SEED. 以下は、2010年のスーパーボウルのハーフタイム中に放送された有名な「Parisian Love（パリっ子の恋）」というGoogle広告の中に表示されたテキストです。 study abroad paris france （留学、パリ、フランス） cafes near the louvre （カフェ、ルーブル美術館近く） translate tu est trés mignon（ tu est trés mignon を翻訳） impress a french girl （フランスの女の子に好印象を与える） chocolate shops near paris france（チョコレート店、フランス、パリ近く） what are truffles（truffles とは何か） who is truffaut（truffautは誰か） long distance relationship advice（遠距離恋愛アドバイス） jobs in paris（パリ仕事） AA120 （AA120) churches in paris（パリの教会） how to assemble a crib（揺り籠の組み立て方）この12の簡単な検索から様々なことが分かります。60秒間という短い時間で、ユーザー（おそらく若い男性）が次々とグーグル検索ボックスに調べたい言葉を入力していきますが、その過程で、その人に芽生えつつある恋愛、結婚前提の交際、そして新しい家族ができたことなどがについて分かってきます。この広告は大好評を得て、タイム（雑誌）などから史上最高のスーパーボール広告の一つとして賞賛されました。別のレンズを通して見ると、「Parisian Love」は史上最も身の毛のよだつような広告の1つでもあります。フェースブックの悪名高いケンブリッジ・アナリティカのデータ収集/調査のスキャンダルと比較すれば、人々が自分自身に関する情報をあまりにも多くさらけ出していて、企業がその人々の貴重な個人データを抽出して悪質な目的で使用している様子が概ねお分かりになるでしょう。「Parisian love」は心温まるような内容でユーチューブで数百万回も再生されましたが、どうか次のことを覚えておいて下さい。誰かが言葉を入力して検索するたびに、その裏ではいつもグーグルが耳をそば立てて聴いているのです。もはやあなたを監視しているのはビッグブラザーだけではなく、グローバル企業も監視しているのです。さあ、あのチョコレートや教会や揺り籠の検索についてどう感じますか？新たな個人情報：会話と感性データ今日、アマゾンのAlexaのようなシステムでは、もはや単語を入力せずとも、データシステムが会話を推定できます。今日、我々はコンピュータやデジタルサービスに話しかけていますが、その会話から感情、声調、音域、リズム、口調、さらには会話の「間（ま）」や「uhm」、「ah」、「ah」などの繋ぎの表現などの迫真性の高い正確な個人データを機械学習（ML）システムが取得しています。そのため、Alexaのようなシステムは、ほとんどの人たちが思っているより遥かに多くの個人データという宝の山を集めることが実際に可能です。私たちが口にする言葉や話し方は、私たちが何を感じ、何を思い、何を信じているかを伝えます。これは「感性データ」と呼ばれ、アマゾンとグーグルが購買意思決定の原動力を明らかにするために利用できる感情のデータの「大鉱脈」です。世界で最も大きく成長している小売業者の1つとして、アマゾンにとってこのデータがどれだけ価値があるか容易に想像できますが、それはサプライヤー、広告主、製造業者、購買動向や個々の意思決定を分析する大勢のアナリストらにとっても同等の価値があります。多くの人が、こうしたデータを使って何が行われているのかアマゾンに疑問を呈しています。 Alexaは、最近のアーカンソー州の殺人事件の証拠から、偶然同僚の一人に転送されたある夫婦のフローリングの床に関する会話に至るまで、私たちのすべての会話を私たちの会話を聞いて、録音し、追跡しています。アマゾンはこの最後の例は技術的な不具合が原因だったと言っていますが、それは極めて納得のいかない言い草です。しかし、そもそもアマゾンはデータポリシーについてはあまり協力的ではなかったし、またそのような必要性もありませんでした。なぜなら、新たに出現しているこうした個人のプライバシー問題に対処するためのデータのプライバシー規制や法的枠組みが米国にはないからです。しかし、ブラウザーのクッキーを使用したユーザーの購入活動に関するデータの収集と、その情報の第三者の小売業者への販売に対して、過去10年間にわたって複数の集団訴訟が既に起されてきたことを忘れてはいけません。 Image credit: Consumer Intelligence Research Partners アマゾンのAlexaのスマートスピーカーシステムは、数百万個も販売されています。米国の家庭では、Alexaがこうしたデバイスの市場全体の70％を占めています。簡単なヘルス関連商品の定期購読でエコードットを無料で入手できるようになっています。そして、アマゾンのエコールックというカメラ付きのデバイスまで存在しています。アマゾンのお勧め通りにエコールックをあなたの寝室に置けば、アマゾンのAIがあなたの出かける時の服装についてアドバイスしてくれます。フェースブックが、基本的にはデータ入力アプリであるアカウントを無料で人々に提供しているように、現在アマゾンは自社に私たちの個人情報を送信するために様々な手段を提供しています。アマゾンが収集しているのは私たちの言葉や話し方だけでなく、私たちの気持ち、私たちの意図、そして私たちの信念に対する手がかりです。アマゾンが収集しているのは私たちの心と頭脳です。彼らは私たちの脈を取りながら、私たちの財布にも触手を伸ばしています。私たちがアマゾンを非難することはほとんど不可能です。それは単に監視型資本主義という彼らのビジネスモデルなのです。 Mark Stephen Meadows は Botanic.ioの創設者兼CEO,seedtoken.ioの創設者兼評議員(また作家、投資家、イラストレーター、船乗りでもある）。本稿はSEED TokenのMedium記事「All Your Searches Belong To Us」の非公式の日本語訳です。SEED Vault Ltdに許可をいただいた上で公開しております。

