How-To Tutorials - Artificial Intelligence

Last week, the House Intelligence Committee held a hearing to examine the public risks posed by “deepfake” videos. Deepfake is identified as a technology that alters audio or video and then is passed off as true or original content. In this hearing, experts on AI and digital policy highlighted to the committee, deepfakes risk to national security, upcoming elections, public trust and the mission of journalism. They also offered potential recommendations on what Congress could do to combat deepfakes and misinformation. The chair of the committee Adam B. Schiff, initiated the hearing by stating that it is time to regulate the technology of deepfake videos as it is enabling sinister forms of deception and disinformation by malicious actors. He adds that “Advances in AI or machine learning have led to the emergence of advance digitally doctored type of media, the so-called deepfakes that enable malicious actors to foment chaos, division or crisis and have the capacity to disrupt entire campaigns including that for the Presidency.” For a quick glance, here’s a TL;DR: Jack Clerk believes that governments should be in the business of measuring and assessing deepfake threats by looking directly at the scientific literature and developing a base knowledge of it. David Doermann suggests that tools and processes which can identify fake content should be made available in the hands of individuals, rather than relying completely on the government or on social media platforms to police content. Danielle Citron warns that the phenomenon of deepfake is going to be increasingly felt by women and minorities and for people from marginalized communities. Clint Watts provides a list of recommendations which should be implemented to prohibit U.S. officials, elected representatives and agencies from creating and distributing false and manipulated content. A unified standard should be followed by all social media platforms. Also they should be pressurized to have a 10-15 seconds delay in all videos, so that they can decide, to label a particular video or not. Regarding 2020 Presidential election: State governments and social media companies should be ready with a response plan, if a fake video surfaces to cause disrupt. It was also recommended that the algorithms to make deepfakes should be open sourced. Laws should be altered, and strict actions should be awarded, to discourage deepfake videos. Being forewarned is forearmed in case of deepfake technology Jack Clerk, OpenAI Policy Director, highlighted in his testimony that he does not think A.I. is the cause of any disruption, but actually is an “accelerant to an issue which has been with us for some time.'' He adds that computer software aligned with A.I. technology has become significantly cheaper and more powerful, due to its increased accessibility. This has led to its usage in audio or video editing, which was previously very difficult. Similar technologies  are being used for production of synthetic media. Also deepfakes are being used in valuable scientific research. Clerk suggests that interventions should be made to avoid its misuse. He believes that “it may be possible for large-scale technology platforms to try and develop and share tools for the detection of malicious synthetic media at both the individual account level and the platform level. We can also increase funding.” He strongly believes that governments should be in the business of measuring and assessing these threats by looking directly at the scientific literature and developing a base knowledge. Clerk concludes saying that “being forewarned is forearmed here.” Make Deepfake detector tools readily availaible David Doermann, the former Project Manager at the Defense Advanced Research Projects Agency mentions that the phrase ‘seeing is believing’ is no longer true. He states that there is nothing fundamentally wrong or evil about the technology, like basic image and video desktop editors, deepfakes is only a tool. There are a lot of positive applications of generative networks just as there are negative ones. He adds that, as of today, there are some solutions that can identify deepfakes reliably. However, Doermann fears that it’s only a matter of time before the current detection capabilities will be rendered less effective. He adds that “it's likely to get much worse before it gets much better.” Doermann suggests that tools and processes which can identify such fake content should be made available in the hands of individuals, rather than relying completely on the government or on social media platforms to police content. At the same time, there should also be ways to verify it or prove it or easily report it. He also hopes that automated detection tools will be developed, in the future, which will help in filtering and detection at the front end of the distribution pipeline. He also adds that “appropriate warning labels should be provided, which suggests that this is not real or not authentic, or not what it's purported to be. This would be independent of whether this is done and the decisions are made, by humans, machines or a combination.” Groups most vulnerable to Deepfake attacks Women and minorities Danielle Citron, a Law Professor at the University of Maryland, describes Deepfake as “particularly troubling when they're provocative and destructive.” She adds that, we as humans, tend to believe what our eyes and ears are telling us and also tend to share information that confirms our biases. It’s particularly true when that information is novel and negative, so the more salacious, we're more willing to pass it on. She also specifies that the deepfakes on social media networks are ad-driven. When all of this is put together, it turns out that the more provocative the deepfake is, the salacious will be the spread virally.  She also informed the panel committee about an incident, involving an investigative journalist in India, who had her posters circulated over the internet and deepfake sex videos, with her face morphed into pornography, over a provocative article. Citron thus states that “the economic and the social and psychological harm is profound”. Also based on her work in cyber stalking, she believes that this phenomenon is going to be increasingly felt by women and minorities and for people from marginalized communities. She also shared other examples explaining the effect of deepfake on trades and businesses. Citron also highlighted that “We need a combination of law, markets and really societal resilience to get through this, but the law has a modest role to play.” She also mentioned that though there are laws to sue for defamation, intentional infliction of emotional distress, privacy torture, these procedures are quite expensive. She adds that criminal law offers very less opportunity for the public to push criminals to the next level. National security Clint Watts, a Senior Fellow at the Foreign Policy Research Institute provided insight into how such technologies can affect national security. He says that “A.I. provides purveyors of disinformation to identify psychological vulnerabilities and to create modified content digital forgeries advancing false narratives against Americans and American interests.” Watts suspects that Russia, “being an enduring purveyor of disinformation is and will continue to pursue the acquisition of synthetic media capability, and employ the output against adversaries around the world.” He also adds that China, being the U.S. rival, will join Russia “to get vast amounts of information stolen from the U.S. The country has already shown a propensity to employ synthetic media in broadcast journalism. They'll likely use it as part of disinformation campaigns to discredit foreign detractors, incite fear inside western-style democracy and then, distort the reality of audiences and the audiences of America's allies.” He also mentions that deepfake proliferation can present a danger to American constituency by demoralizing it. Watts suspects that the U.S. diplomats and military personnel deployed overseas, will be prime target for deepfake driven disinformation planted by adversaries. Watts provided a list of recommendations which should be implemented to “prohibit U.S. officials, elected representatives and agencies from creating and distributing false and manipulated content.” The U.S. government must be the sole purveyor of facts and truth to constituents, assuring the effective administration of democracy via productive policy debate from a shared basis of reality. Policy makers should work jointly with social media companies to develop standards for content and accountability. The U.S. government should partner with private sectors to implement digital verification designating a date, time and physical origination of the content. Social media companies should start labeling videos, and forward the same across all platforms. Consumers should be able to determine the source of the information and whether it's the authentic depiction of people and events. The U.S. government from a national security perspective, should maintain intelligence on capabilities of adversaries to conduct such information. The departments of defense and state should immediately develop response plans, for deepfake smear campaigns and mobilizations overseas, in an attempt to mitigate harm. Lastly he also added that public awareness of deepfakes and signatures, will