あなたの検索はすべて我々のものだ

0

all-your-searches-belong-to-usあなたの全ての検索は我々のもの-1e44d7d23d76

2018-09-17

2018-09-17 23:46:51

https://medium.com/s/story/all-your-searches-belong-to-usあなたの全ての検索は我々のもの-1e44d7d23d76

false

107

オープンでグローバルな信頼できるbotエコノミーを実現する

null

[非公式]SEED Token Japan

null

非公式-seed-token-japan

AI,人工知能,ブロックチェーン,BOTS,CONVERSATIONAL UI

null

Amazon Echo

amazon-echo

Amazon Echo

3,511

Sky Thunder

null

15e3f4ac3db9

SkyThunder123

1

3

20,181,104

null

0

null

0

f5af2b715248

2018-04-20

2018-04-20 05:48:18

2018-04-20

2018-04-20 12:42:08

5

false

en

2018-05-06

2018-05-06 06:08:49

6

1e499d1dfadb

5.961635

4

0

The flatness of a photograph can’t activate even a fraction of the multidimensional experience of a well-designed scent.

5

Harnessing Olfactory Memory in Product Design In January 2018, I gave a concise presentation exploring olfactory memory at PechaKucha Jerusalem 3.0. In case you missed it, these were the main points: “Smell is a potent wizard that transports you across thousands of miles and all the years you have lived.” — Helen Keller. Let that sink in. Smellfies > Selfies Selfies (or any photograph taken for the purpose of documenting memory) effectively capture specific events. However, in their specificity, they can also have the unwanted effect of diminishing all but the photographed event, as if the surrounding memories, or those like it which came before and after but were not photographed, did not happen. Smellfies: These are taken with your brain, via your nose. Your mind has taken smellfies consistently throughout your entire life and grouped them according to scent, not event. Therefore, every time you smell cinnamon, you have the opportunity to vividly recall every single life experience you’ve had with cinnamon- and the earlier the memory, the more strongly it has been stored. In the way that Google Photos can group your photographs by face, your olfactory memory groups your memories by scent. The flatness of a photograph can’t activate even a fraction of the multidimensional experience of a well-designed scent. The audience then smelled 8 unidentified fragrances on numbered cards for 40 seconds each, coordinating with numbered slides (familiar scents such as olive, cedar, and sweet pea), while allowing memories to surface through the scent experience. I concluded with a recommendation to take advantage of their olfactory recall by remembering to breathe, and breathing to remember. A deliciously honest friend informed me afterward that a big question was left unanswered: How do I actually apply this awareness of the olfactory memory to product design? I’ll warn you that the process goes bit off the deep end. Image by Photo by Sweet Ice Cream Photography. Still here? Great! Read on: Step 1: Find a story worth telling. When creating products for my soap and candle business, I always began with a story/experience that was at once universal and individual (within my target market). In the case of the Sweet Orange + Clove Candle, I chose to create a scented candle to be used in the ceremony of Havdalah by which observant Jews, every Saturday night, simultaneously bid farewell to their weekly day of rest, and greet the coming week. The Havdalah Ceremony involves a series of sensory blessings: on a cup of wine, a scent, and a burning flame made of symbolically intersecting wicks. I wanted the scented candle to functionally qualify to be used in the ceremony (by having intersecting wicks) and support the overall experience through story-specific scent. Step 2: Find the scents present in the story. My own memories of Havdalah included the scents of wine/grape juice, but also those of a smokey extinguished candle, and fresh rosemary. In some streams of the Jewish tradition, fresh herbs are considered ideal for making the ‘scent’ blessing of the ceremony, and I’d been to many homes throughout Israel where the hosts picked pungent rosemary from their gardens minutes before the blessings were made over them at the Havdalah ceremony. I also had the memory of dried orange peels and dried Etrog (Citron) which I’d seen used for the ‘scent’ blessing. Then there were memories of cinnamon sticks, some weak, some potent, as well as sachets of prickly cloves and rounded allspice. I compiled a list of my memories in their scent form: wine, smoke, rosemary, sweet orange, citron, cinnamon, clove, and allspice. Step 3: Search for the memory flood. This part is the hardest to articulate as it involves synesthesia, defined as a