assist in tamping down attempts to subvert the  U.S. democracy and incite violence. Schiff asked the witnesses, if it's “time to do away with the immunity that social media platforms enjoy”, Watts replied in the affirmative and listed suggestions in three particular areas. If social media platforms see something spiking in terms of virality, it should be put in a queue for human review, linked to fact checkers, then down rate it and don't let it into news feeds. Also make the mainstream understand what is manipulated content. Anything related to outbreaks of violence and public safety should be regulated immediately. Anything related to elected officials or public institutions, should immediately be flagged and pulled down and checked and then a context should be given to it. Co-chair of the committee, Devin Nunes asked Citron what kind of filters can be placed on these tech companies, as “it's not developed by partisan left wing like it is now, where most of the time, it's conservatives who get banned and not democrats”. Citron suggested that proactive filtering won’t be possible and hence companies should react responsibly and should be bipartisan. She added that “but rather, is this a misrepresentation in a defamatory way, right, that we would say it's a falsehood that is harmful to reputation. that's an impersonation, then we should take it down. This is the default I am imagining.” How laws could be altered according to the changing times, to discourage deepfake videos Citron says that laws could be altered, like in the case of Section 230 C. It states that “No speaker or publisher -- or no online service shall be treated as a speaker or publisher of someone else's content.” This law can be altered to “No online service that engages in reasonable content moderation practices shall be treated as a speaker or publisher of somebody else's content.” Citron believes that avoiding reasonability could lead to negligence of law. She also adds that “I've been advising Twitter and Facebook all of the time. There is meaningful reasonable practices that are emerging and have emerged in the last ten years. We already have a guide, it's not as if this is a new issue in 2019. So we can come up with reasonable practices.” Also Watts added that if any adversary from big countries like China, Iran, Russia makes a deepfake video to push the US downwards, we can trace them back if we have aggressive laws at our hand. He says it could be anything from an “arrest and extradition, if the sanction permits, response should be individually, or in terms of cyber response”, could help us to discourage deepfake. How to slow down the spread of videos One of the reasons that these types of manipulated images gain traction is because it's almost instantaneous - they can be shared around the world, shared across platforms in a few seconds. Doermann says that these social media platforms must be pressurized to have a 10-15 seconds delay, so that it can be decided whether to label a particular video or not. He adds that “We've done it for child pornography, we've done it for human trafficking, they're serious about those things. This is another area that's a little bit more in the middle, but I think they can take the same effort in these areas to do that type of triage.” This delay will allow third parties or fact checkers to decide on the authenticity of videos and label them. Citron adds that this is where labelling a particular video can help, “I think it is incredibly important and there are times in which, that's the perfect rather than second best, and we should err on the side of inclusion and label it as synthetic.” The representative of Ohio, Brad Wenstrup added that we can have internal extradition laws, which can punish somebody when “something comes from some other country, maybe even a friendly country, that defames and hurts someone here”. There should be an agreement among nations that “we'll extradite those people and they can be punished in your country for what they did to one of your citizens.” Terri Sewell, the Representative of Alabama further probed about the current scenario of detecting fake videos, to which Doermann replied that currently we have enough solutions to detect a fake video, however with a constant delay of 15-20 minutes. Deepfakes and 2020 Presidential elections Watts says that he’s concerned about deepfakes acting on the eve of election day 2020. Foreign adversaries may use a standard disinformation approach by “using an organic content that suits their narrative and inject it back.” This can escalate as more people are making deepfakes each year. He also added that “Right now I would be very worried about someone making a fake video about electoral systems being out or broken down on election day 2020.” So state governments and social media companies should be ready with a response plan in the wake of such an event. Sewell then asked the witnesses for suggestions on campaigns to political parties/candidates so that they are prepared for the possibility of deepfake content. Watts replied that the most important thing to counter fake content would be a unified standard, that all the social media industries should follow. He added that “if you're a manipulator, domestic or international, and you're making deep fakes, you're going to go to whatever platform allows you to post anything from inauthentic accounts. they go to wherever the weak point is and it spreads throughout the system.” He believes that this system would help counter extremism, disinformation and political smear campaigns. Watts added any sort of lag in responding to such videos should be avoided as “any sort of lag in terms of response allows that conspiracy to grow.” Citron also pointed out that firstly all candidates should have a clear policy about deep fakes and should commit that they won’t use them or spread them. Should the algorithms to make deepfakes be open sourced? Doermann answered that the algorithms of deepfakes have to be absolutely open sourced. He says that though this might help adversaries, but they are anyway going to learn about it. He believes this is significant as, “We need to get this type of stuff out there. We need to get it into the hands of users. There are companies out there that are starting to make these types of things.” He also states that people should be able to use this technology. The more we educate them, more the tools they learn, more the correct choices people can make. On Mark Zuckerberg’s deepfake video On being asked to comment on the decision of Mark Zuckerberg to not take down his deepfake video from his own platform, Facebook, Citron replied that Mark gave a perfect example of “satire and parody”, by not taking down the video. She added that private companies can make these kinds of choices, as they have an incredible amount of power, without any liability, “it seemed to be a conversation about the choices they make and what does that mean for society. So it was incredibly productive, I think.” Watts also opined that he likes Facebook for its consistency in terms of enforcement and that they are always trying to learn better things and implement it. He adds that he really like Facebook as its always ready to hear “from legislatures about what falls inside those parameters. The one thing that I really like is that they're doing is identifying inauthentic account creation and inauthentic content generation, they are enforcing it, they have increased the scale,and it is very very good in terms of how they have scaled it up, it’s not perfect, but it is better.”   Read More: Zuckberg just became the target of the world’s first high profile white hat deepfake op. Can Facebook come out unscathed? On the Nancy Pelosi doctored video Schiff asked the witnesses if there is any account on the number of millions of people who have watched the doctored video of Nancy Pelosi, and an account of how many of them ultimately got to know that it was not a real video. He said he’s asking this as according to psychologists, people never really forget their once constructed negative impression. Clarke replied that “Fact checks and clarifications tend not to travel nearly as far as the initial news.” He added that its becomes a very general thing as “If you care, you care about clarifications and fact checks. but if you're just enjoying media, you're enjoying media. You enjoy the experience of the media and the absolute minority doesn’t care whether it's true.” Schiff also recalled how in 2016, “some foreign actresses, particularly Russia had mimicked black lives matter to push out continent to racially divide people.” Such videos gave the impression of police violence, on people of colour. They “certainly push out videos that are enormously jarring and disruptive.” All the information revealed in the hearing was described as “scary and worrying”, by one of the representatives. The hearing was ended by Schiff, the chair of the committee, after thanking all the witnesses for their testimonies and recommendations. For more details, head over to the full Hearing on deepfake videos by the House Intelligence Committee. Worried about Deepfakes? Check out the new algorithm that manipulate talking-head videos by altering the transcripts Lawmakers introduce new Consumer privacy bill and Malicious Deep Fake Prohibition Act to support consumer privacy and battle deepfakes Machine generated videos like Deepfakes – Trick or Treat?