perceptual phenomenon in which stimulation of one sensory or cognitive pathway leads to automatic, involuntary experiences in a second sensory or cognitive pathway. I assembled the following tools- a well-calibrated digital scale a beaker the 8 fragrance notes (wine, smoke, rosemary, sweet orange, citron, cinnamon, clove, and allspice) pipettes a recipe notebook and pencil to record weight ratios and degrees of memory recall side-by-side a jar of ground coffee to clear the olfactory palette between blends, not unlike the role of ginger in the sushi experience a relaxed and open state of mind in which to allow a flood of memories, good and bad, expected and unexpected, to surface Then I blended, recorded the ratios, tested the scent, and recorded the level of memory-flood on a scale of 1–10. And reblended, and repeated. Ratio is everything and it’s never what you’d expect. Here’s why: Although the Havdalah subject has the elements I’ve listed, they are not present in equal proportion. A cake is not baked with equal parts flour, sugar, salt, eggs, oil, baking soda, and water. Similarly, my sensory experiences of the Havdalah ceremony were not baked with equal parts rosemary, smoke, orange, etc. I have one criteria for the scent blend to test its accuracy (which some would call its marketability). And that is- Does this cause a flood of memories without manipulation? With a clear and open mind, am I recalling more details of my experiences while I smell this blend? Are these memories being imposed by my cognitive mind, or evoked through my open heart in response to this blend? When to stop blending: I look for physical signs- if I am moved to tears by the scent, that is a good sign that the fragrance is a true translation of the experience. This is because I generally cry only in response to truth and beauty. I also notice that my hands become very cold when the accurate blend comes into focus. So I stop (not as easy as it sounds), write down the ratios in the final blend, and put away my tools in a state of awe and gratitude for the birth of this fragrance, which has not existed in the world before this moment. Step 4: Test in the Open-Air Market This autobiographical approach to product creation begs the question- Aren’t we all different? Don’t we lead different lives? Who is to say that what speaks to you, will speak to another? That was the most exciting part of the product process- finding out the degree to which my resonance resonated with others. In order to find that out, I created a vessel for experiencing the scent- the wooden-wick candle. Then it needed a name- the name of a scent frames the experience of it. The earliest incarnation of “Sweet Orange & Clove”. Naming the scent I learned that the name of the scent, like the name of a restaurant salad, should be as straightforward as possible (think ‘Cucumber Melon’ instead of ‘High Water Content Bliss’), but leave out unnecessary details which would clutter the concept. Customers must be willing to pick up the tester, and they’re more likely to take a chance on a tester called ‘Sweet Orange + Clove’ than a tester called ‘Orange, Smoke, Clove + Citron’, even if all four notes are present in various degrees. Sweet Orange + Clove — The Verdict Scent sells. As a formulator, I knew that customers decided whether or not to buy my products by actively compared the scent I’d created, which was the creative interpretation of a story, with their OWN stories, through their olfactory recall. I learned that the more I solved for an unbiased memory-flood in my design process, the more they experienced the same from the product. They’d take home the scents that were already part of the best of their life experiences, and leave the ones which were less familiar or which they didn’t want to reexperience. The day I brought our case study of ‘Sweet Orange + Clove’ to the open-air market, I saw shoppers lift up and smell the test candle, gasp, and immediately share newly-recovered memories with their shopping partners which had flooded forth from their memory cache. When this happened consistently, until the test batch sold out and a customer was asking me if she could just buy the tester, I knew I’d succeeded in translating a story to scent. ___________ Shoshana Rubli is an entrepreneur and blockchain enthusiast who enjoys green smoothies and going #meta. This story is published in The Startup, Medium’s largest entrepreneurship publication followed by 318,120+ people. Subscribe to receive our top stories here.