As algorithms are increasingly used to make decisions of social consequence, the social values encoded in these decision-making procedures are the subject of increasing study, with fairness being a chief concern. The Conference on Fairness, Accountability, and Transparency (FAT) scheduled on Feb 23 and 24 this year in New York is an annual conference dedicated to bringing theory and practice of fair and interpretable Machine Learning, Information Retrieval, NLP, Computer Vision, Recommender systems, and other technical disciplines. This year's program includes 17 peer-reviewed papers and 6 tutorials from leading experts in the field. The conference will have three sessions. Session 4 of the two-day conference on Saturday, February 24, is in the field of fair classification. In this article, we give our readers a peek into the four papers that have been selected for presentation in Session 4. You can also check out Session 1,  Session 2, and Session 3 summaries in case you’ve missed them. The cost of fairness in binary classification What is the paper about? This paper provides a simple approach to the Fairness-aware problem which involves suitably thresholding class-probability estimates. It has been awarded Best paper in Technical contribution category. The authors have studied the inherent tradeoffs in learning classifiers with a fairness constraint in the form of two questions: What is the best accuracy we can expect for a given level of fairness? What is the nature of these optimal fairness aware classifiers? The authors showed that for cost-sensitive approximate fairness measures, the optimal classifier is an instance-dependent thresholding of the class probability function. They have quantified the degradation in performance by a measure of alignment of the target and sensitive variable. This analysis is then used to derive a simple plugin approach for the fairness problem. Key takeaways For Fairness-aware learning, the authors have designed an algorithm targeting a particular measure of fairness. They have reduced two popular fairness measures (disparate impact and mean difference) to cost-sensitive risks. They show that for cost-sensitive fairness measures, the optimal Fairness-aware classifier is an instance-dependent thresholding of the class-probability function. They quantify the intrinsic, method independent impact of the fairness requirement on accuracy via a notion of alignment between the target and sensitive feature. The ability to theoretically compute the tradeoffs between fairness and utility is perhaps the most interesting aspect of their technical results. They have stressed that the tradeoff is intrinsic to the underlying data. That is, any fairness or unfairness, is a property of the data, not of any particular technique. They have theoretically computed what price one has to pay (in utility) in order to achieve a desired degree of fairness: in other words, they have computed the cost of fairness. Decoupled Classifiers for Group-Fair and Efficient Machine Learning What is the paper about? This paper considers how to use a sensitive attribute such as gender or race to maximize fairness and accuracy, assuming that it is legal and ethical. Simple linear classifiers may use the raw data, upweight/oversample data from minority groups, or employ advanced approaches to fitting linear classifiers that aim to be accurate and fair. However, an inherent tradeoff between accuracy on one group and accuracy on another still prevails. This paper defines and explores decoupled classification systems, in which a separate classifier is trained on each group. The authors present experiments on 47 datasets. The experiments are “semi-synthetic” in the sense that the first binary feature was used as a substitute sensitive feature. The authors found that on many data sets the decoupling algorithm improves performance while less often decreasing performance. Key takeaways The paper describes a simple technical approach for a practitioner using ML to incorporate sensitive attributes. This approach avoids unnecessary accuracy tradeoffs between groups and can accommodate an application-specific objective, generalizing the standard ML notion of loss. For a certain family of “weakly monotonic” fairness objectives, the authors provide a black-box reduction that can use any off-the-shelf classifier to efficiently optimize the objective. This work requires the application designer to pin down a specific loss function that trades off accuracy for fairness. Experiments demonstrate that decoupling can reduce the loss on some datasets for some potentially sensitive features A case study of algorithm-assisted decision making in child maltreatment hotline screening decisions What is the paper about? The work is based on the use of predictive analytics in the area of child welfare. It won the best paper award in the Technical and Interdisciplinary Contribution. The authors have worked on developing, validating, fairness auditing, and deploying a risk prediction model in Allegheny County, PA, USA. The authors have described competing models that are being developed in the Allegheny County as part of an ongoing redesign process in comparison to the previous models. Next, they investigate the predictive bias properties of the current tool and a Random forest model that has emerged as one of the best performing competing models. Their predictive bias assessment is motivated both by considerations of human bias and recent work on fairness criteria. They then discuss some of the challenges in incorporating algorithms into human decision-making processes and reflect on the predictive bias analysis in the context of how the model is actually being used. They also propose an “oracle test” as a tool for clarifying whether particular concerns pertain to the statistical properties of a model or if these concerns are targeted at other potential deficiencies. Key takeaways The goal in Allegheny County is to improve both the accuracy and equity of screening decisions by taking a Fairness-aware approach to incorporating prediction models into the decision-making pipeline. The paper reports on the lessons learned so far by the authors, their approaches to predictive bias assessment, and several outstanding challenges in the child maltreatment hotline context. This report contributes to the ongoing conversation concerning the use of algorithms in supporting critical decisions in government—and the importance of considering fairness and discrimination in data-driven decision making. The paper discussion and general analytic approach are also broadly applicable to other domains where predictive risk modeling may be used. Fairness in Machine Learning: Lessons from Political Philosophy What is the paper about? Plenty of moral and political philosophers have expended significant efforts in formalizing and defending the central concepts of discrimination, egalitarianism, and justice. Thus it is unsurprising to know that the attempts to formalize ‘fairness’ in machine learning contain echoes of these old philosophical debates. This paper draws on existing work in moral and political philosophy in order to elucidate emerging debates about fair machine learning. It answers the following questions: What does it mean for a machine learning model to be ‘fair’, in terms which can be operationalized? Should fairness consist of ensuring everyone has an equal probability of obtaining some benefit, or should we aim instead to minimize the harms to the least advantaged? Can the relevant ideal be determined by reference to some alternative state of affairs in which a particular social pattern of discrimination does not exist? Key takeaways This paper aims to provide an overview of some of the relevant philosophical literature on discrimination, fairness, and egalitarianism in order to clarify and situate the emerging debate within fair machine learning literature. The author addresses the conceptual distinctions drawn between terms frequently used in the fair ML literature–including ‘discrimination’ and ‘fairness’–and the use of related terms in the philosophical literature. He suggests that ‘fairness’ as used in the fair machine learning community is best understood as a placeholder term for a variety of normative egalitarian considerations. He also provides an overview of implications for the incorporation of ‘fairness’ into algorithmic decision-making systems. We hope you like the coverage of Session 4. Don’t miss our coverage on Session 5 on Fat recommenders and more.