Harnessing Olfactory Memory in Product Design

133

taking-smellfies-olfactory-memory-in-product-design-1e499d1dfadb

2018-05-06

2018-05-06 06:08:50

https://medium.com/s/story/taking-smellfies-olfactory-memory-in-product-design-1e499d1dfadb

false

1,359

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_

Product Management

product-management

Product Management

25,668

Shoshana Rubli

Millennial Expat + Entrepreneur | US IL CH | Contributor @TheStartup @TheEvolution @4rev | shoshanarubli@gmail.com

25f9d76675df

shoshanarubli

275

128

20,181,104

null

0

null

0

700ca364d424

2018-06-26

2018-06-26 12:37:30

2018-07-11

2018-07-11 08:55:17

2

false

en

2018-07-11

2018-07-11 09:09:05

4

1e49c57fe5e3

4.824843

8

0

Where is the divide between reality and representation?

5

Virtual realities, the simulacrum and electric sheep Where is the divide between reality and representation? As Gilles Deleuze notes in his 1983 work Plato and the Simulacrum, the power of the simulation is not in its opposition or otherness, but in its almost imperceptible similarity to reality. A radical French thinker, Deleuze’s philosophy centres upon concepts of multiplicity and difference. Writing with partner and psychoanalyst Felix Guattari, Deleuze championed a rhizomatic approach to knowledge and understanding; allowing for multiple, non hierarchal ideas to coexist and feed into one another. His writing rejects a dualistic or binary understanding of the world, claiming that the “rhizome has no beginning or end; it is always in the middle, between things, interbeing, intermezzo”— it is perhaps this ambiguity of origin which endows much of popular culture’s rendering of AI, VR and the cyborg body with an innate sense of dystopia. Rhizome, or an “image of thought”, is a term Deleuze and Guattari use to describe theory and research that allows for multiple, non-hierarchical entry and exit points in data representation and interpretation. It presents history and culture as a wide array of attractions and influences with no specific origin. Ridley Scott’s Blade Runner (1982) follows the story of Rick Deckard, a bounty hunter, or blade runner, hired to retire (a dehumanised translation for kill) replicant cyborg beings. These cyborgs are designed to be “more human than human”, albeit lacking in emotional depth and sense of empathy, and in theory distinct from humans. However, when he encounters an experimental replicant named Rachael — who has been implanted with false memories, and thus believes herself to be human — the border between human and machine becomes hazy. The film never clarifies whether Deckard is himself a replicant, an ambivalence which problematises this grey area between appearances and truth. Take Philip K. Dick’s Do Androids Dream of Electric Sheep? (1968), a loose basis for Scott’s film; the opening of the novel documents a conversation between Rick Deckard and his wife Iran surrounding their Penfield mood organ, a device used to regulate the emotions of users. Iran tells Rick that her schedule lists “a six hour self-accusatory depression,” much to her husband’s despair. She informs him that she was tinkering with the machine one day when she stumbled across the setting for despair, and has scheduled it for twice a month — “I think that’s a reasonable amount of time to feel hopeless about everything… don’t you think?” There is almost a tongue in cheek approach to these slippages between reality and simulation — moments of dark humour can be found throughout the text, in which a point is made of such uncanny moments in which our characters break a fourth wall, and discuss a collision between reality and its simulation. What draws my attention is an awareness of the constructed nature around them, and how the characters choose to inhabit the world in light of this knowledge. Deckard yearns for a real animal to replace his mechanical sheep, rather than tending to a “fake” he finds demoralising; he discloses to his neighbour of how he used to own a real sheep which died suddenly, and had a replica of him made without arousing suspicion; not the same, but “almost”. I have come to think of the novel’s title as an extension into this territory of the “almost,” a world in which androids count electric sheep to fall asleep, almost perfectly emulating human characteristics but slightly off kilter — a technology which parodies and simulates. We’re told that we must fear the ability of these technologies to replicate, and to position them as the uncanny doppelgänger. A positioning of the virtual as a false image (distorted to the point of seeming real), attempts to strip the copy or the simulacrum of its power for productivity. But why are we conditioned to view that which is, arguably, made in our image with such hesitation and trepidation? §The artist Terence Broad engages with such themes of replication, simulation and representation within his work Autoencoded (2016), a technologically manufactured replication of Scott’s Blade Runner. Broad writes extensively about his thinking and practice, sighting the mind/body split of philosopher René Decartes as a focal point of his artistic motivation. Descartes assertion of ‘cogito ergo sum’/‘I think therefore I am’, hoped to reconcile a unification between a thinking mind and existence. Using an autoencoder, Broad constructed the film by teaching a neural network to distinguish between data fed to it, to memorise and then relay from memory a string of images which present a hazy and dream like impression of Scott’s work. This method of feeding the network compressed data, and allowing it to reconstruct what it has “seen” perhaps attempts to endow the technology with a sense of agency. The artist presents the ability for such a virtual system which ‘perceives images but is not embodied within the environment that the images represent’ to exacerbate the assertion of the thinking mind as the root of reality. The film picks up on a disparity between the human and its representations, and both Broad’s work and the materialisation of A.I. represented within sci-fi such as Blade Runner become part of a process of humanising technology and giving rise to the possibility for a subjective machine. This draws an interesting parallel with Warner Brother’s response to the piece, issuing a DMCA takedown notice for Broad’s work on Vimeo, on the grounds of copyright infringement —in other words: ‘Warner had just DMCA’d an artificial reconstruction of a film about artificial intelligence being indistinguishable from humans, because it couldn’t distinguish between the simulation and the real thing’. A key aspect of Scott’s film, at least philosophically, is the attempt to distinguish between the representations of the real and “reality” itself — and whether this is can ever be clear cut. The importance of distinguishing between real and fake, origin (primary) and offspring (secondary/copy), original and implanted resonate not only throughout science fiction dedicated to the future of humanity and cyborg bodies, but more broadly to discussions of the nature of the human and of ontology. Creating machines with a sense of human instability and irrationality is potentially the most challenging and yet exciting aspect of the android and of A.I. Before taking his own life in November 1995, Deleuze penned an essay entitled The Actual and The Virtual — defining the virtual as a kind of potentiality, and of continuous multiplicities, we can begin to think of the virtual space perhaps not as an extension of the physical world, but as a arena for potentiality, and a space to work through, to confront and to challenge. And maybe the fear of the doppelgänger or the android stems from precisely this — that it forces us to confront ourselves, to rethink that which we believe we fully grasp, and the basis of all our understanding. Perhaps the virtual challenges our grasp upon our own perception. Do our attempts to control simulation and mimicry enable a potentially more objective view of our domain?

Virtual realities, the simulacrum and electric sheep

352

virtual-realities-the-simulacrum-and-electric-sheep-1e49c57fe5e3

2018-07-11

2018-07-11 09:09:05

https://medium.com/s/story/virtual-realities-the-simulacrum-and-electric-sheep-1e49c57fe5e3

false

1,177

We know there should be no boundaries to the reach of human connection.