2017 has been a wonderful year for Machine Learning. Developing smart, intelligent models has now become easier than ever thanks to the extensive research into and development of newer and more efficient tools and frameworks. While the likes of Tensorflow, Keras, PyTorch and some more have ruled the roost in 2017 as the top machine learning and deep learning libraries, 2018 promises to be even more exciting with a strong line-up of open source and enterprise tools ready to take over - or at least compete with - the current lot. In this article, we take a look at 10 such tools and frameworks which are expected to make it big in 2018. Amazon Sagemaker One of the major announcements in the AWS re:Invent 2017 was the general availability of Amazon Sagemaker - a new framework that eases the building and deployment of machine learning models on the cloud. This service will be of great use to developers who don’t have a deep exposure to machine learning, by giving them a variety of pre-built development environments, based on the popular Jupyter notebook format. Data scientists looking to build effective machine learning systems on AWS and to fine-tune their performance without spending a lot of time will also find this service useful. DSSTNE Yet another offering by Amazon, DSSTNE (popularly called as Destiny) is an open source library for developing machine learning models. It’s primary strength lies in the fact that it can be used to train and deploy recommendation models which work with sparse inputs. The models developed using DSSTNE can be trained to use multiple GPUs, are scalable and are optimized for fast performance. Boasting close to 4000 stars on GitHub, this library is yet another tool to look out for in 2018! Azure Machine Learning Workbench Way back in 2014, Microsoft put Machine Learning and AI capabilities on the cloud by releasing Azure Machine Learning. However, this was strictly a cloud-only service. During the Ignite 2017 conference held in September, Microsoft announced the next generation of Machine Learning on Azure - bringing machine learning capabilities to the organizations through their Azure Machine Learning Workbench. Azure ML Workbench is a cross-platform client which can run on both Windows and Apple machines. It is tailor-made for data scientists and machine learning developers who want to perform their data manipulation and wrangling tasks. Built for scalability, users can get intuitive insights from a broad range of data sources and use them for their data modeling tasks. Neon Way back in 2016, Intel announced their intentions to become a major player in the AI market with the $350 million acquisition of Nervana, an AI startup which had been developing both hardware and software for effective machine learning. With Neon, they now have a fast, high-performance deep learning framework designed specifically to run on top of the recently announced Nervana Neural Network Processor. Designed for ease of use and supporting integration with the iPython notebook, Neon supports training of common deep learning models such as CNN, RNN, LSTM and others. The framework is showing signs of continuous improvement and with over 3000 stars on GitHub, Neon looks set to challenge the major league of deep learning libraries in the years to come. Microsoft DMLT One of the major challenges with machine learning for enterprises is the need to scale out the models quickly, without compromising on the performance while minimising significant resource consumption. Microsoft’s Distributed Machine Learning framework is designed to do just that. Open sourced by Microsoft so that it can receive a much wider support from the community, DMLT allows machine learning developers and data scientists to take their single-machine algorithms and scale them out to build high performance distributed models. DMLT mostly focuses on distributed machine learning algorithms and allows you to perform tasks such as word embedding, sampling, and gradient boosting with ease. The framework does not have support for training deep learning models yet, however, we can expect this capability to be added to the framework very soon. Google Cloud Machine Learning Engine Considered to be Google’s premium machine learning offering, the Cloud Machine Learning Engine allows you to build machine learning models on all kinds of data with relative ease. Leveraging the popular Tensorflow machine learning framework, this platform can be used to perform predictive analytics at scale. It also lets you fine-tune and optimize the performance of your machine learning models using the popular HyperTune feature. With a serverless architecture supporting automated monitoring, provisioning and scaling, the Machine Learning Engine ensures you only have to worry about the kind of machine learning models you want to train. This feature is especially useful for machine learning developers looking to build large-scale models on the go. Apple Core ML Developed by Apple to help iOS developers build smarter applications, the Core ML framework is what makes Siri smarter. It takes advantage of both CPU and GPU capabilities to allow the developers to build different kinds of machine learning and deep learning models, which can then be integrated seamlessly into the iOS applications. Core ML supports all popularly used machine learning algorithms such as decision trees, Support Vector Machines, linear models and more. Targeting a variety of real-world use-cases such as natural language processing, computer vision and more, Core ML’s capabilities make it possible to analyze data on the Apple devices on the go, without having to import to the models for learning. Apple Turi Create In many cases, the iOS developers want to customize the machine learning models they want to integrate into their apps. For this, Apple has come up with Turi Create. This library allows you to focus on the task at hand rather than deciding which algorithm to use. You can be flexible in terms of the data set, the scale at which the model needs to operate and what platform the models need to be deployed to. Turi Create comes in very handy for building custom models for recommendations, image processing, text classification and many more tasks. All you need is some knowledge of Python to get started! Convnetjs Move over supercomputers and clusters of machines, deep learning is well and truly here - on your web browsers! You can now train your advanced machine learning and deep learning models directly on your browser, without needing a CPU or a GPU, using the popular Javascript-based Convnetjs library. Originally written by Andrej Karpathy, the current director of AI at Tesla, the library has since been open sourced and extended by the contributions of the community. You can easily train deep neural