null

From The Edge

hello@ori.co

from-the-edge

5G,EDGE COMPUTING,MESH NETWORKS,DIGITAL TRANSFORMATION

ori_edge

Science Fiction

science-fiction

Science Fiction

15,600

Gemma Murray

Feminist philosopher & art theorist

e3a3ea1829d6

gemma_murray_

15

19

20,181,104

null

0

null

0

c090788950fb

2018-01-25

2018-01-25 04:23:10

2018-01-25

2018-01-25 04:24:44

1

false

en

2018-01-25

2018-01-25 04:26:21

6

1e4b5fd72bb1

3.550943

0

Written by Guest Blogger: Dana Hauk

5

4 Ways Machine Learning Is Changing the Marketing Game Written by Guest Blogger: Dana Hauk Marketing trends come and go, but when something truly innovative comes around, it doesn’t just pass as a fad; it incorporates itself into the industry for good. For the past few years, that thing appears to be the use of artificial intelligence, specifically machine learning, in marketing. When used correctly, machine learning can take your marketing to the next level and exponentially improve your customer experience. Many companies are just starting to take an interest in AI, as it trickles down from being used only by the tech giants who can afford it, like Google, and into every company’s grasp. However, because AI is such a vast and complex field, it can be difficult for a marketing professional to decipher exactly what components of AI are catered specifically to the marketing industry, and which ones are left for other purposes. This blog post outlines the four main ways that AI applies to the world of marketing, and will help you inform yourself on what you need to know to get your company started on the trend. Highly targeted ads The most important way that AI is entering the field of marketing is with its ability to target customers with impressive accuracy. In this case study, AI technology helped Delta Faucet increased page views per user of their online content by 49% using natural language processing algorithms. The algorithms predicted which content would be the most popular and then determined who the content should be targeted to. The use of machine learning was the foundation of a game-changing marketing campaign for Delta that focused on intense personalization of ads to bring in more new users to their website and keep them there. This kind of targeted content success is typical with AI, because it automates the process of segmenting your content and your customers to make sure that the right content gets created in the first place, and that it gets targeted to the right people at the right time on the right channel. The amount of manpower it would take to do this kind of targeting is an unrealistic load for most marketing teams to undertake-but with AI, it suddenly becomes possible. Recommender systems True or False: When a friend recommends a book to you that you fall in love with, or a restaurant that you enjoy, it endears you to the friend and makes you feel comforted and understood. If you answered true, then you understand the genius method behind the growing use of recommender systems by companies like Netflix, and Audible. Recommender systems use machine learning to analyze a customer’s past preferences so that they can predict future preferences on entertainment, products, or services. By Netflix allowing you to set up different profiles (When they ask, “Who’s watching?”) they are able to personalize these recommender systems to your specific viewing preferences, leaving out your spouse or child’s viewing history that would skew the data. Recommender systems allow a company to essentially read a customer’s mind by figuring out what they like and what they are most likely to do. By having a system recommend things to you, they are improving your customer experience and increasing your loyalty to their brand. Sentiment Analysis Sentiment analysis is the machine learning technique that determines whether a piece of writing is positive, negative or neutral. By doing this, bots can sift through customer feedback in order to categorize feedback into positive or negative, and can even judge the “mood” of an email so as to help the company properly respond to it. Sentiment analysis is also used to read comments on social media in order to determine a company’s online reputation. A bot that is measuring sentiment analysis will be able to inform your company if people are complaining about your brand or praising it, allowing you to target detractors and solve their problems, or to find your promoters to turn them into brand ambassadors. The ability for a machine to read emotion has endless possibilities for your marketing efforts and to help you improve your customer’s all around experience with your company. Chatbots One of the most popular ways that brands are taking advantage of artificial intelligence is the use of chatbots. Chatbots get “smart” by integrating with social media, allowing them to read your basic info, as well as your mood and past activity. In this way, chatbots take advantage of the three AI techniques listed above: targeting, recommending, and analyzing sentiment; in order to communicate with the customer in a highly personalized and coherent way. Chatbots are a great tool for keeping customers on your page, improving wait times for customer service inquiries, and keeping your customer experience fresh and fun. Artificial intelligence is offering marketing departments the opportunity to personalize and automate their marketing like never before. By tailoring your marketing to the individual customer experience, you can set yourself apart from the competition and win loyal customers for life. — Want to be more effective with the time you have creating marketing strategies? You can read The 5 Step Process For a More Structured Marketing Strategy eBook for a more in-depth discussion of these concepts and how you can begin to implement them. Originally published at www.thoughtsofabrandstrategist.com.

4 Ways Machine Learning Is Changing the Marketing Game

0

4-ways-machine-learning-is-changing-the-marketing-game-1e4b5fd72bb1

2018-01-25

2018-01-25 09:21:36

https://medium.com/s/story/4-ways-machine-learning-is-changing-the-marketing-game-1e4b5fd72bb1

false

888

A project of Multiple Atlanta: a strategic design firm specializing in brand and digital marketing strategy, consulting, and execution.

null

Thoughts of a Brand Strategist

atlanta@multipleinc.com

thoughts-of-a-brand-strategist

DIGITAL MARKETING,BRAND STRATEGY,GRAPHIC DESIGN,INBOUND MARKETING,BRANDING

skotwaldron

Machine Learning

machine-learning

Machine Learning

51,320

Skot Waldron

Principal www.multipleinc.com / Creator www.ThoughtsofaBrandStrategist.com / Brand Coach / Digital Marketing Consultant / Speaker

ed2472ae5e55

skotwaldron

114

138

20,181,104

null

0

null

0

null

2018-02-24

2018-02-24 00:03:17

2018-02-23

2018-02-23 23:56:55

2

false

en

2018-02-24

2018-02-24 00:05:33

21

1e4b663c481e

2.590881

1

0

Enjoy your weekend and save this for a rainy d — OK, it’s meant to rain all weekend, but just save this for Monday OK?

4

#TwitterLockOut, Voice Search and more in Do Not Read Until Monday Enjoy your weekend and save this for a rainy d — OK, it’s meant to rain all weekend, but just save this for Monday OK? Hey, we got a new logo! Twitter ruffles feathers following bot purge Twitter — either the ultimate achievement of having valid opinions voiced from all social strata … or an experiment to do so gone horribly, horribly wrong — made a few moves to purge the bot presence on its platform this week. Dramatization. The most talked-about action was the so-called #TwitterLockOut, where Twitter required certain suspect accounts to verify a phone number to log in. Totally not something to freak out about right? Haha, I kid — this is Twitter, of course people freaked out. Far-right personalities quickly decried it as a strike on free speech, before Twitter clarified its intent. Then the commotion mostly died down. Twitter likely had the crackdown planned previously, but it did come closely after bot propaganda about the Florida school shooting flooded the platform. In other bot-battling news, Twitter will be barring ‘bulk posting’ — posting the same or similar content across multiple accounts — imminently. That could have effects for brands using multiple accounts — so watch out! And for customer-service focused brands, it also made a tweak to DM rules to make it a bit friendlier. Voice search is going to be huuuuuge You don’t need to be best pals with Siri, Alexa, Cortana or — Assistant? WTF Google, you can’t even give it real a name? — to know voice search is big. Gartner even stated this week that voice search should take up about 30 percent of overall searches by 2020. Shortly following, it leaked that Apple wants to get Siri all up in your earholes with an upgraded AirPod — simply say “Hey Siri” and she will answer you with some reply irrelevant to your inquiry. Well, at least they’ll be waterproof. More light-rain than help-I’m-drowning, but still. Google also updated its, ahem, Assistant, integrating multilingual speech recognition, as well as debuting Routines, which help users execute multiple tasks from just one command. This shift has big implications for both organic and paid search — where content optimizations will need to modernize and impressions should likely decline, but also opens up the possibility of new ads platforms. Odds + The End More people-less Amazon stores? This is fine, so long as they’re not called Bodega. The net neutrality repeal now has a set date. Someone launched a fake news game. I mean, it seems someone already won this game but whatever. If you want to know what milking it looks like, McDonalds is bringing Szechuan sauce back, again. Uber takes on — the bus, I guess? Need to borrow some loot? Ask Evan Spiegel — he got a sick $637M bonus last year brah. His company, Snap, also delved into ecommerce as a new revenue stream. Hey I mean they did pretty good at selling other stuff like glasses, right? I guess I should mention Kylie Jenner, but the Maybelline move was way more savage. Facebook has pulled it’s seemingly annual “but I can change” routine with advertisers, promising to update questionable metrics. Dang, some people are really serious about lobsters. Feel good hits of the week: Snap Maps shows the scale of high-school walkouts, and people are using the popularity ofBlack Panther to register voters. Originally published at mry.com on February 23, 2018.

#TwitterLockOut, Voice Search and more in Do Not Read Until Monday

1

twitterlockout-voice-search-and-more-in-do-not-read-until-monday-1e4b663c481e

2018-02-27

2018-02-27 15:23:09

https://medium.com/s/story/twitterlockout-voice-search-and-more-in-do-not-read-until-monday-1e4b663c481e

false

585

null

Artificial Intelligence

artificial-intelligence

Artificial Intelligence

66,154

MRY

Founded in 2002, MRY is a global digital marketing agency offering creative strategies to reach elusive audiences. Visit our blog: https://mry.com/blog