networks and even reinforcement learning models on your browser directly, powered by this very unique and useful library. This library is suited for those who do not wish to purchase serious hardware for training computationally-intensive models. With close to 9000 stars on GitHub, Convnetjs has been one of the rising stars in 2017 and is quickly becoming THE go-to library for deep learning. BigML BigML is a popular machine learning company that provides an easy to use platform for developing machine learning models. Using BigML’s REST API, you can seamlessly train your machine learning models on their platform. It allows you to perform different tasks such as anomaly detection, time series forecasting, and build apps that perform real-time predictive analytics. With BigML, you can deploy your models on-premise or on the cloud, giving you the flexibility of selecting the kind of environment you need to run your machine learning models. True to their promise, BigML really do make ‘machine learning beautifully simple for everyone’. So there you have it! With Microsoft, Amazon, and Google all fighting for supremacy in the AI space, 2018 could prove to be a breakthrough year for developments in Artificial Intelligence. Add to this mix the various open source libraries that aim to simplify machine learning for the users, and you get a very interesting list of tools and frameworks to keep a tab on. The exciting thing about all this is - all of them possess the capability to become the next TensorFlow and cause the next AI disruption.  

Conversational AI is one of the most interesting applications of artificial intelligence in recent years. While the trend isn’t yet ubiquitous in the way that recommendation systems are (perhaps unsurprising), it has been successfully productized by a number of tech giants, in the form of Google Home and Amazon Echo (which is ‘powered by’ Alexa). The conversational AI arms race Arguably, 2018 has seen a bit of an arms race in conversational AI. As well as Google and Amazon, the likes of IBM, Microsoft, and Apple have wanted a piece of the action. Here are some of the new conversational AI tools and products these companies introduced this year: Google Google worked towards enhancing its conversational interface development platform, Dialogflow. In July, at the Google Cloud Next event, it announced several improvements and new capabilities to Dialogflow including Text to Speech via DeepMind's WaveNet and Dialogflow Phone Gateway for telephony integration. It also launched a new product called Contact Center AI that comes with Dialogflow Enterprise Edition and additional capabilities to assist live agents and perform analytics. Google Assistant became better in having a back-and-forth conversation with the help of Continued Conversation, which was unveiled at the Google I/O conference. The assistant became multilingual in August, which means users can speak to it in more than one language at a time, without having to adjust their language settings. Users can enable this multilingual functionality by selecting two of the supported languages. Following the footsteps of Amazon, Google also launched its own smart display named Google Home Hub at the ‘Made by Google’ event held in October. Microsoft Microsoft in 2018 introduced and improved various bot-building tools for developers. In May, at the Build conference, Microsoft announced major updates in their conversational AI tools: Azure Bot Service, Microsoft Cognitive Services Language Understanding, and QnAMaker. To enable intelligent bots to learn from example interactions and handle common small talk, it launched new experimental projects from named Conversation Learner and Personality Chat. At Microsoft Ignite, Bot Framework SDK V4.0 was made generally available. Later in November, Microsoft announced the general availability of the Bot Framework Emulator V4 and Web Chat control. In May, to drive more research and development in its conversational AI products, Microsoft acquired Semantic Machines and established conversational AI center of excellence in Berkeley. In November, the organization's acquisition of Austin-based bot startup XOXCO was a clear indication that it wants to get serious about using artificial intelligence for conversational bots. Producing guidelines on developing ‘responsible’ conversational AI further confirmed Microsoft wants to play a big part in the future evolution of the area. Microsoft were the chosen tech partner by UK based conversational AI startup ICS.ai. The team at ICS are using Azure and LUIS from Microsoft in their public sector AI chatbots, aimed at higher education, healthcare trusts and county councils. Amazon Amazon with the aims to improve Alexa’s capabilities released Alexa Skills Kit (ASK) which consists of APIs, tools, documentation, and code samples using which developers can build new skills for Alexa. In September, it announced a preview of a new design language named Alexa Presentation Language (APL). With APL, developers can build visual skills that include graphics, images, slideshows, and video, and to customize them for different device types. Amazon’s smart speaker Echo Dot saw amazing success with becoming the best seller in smart speaker category on Amazon. At its 2018 hardware event in Seattle, Amazon announced the launch of redesigned Echo Dot and a new addition to Alexa-powered A/V device called Echo Plus. As well as the continuing success of Alexa and the Amazon Echo, Amazon’s decision to launch the Alexa Fellowship at a number of leading academic institutions also highlights that for the biggest companies conversational AI is as much about research and exploration as it is products. Like Microsoft, it appears that Amazon is well aware that conversational AI is an area only in its infancy, still in development - as much as great products, it requires clear thinking and cutting-edge insight to ensure that it develops in a way that is both safe and impactful. What’s next? This huge array of products is a result of advances in deep learning researches. Now conversational AI is not just limited to small tasks like setting an alarm or searching the best restaurant. We can have a back and forth conversation with the conversational agent. But, needless to say, it still needs more work. Conversational agents are yet to meet user expectations related to sensing and responding with emotion. In the coming years, we will see these systems understand and do a good job at generating natural language. They will be able to have reasonably natural conversations with humans in certain domains, grounded in context. Also, the continuous development in IoT will provide AI systems with more context. Apple has introduced Shortcuts for iOS 12 to automate your everyday tasks Microsoft amplifies focus on conversational AI: Acquires XOXCO; shares guide to developing responsible bots Amazon is supporting research into conversational AI with Alexa fellowships