84760fbbae42

MRY

3,440

820

20,181,104

null

0

null

0

null

2018-02-06

2018-02-06 09:30:21

2018-02-07

2018-02-07 08:11:24

1

false

en

2018-02-07

2018-02-07 08:11:24

1

1e4b97682ec

3.943396

0

null

3

Four key tips for getting digital transformation right Digital transformation is vitally important for businesses, as it allows them to keep their competitive edge in the era of disruption It is the digital age of engagement, interaction and collaboration. Older IT architecture is typically outdated when it comes to security, making it more exposed to potential breaches. With this in mind, businesses need to focus their attention towards digital transformation and restructure their components with digital technologies. This will not only safeguard against the mounding security threats but will also ensure that businesses can streamline processes and ultimately remain competitive. Digital transformation is a visible whole scale restructuring of business processes to avoid a tipping point caused by digital technologies and downstream market efforts. It can often take a significant amount of time to entirely restructure a business and can use a huge amount of resources, especially for large organisations. Keeping a competitive edge Businesses of all industries and sizes can benefit from undergoing digital transformation. One key benefit is that it allows businesses to optimise their end-user engagement and experience through providing them with the best systems and tools. From dwindling budgets to time restraints, there are various reasons why businesses are yet to embrace and embark on digital transformation. On the whole, there is currently a lack of organisation wide information and collaborative approach. Each individual division within an organisation needs to ensure that they are working together through the process of digital transformation in order to efficiently upgrade their systems. Digital transformation can be a time-consuming process, which can also be costly and require a significant amount of resources. It is important that business correctly implement the process in order to optimise their resources. In order to ensure that businesses can get digital transformation right, they should bear in mind the following four key points. 1. Automation is key In the last few years, there has been a significant rise in processes that simplify manual tasks. Artificial Intelligence (AI) and Machine Learning have already made a huge impact and technology such as chatbots are enabling AI to integrate to everyday life. Companies across all industry sectors are now able to use AI and machine learning to utilise the data available in a business to effectively complete tasks. Before embarking of business transformation businesses must consider the capabilities of AI and Machine Learning and evaluate the tasks that can be ultimately automated and ensure that these are at the forefront of their mind during the entire process. Businesses must ensure that the vision has to be the guiding force behind each decision during transformation. Automating manual tasks ensures that day to day processes are streamlined, saving businesses time and money and ultimately giving staff more time to focus on other tasks which will help to drive the business forward. 2. Enable innovative working The workforce is the most valuable asset of any business, and it is vital to ensure that employees are prioritised during the process of digital transformation. Technological progress offers the chance to introduce a new, innovative type of working. Working hours are becoming more flexible and employees are increasingly able to organise their working day themselves. Besides interpersonal relationships with colleagues and superiors, flexible working times are the most important factor for employee motivation. They not only want to decide when they want to do their work but also from where. Flexible working is now easier than ever due to a huge increase in innovative solutions, which enable staff to work from anywhere. There are now readily available collaborative tools which help employees who are geographically separated, to work together virtually and be more productive without compromising on their work-life balance. These technologies can grant employees more freedom by letting them organise their own working hours. This requires a relationship of trust between staff and employers. Since employees are only sporadically present at work, their companies experience a loss of control over them. In addition, a technical foundation has to be created to be able to introduce and offer modern modes of working. 3. Data led decision making Along with allowing automation and flexible working, digital transformation also allows data led decision making. Data is extremely valuable to every business and it is crucial that organisations wake up to this and utilise the data at their fingertips in order to monitor their customer’s behavioural patterns Digital transformation allows businesses to make real-time decisions based on updated data. One example of how this can be beneficial is for sales teams. They are able to view any changes in the buying behaviour and make quick strategic decisions to accommodate for such changes. Businesses can change the pace of their strategy in line with the behaviour of their customers. 4. Enhancing customer needs Businesses should also acknowledge that digital transformation is not just about a technology refresh and updating typically outdated systems but it is also about the human aspects and an improvement on workplace culture. However, there can often be negative backlashes from the digital transformation from employees, who do not see the immediate benefit and are concerned about getting to grips with the new systems. Therefore, it is important that customers and employees are kept at the centre of any decisions that are made. However, once employees see how upgraded infrastructure can ease their daily tasks and free up their time, employees will quickly realise the value that digital transformation has. It is important that organisations begin to focus their efforts towards digital transformation and move away from out-dated processes and technology that are typically more prone to security risks. By following these four tips, businesses can ensure that they correctly implement the process of digital transformation and through carrying out the process in the best possible way, they can avoid an unnecessary drain on their time and resources. Also — have a look at the seven steps to address when going digital

Four key tips for getting digital transformation right

0

four-key-tips-for-getting-digital-transformation-right-1e4b97682ec

2018-02-07

2018-02-07 08:11:25

https://medium.com/s/story/four-key-tips-for-getting-digital-transformation-right-1e4b97682ec

false

992

null

Sap

sap

Sap

3,303

Neil How

Founder of Limelight Solutions, an IT consultancy specialising in the implementation of SAP under accelerated timelines. www.limelightsolutions.co.uk

a30625d82dac

neilhow

8

9

20,181,104

null

0

# Represent sentence with word index, using word index to represent a sentence x = data_helpers.build_index_sentence(x_text, vocabulary) y = y.argmax(axis=1) # y: [1, 1, 1, ...., 0, 0, 0]. 1 for positive, 0 for negative # Output output_layer = Dense(1, activation='softmax', name='softmax_output')(output_layer) output_layer = Dense(1, activation='sigmoid')(output_layer) # Represent sentence with word index, using word index to represent a sentence x = data_helpers.build_index_sentence(x_text, vocabulary) # y = y.argmax(axis=1) # comment this to make the dimension of y is 2 output_layer = Dense(2, activation='softmax')(output_layer) # change 1 to 2 as the output neuron

5

null

2018-07-03

2018-07-03 06:47:21

2018-07-03

2018-07-03 07:16:30

5

false

en

2018-07-03

2018-07-03 07:16:30

1

1e4bf91a2ffe

1.833333

3

0

If we use softmax as the activation function to do a binary classification, we should pay attention to the number of neuron in output…

3

Softmax output neurons number for Binary Classification? If we use softmax as the activation function to do a binary classification, we should pay attention to the number of neuron in output layer. you can find the detail implementation with Keras in this notebook One output neuron y should be 1 dimension In the model part, we set the output neuron number as 1. If we train this model, the loss will be big and hard to converge. The reason is that softmax will assign probability for each class, and the total sum of the probabilities over all classes equals to one. And the number of neuron in output layer is only 1, this will cause the output of softamx will all become 1. the outputs of softamx are all 1 But in this case, if we set activation function as sigmoid, the loss will become small and converge. Two output neuron The solution is pretty simply, we set y as two dimension, and set the number of output neuron as 2. Now the loss is small and can be decreased normally.

Softmax output neurons number for Binary Classification?