Machine learning has always been dependent on the selection of the right features within a given model; even the selection of the right algorithm. But deep learning changed this. The selection process is now built into the models themselves. Researchers and engineers are now shofting their focus from feature engineering to network engineering. Out of this, AutoML, or meta learning, has become an increasingly important part of deep learning. AutoML is an emerging research topic which aims at auto-selecting the most efficient neural network for a given learning task. In other words, AutoML represents a set of methodologies for learning how to learn efficiently. Consider for instance the tasks of machine translation, image recognition, or game playing. Typically, the models are manually designed by a team of engineers, data scientist, and domain experts. If you consider that a typical 10-layer network can have ~1010 candidate network, you understand how expensive, error prone, and ultimately sub-optimal the process can be. This article is an excerpt from a book written by Antonio Gulli and Amita Kapoor titled TensorFlow 1.x Deep Learning Cookbook. This book is an easy-to-follow guide that lets you explore reinforcement learning, GANs, autoencoders, multilayer perceptrons and more. AutoML with recurrent networks and with reinforcement learning The key idea to tackle this problem is to have a controller network which proposes a child model architecture with probability p, given a particular network given in input. The child is trained and evaluated for the particular task to be solved (say for instance that the child gets accuracy R). This evaluation R is passed back to the controller which, in turn, uses R to improve the next candidate architecture. Given this framework, it is possible to model the feedback from the candidate child to the controller as the task of computing the gradient of p and then scale this gradient by R. The controller can be implemented as a Recurrent Neural Network (see the following figure). In doing so, the controller will tend to privilege iteration after iterations candidate areas of architecture that achieve better R and will tend to assign a lower probability to candidate areas that do not score so well. For instance, a controller recurrent neural network can sample a convolutional network. The controller can predict many hyper-parameters such as filter height, filter width, stride height, stride width, and the number of filters for one layer and then can repeat. Every prediction can be carried out by a softmax classifier and then fed into the next RNN time step as input. This is well expressed by the following images taken from Neural Architecture Search with Reinforcement Learning, Barret Zoph, Quoc V. Le: Predicting hyperparameters is not enough as it would be optimal to define a set of actions to create new layers in the network. This is particularly difficult because the reward function that describes the new layers is most likely not differentiable. This makes it impossible to optimize using standard techniques such as SGD. The solution comes from reinforcement learning. It consists of adopting a policy gradient network. Besides that, parallelism can be used for optimizing the parameters of the controller RNN. Quoc Le & Barret Zoph proposed to adopt a parameter-server scheme where we have a parameter server of S shards, that store the shared parameters for K controller replicas. Each controller replica samples m different child architectures that are trained in parallel as illustrated in the following images, taken from Neural Architecture Search with Reinforcement Learning, Barret Zoph, Quoc V. Le: Quoc and Barret applied AutoML techniques for Neural Architecture Search to the Penn Treebank dataset, a well-known benchmark for language modeling. Their results improve the manually designed networks currently considered the state-of-the-art. In particular, they achieve a test set perplexity of 62.4 on the Penn Treebank, which is 3.6 perplexity better than the previous state-of-the-art model. Similarly, on the CIFAR-10 dataset, starting from scratch, the method can design a novel network architecture that rivals the best human-invented architecture in terms of test set accuracy. The proposed CIFAR-10 model achieves a test error rate of 3.65, which is 0.09 percent better and 1.05x faster than the previous state-of-the-art model that used a similar architectural scheme. Meta-learning blocks In Learning Transferable Architectures for Scalable Image Recognition, Barret Zoph, Vijay Vasudevan, Jonathon Shlens, Quoc V. Le, 2017. propose to learn an architectural building block on a small dataset that can be transferred to a large dataset. The authors propose to search for the best convolutional layer (or cell) on the CIFAR-10 dataset and then apply this learned cell to the ImageNet dataset by stacking together more copies of this cell, each with their own parameters. Precisely, all convolutional networks are made of convolutional layers (or cells) with identical structures but different weights. Searching for the best convolutional architectures is therefore reduced to searching for the best cell structures, which is faster more likely to generalize to other problems. Although the cell is not learned directly on ImageNet, an architecture constructed from the best learned cell achieves, among the published work, state-of-the-art accuracy of 82.7 percent top-1 and 96.2 percent top-5 on ImageNet. The model is 1.2 percent better in top-1 accuracy than the best human-invented architectures while having 9 billion fewer FLOPS—a reduction of 28% from the previous state of the art model. What is also important to notice is that the model learned with RNN+RL (Recurrent Neural Networks + Reinforcement Learning) is beating the baseline represented by Random Search (RS) as shown in the figure taken from the paper. In the mean performance of the top-5 and top-25 models identified in RL versus RS, RL is always winning: AutoML and learning new tasks Meta-learning systems can be trained to achieve a large number of tasks and are then tested for their ability to learn new tasks. A famous example of this kind of meta-learning is transfer learning, where networks can successfully learn new image-based tasks from relatively small datasets. However, there is no analogous pre-training scheme for non-vision domains such as speech, language, and text. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, Chelsea Finn, Pieter Abbeel, Sergey Levine, 2017, proposes a model- agnostic approach names MAML, compatible with any model trained with gradient descent and applicable to a variety of different learning problems, including classification, regression, and reinforcement learning. The goal of meta-learning is to train a model on a variety of learning tasks, such that it can solve new learning tasks using only a small number of training samples. The meta-learner aims at finding an initialization that rapidly adapts to various problems quickly (in a small number of steps) and efficiently (using only a few examples). A model represented by a parametrized function fθ with parameters θ.When adapting to a new task Ti, the model's parameters θ become θi  . In MAML, the updated parameter vector θi  is computed using one or more gradient descent updates on task Ti. For example, when using one gradient update, θ ~ = θ − α∇θLTi (fθ) where LTi is the loss function for the task T and α is a meta-learning parameter. The MAML algorithm is reported in this figure: MAML was able to substantially outperform a number of existing approaches on popular few-shot image classification benchmark. Few shot image is a quite challenging problem aiming at learning new concepts from one or a few instances of that concept. As an example, Human-level concept learning through probabilistic program induction, Brenden M. Lake, Ruslan Salakhutdinov, Joshua B. Tenenbaum, 2015, suggested that humans can learn to identify novel two-wheel vehicles from a single picture such as the one contained in the box as follows: If you enjoyed this excerpt, check out the book TensorFlow 1.x Deep Learning Cookbook, to skill up and implement tricky neural networks using Google's TensorFlow 1.x AmoebaNets: Google’s new evolutionary AutoML AutoML : Developments and where is it heading to What is Automated Machine Learning (AutoML)?