3

softmax-output-neurons-number-for-binary-classification-1e4bf91a2ffe

2018-07-03

2018-07-03 07:16:30

https://medium.com/s/story/softmax-output-neurons-number-for-binary-classification-1e4bf91a2ffe

false

265

null

Machine Learning

machine-learning

Machine Learning

51,320

Xu LIANG

I KNOW NOTHING!

ee86e6752cb4

zhuixiyou

10

42

20,181,104

null

0

null

0

8e54d041f21b

2018-08-29

2018-08-29 13:37:58

2018-08-29

2018-08-29 19:22:04

1

false

en

2018-08-29

2018-08-29 19:23:49

8

1e4c5b68f38

2.879245

6

0

1

The Future of AR/VR, How to Apply AI in Your Business, and Startups to Watch Out For

5

The Top 3 Sessions at TakeOver Innovation Conference 2018 The Future of AR/VR, How to Apply AI in Your Business, and Startups to Watch Out For By The TakeOver Team Augmented Reality and Virtual Reality: Snap Back to Reality From a 14-year-old founder of an AR/VR company to a film and television strategist, to an author of a book that blurs the lines between the physical and digital worlds, this panel explored the current and future landscape of AR/VR technology. The panelists, Sabarish Gnanamoorthy (Founder of Waypoint AR), Ted Biggs (VP of Convergent and Technology at Shaftesbury), and Dr. Helen Papagiannis (Author of Augmented Human), discussed and debated mass consumption of AR/VR, trends, and how the tech can make a greater impact. We heard about some of the cutting-edge applications of AR/VR, for example: the technology is used to calm medical patients by 30%, and in just 8 minutes! Gnanamoorthy also challenged the panelists with the ethics of data collection. He explained that devices can collect and store what a user sees and feels in an AR/VR experience, so we have to be cautious and mindful of how and why data is collected. In short, it should only be used to enhance the user experience while maintaining privacy. This is just a taste of what the audience learned and uncovered; watch the panel to learn more about AR and VR, and how it’s shaping the future. Watch the full session! “I’d say, if not all of you, but most of you have the power of AR in your pocket” — Helen Papagiannis Element AI Workshop It’s not everyday that you see CEOs and top-tier execs cramming into a room and squeezing to grab a seat on the floor, but this was the scene at the Element AI Workshop. Karthik Ramakrishnan, the Vice President and Head of Industry Solutions at Element AI, took participants through an interactive workshop. He helped them uncover the value and need for an Artificial Intelligence (AI)-based solution to solve one of their own distinctive business problems. Karthik broke down some of the fuzziness surrounding AI, shared his business problem matrix, framing tactics, and took the audience through his tried-and-true process to identifying an AI-based solution. Watch and find out how you can identify a pressing business problem, uncover an AI solution, and turn it into something scalable. Get your organization started with AI! Venture Studios Showcase TribalScale is pushing the boundaries of how and what it means to innovate. In February, we launched Venture Studios, which leverages our teams, processes, and tools to create, source, and scale disruptive startups in partnership with corporates. Just 14 weeks later at the TakeOver Innovation Conference, 4 disruptive startups were showcased. The founders and leaders of the studios startups, Troüpe, MESH, Senso.ai, and Every, took the stage, introduced their companies, and told the audience why and how they’re going to make a BIG impact on the automotive and financial industries. The audience also heard from all-star innovators: Roger Chabra (CIO and Head of Venture Studios), Peter Aceto (former CEO of Tangerine), Farid Kassam (Principal of Akeelee, who also helped build the CANADARM1), Preeti Malik (former National Lead of Risk and Compliance at Accenture), and Matt Crossley (former Director of Engineering and Electrification at GM China). Spoiler alert: we might think we know how to innovate, but we’re wrong. Watch and learn how our Venture Studios startups are disrupting the auto and finance space! About TribalScale’s TakeOver Innovation Conference TakeOver Innovation Conference is Canada’s largest one-day innovation conference and attracts over 1,200 business executives and thought leaders from around the world. TakeOver delegates come together from a range of industries to discuss current and future trends that will enable individuals and companies to grow and thrive. Through hands-on learning, networking opportunities, and exposure to emerging technologies, participants will discover how to transform and innovate for long-term success. With truly diverse and inclusive aims, brand new panel sessions, interactive workshops and demos, TribalScale’s TakeOver Conference is the ultimate place for innovation. For more information on TakeOver please visit https://takeoverinnovationconference.com/ Join our fast growing team and connect with us on Twitter, LinkedIn & Facebook! Learn more about us on our website.

The Top 3 Sessions at TakeOver Innovation Conference 2018

143

the-top-3-sessions-at-takeover-innovation-conference-2018-1e4c5b68f38

2018-08-29

2018-08-29 19:23:49

https://medium.com/s/story/the-top-3-sessions-at-takeover-innovation-conference-2018-1e4c5b68f38

false

710

TribalScale is a global innovation partner that helps enterprises adapt and thrive in the digital era. We do this by transforming teams and processes, building best-in-class digital products, and co-creating disruptive startups.

null

tribalscale

null

TribalScale

info@tribalscale.com

tribalscale

TECH,INNOVATION,EMERGING TECHNOLOGY,MOBILE,WEB

tribalscale

Virtual Reality

virtual-reality

Virtual Reality

30,193

TribalScale Inc.

Right the future.

3d4f4d2f3e6

tribalscale

1,524

27

20,181,104

null