Atlantic Council CEO, Frederick Kempe spoke in the World Economic Forum (WEF) in Davos, Switzerland. He talked about the Cold war between the US and China and why the countries need to co-operate and not compete in the tech arms race, in his presentation Future Frontiers of Technology Control. He began his presentation by posing a question set forth by Former US Foreign National Security Advisor Stephen Hadley, “Can the incumbent US and insurgent China become strategic collaborators and strategic competitors in this tech space at the same time?” Read also: The New AI Cold War Between China and the USA Kempe’s three framing arguments Geopolitical Competition This fusion of tech breakthroughs blurring lines of the physical, digital, and biological space is reaching an inflection point that makes it already clear that they will usher in a revolution that will determine the shape of the global economy. It will also determine which nations and political constructs may assume the commanding heights of global politics in the coming decade. Technological superiority Over the course of history, societies that dominated economic innovation and progress have dominated in international relations — from military superiority to societal progress and prosperity. On balance, technological progress has contributed to higher standards of living in most parts of the world; however, the disproportionate benefit goes to first movers. Commanding Heights The technological arms race for supremacy in the fourth industrial revolution has essentially become a two-horse contest between the United States and China. We are in the early stages of this race, but how it unfolds and is conducted will do much to shape global human relations. The shift in 2018 in US-China relations from a period of strategic engagement to greater strategic competition has also significantly accelerated the Tech arms race. China vs the US: Why China has the edge? It was Vladimir Putin, President of the Russian Federation who said that “The one who becomes the leader in Artificial Intelligence, will rule the world.” In 2017, DeepMind’s AlphaGo defeated a Chinese master in Go, a traditional Chinese game. Following this defeat, China launched an ambitious roadmap, called the next generation AI plan. The goal was to become the Global leader in AI by 2030 in theory, technology, and application. On current trajectories, in the four primary areas of AI over the next 5 years, China will emerge the winner of this new technology race. Kempe also quotes, author of the book, AI superpowers, Kai-fu Lee who argues that harnessing of the power of AI today- the electricity of the 21st century- requires abundant data, hungry entrepreneurs, AI scientists, and an AI friendly policy. He believes that China has the edge in all of these. The current AI has translated from out of the box research, where the US has expertise in, to actual implementation, where China has the edge. Per, Kai-fu Lee China already has the edge in entrepreneurship, data, and government support, and is rapidly catching up to the U.S. in expertise. The world has translated from the age of world-leading expertise (US department) to the age of data, where China wins hands down. Economists call China the Saudi Arabia of Data and with that as the fuel for AI, it has an enormous advantage. The Chinese government without privacy restrictions can gain and use data in a manner that is out of reach of any democracy. Kemper concludes that the nature of this technological arms contest may favor insurgent China rather than the incumbent US. What are the societal implications of this tech cold war He also touched upon the societal implications of AI and the cold war between the US and China. A number of jobs will be lost by 2030. Quoting from Kai-fu Lee’s book, Kempe says that Job displacement caused by artificial intelligence and advanced robotics could possibly displace up to 54 million US workers which comprise 30% of the US labor force. It could also displace up to 100 million Chinese workers which are 12% of the Chinese labor force. What is the way forward with these huge societal implications of a bi-lateral race underway? Kempe sees three possibilities. A sloppy Status Quo A status quo where China and the US will continue to cooperate but increasingly view each other with suspicion. They will manage their rising differences and distrust imperfectly, never bridging them entirely, but also not burning bridges, either between researchers, cooperations, or others. Techno Cold War China and the US turn the global tech contest into more of a zero-sum battle for global domination. They organize themselves in a manner that separates their tech sectors from each other and ultimately divides up the world. Collaborative Future - the one we hope for Nicholas Thompson and Ian Bremmer argued in a wired interview that despite the two countries’ societal difference, the US should wrap China in a tech embrace. The two countries should work together to establish international standards to ensure that the algorithms governing people’s lives and livelihoods are transparent and accountable. They should recognize that while the geopolitics of technological change is significant, even more important will be the challenges AI poses to all societies across the world in terms of job automation and the social disruptions that may come with it. It may sound utopian to expect US and China to cooperate in this manner, but this is what we should hope for. To do otherwise would be self-defeating and at the cost of others in the global community which needs our best thinking to navigate the challenges of the fourth industrial revolution. Kempe concludes his presentation with a quote by Henry Kissinger, Former US Secretary of State and National Security Advisor, “We’re in a position in which the peace and prosperity of the world depend on whether China and the US can find a method to work together, not always in agreement, but to handle our disagreements...This is the key problem of our time.” Note: All images in this article are taken from Frederick Kempe’s presentation. We must change how we think about AI, urge AI founding fathers Does AI deserve to be so Overhyped? Alarming ways governments are using surveillance tech to watch you

Pieter Abbeel is a professor at UC Berkeley and a former Research Scientist at OpenAI. His current research focuses on robotics and machine learning with particular focus on deep reinforcement learning, deep imitation learning, deep unsupervised learning, meta-learning, learning-to-learn, and AI safety. This article attempts to bring our readers to Pieter’s fantastic Keynote speech at NIPS 2017. It talks about the implementation of Deep Reinforcement Learning in Robotics, what challenges exist and how these challenges can be overcome. Once you’ve been through this article, we’re certain you’d be extremely interested in watching the entire video on the NIPS Facebook page. All images in this article come from his presentation slides and do not belong to us. Robotics in ML has been growing in leaps and bounds with several companies investing huge amounts to tie both these technologies together in the best way possible. Although, there are still several aspects that are not thoroughly accomplished when it comes to AI Robotics. Here are a few of them: Maximize Signal Extracted from Real World Experience Faster/Data efficient Reinforcement Learning Long Horizon Reasoning Taskability (Imitation Learning) Lifelong Learning (Continuous Adaptation) Leverage Simulation Maximise signal extracted from real world experience We need more real world data, so we need to extract as much signal from it. In the diagram below, are the different layers of machine learning that engineers perform. There are engineers who look at the entire cake and train the agent to take both the learning from the reward and from auxiliary signals. This is because using only Reinforcement Learning doesn’t give you a lot of signal. Is there then, a possibility of having a Reward Signal in RL that ties more RL into the system? There’s something known as Hindsight Experience Replay. The idea is to get a reward signal from any experience by assuming the goal equals whatever happened, and not just from success like in usual RL. For this, we need to assume that whatever the agent does is a success. We use Q-learning and instead of a standard Q function, we use multiple goals even though they were not really a goal when you were acting.  Here, a replay buffer collects experience, Q-learning is then applied and a hindsight replay is performed to infuse a new reward for everything the agent has done. For various robotic tasks like pushing, sliding and picking and placing objects, this does very well. Faster Reinforcement Learning When we’re talking about faster RL, we’re talking about much more data efficient RL. Here is a diagram that demonstrates standard RL: An agent lets a robot perform an action in a particular environment or situation in order to achieve a reward. Here, the goal is to maximise the reward. As against Supervised Learning, there is no supervision as to whether the actions taken by the agents are right or wrong. That brings in a few additional challenges in RL, which are: Credit assignment: This is a major problem and is where you get the signal from in RL Stability: Because of the feedback loop, the system could destabilize and destroy itself Exploration: Doing things you’ve never done before when the only way to learn is based on what you’ve done before Despite this, there have been great improvements in Reinforcement Learning in the past few years, enabling AI systems to play games like Go, Dota, etc. It has also been implemented in building robots by NASA for planetary exploration. But the question still exists: “How good is learning?” In the game of pong, a human takes roughly 2 hours to learn what Deep Q-Network (DQN) learns in 40 hours! A more careful study reveals that after 15 minutes, humans tend to outperform DDQN that has trained for 115 hours. This is a tremendous gap in terms of learning efficiency. So, how do we overcome the challenge? Several fully generalised algorithms like Trust Region Policy Optimization (TRPO), DQN, Asynchronous Actor-Critic Agents (A3C) and Rainbow are available, meaning that they can be applied to any kind of environment. Although, only a very small subset of environments are actually encountered in the real world. Can we develop fast RL algorithms that take advantage of this situation? RL Agents can be reused to train various policies. The RL algorithm is developed to train the policy to adapt to a particular environment A. This can then be replicated to environment B and so on. Humans develop the RL algorithm and then rely on it to train the policy. Despite this, none of the algorithms are as good as human learners. Do we have an alternative then? Indeed, yes! Why not let the system learn not just the policy but the algorithm as well or in other words, the entire agent? Enter Meta-Reinforcement Learning In Meta-RL, the learning algorithm itself is being learnt. You could relate this to meta-programming, where one program is trained to write another. This process helps a system learn the world better so it can pick up on learning a new situation quicker. So how does this work? The system is faced with many environments, so that it learns the algorithms and then outputs a faster RL Agent. So, when faced with a new environment, it quickly adapts to it. For evaluating the actual performance, the Multi-armed bandits problem can be considered. Here’s the setting: each bandit has its own distribution over payouts, and in each episode you can choose one bandit. A good RL agent should be able to explore a sufficient number of bandits and exploit the best ones. We need to come up with an algorithm that pulls a higher probability of payoff, rather than a low probability. There are already several asymptotically optimal algorithms like Gittins index, UCB1, Thompson Sampling, that have been created to solve this problem. Here’s a comparison of some of them with the Meta-RL algorithm. The result is quite impressive. The Meta-RL algorithm is equally competitive with Gittins. In a case where the task is to obtain an on target running direction as well as attain the maximum speed, the agent when dropped into an environment is able to master the the task almost instantly. However, meta-learning succeeds only 2/3rd of the time. It doesn’t succeed the rest of the time due to two main reasons. Overfitting: You would usually tend to overfit to the current situation rather than generically fitting to situations Underfitting: This is when you don’t get enough signal to get any rewards The solution is to put a different structure underneath the system. Instead of using an RNN, we use a wavenet like architecture or maybe Simple Neural Attentive Meta-Learner (SNAIL). SNAIL is able to perform a bit better than RL2 in the same Bandits problem. Longer Horizon Reasoning We need to learn to reason over longer horizons than what canonical algorithms do. For this, we need hierarchy. For example, suppose a robot has to perform 10 tasks in a day. This would mean it has 10 timesteps per day? Each of these 10 tasks would have subtasks under them. Let’s assume that would make it a total of 1000 time steps. To perform these tasks, the robot would need footstep planning, which would amount to 100,000 time steps. Footsteps in turn require commands to be sent to motors, which would make it 100,000,000 time steps. This is a very long horizon. We can formulate this as a meta-learning problem. The agent has to solve a distribution of related long-horizon tasks with the goal of learning new tasks in the distribution quickly. If that is our objective, hierarchy would fall out. Taskability (Imitation Learning) There are several things we want from robots. We need to be able to tell them what to do and we can do this by giving them examples. This is called Imitation Learning, which can be successfully implemented to a variety of use cases. The idea is to collect many demonstrations, then train something from those demonstrations, then deploy the learn policy. The problem with this is that everytime there is a new task, you start from scratch. The solution to this problem is experience through several demonstrations, as in the case of humans. Although, instead of running the agent through several demos, it is trained completely on one, then showed a frame of a second demo, where it uses it to predict what the outcome would be. This is known as One-Shot imitation learning which is a part of supervised learning, where in several demonstrations are used to train the system to be able to handle any new environment it is put into. Lifelong learning (Continuous Adaptation) What we usually do in ML can be divided into two broad steps: Run Machine Learning Deploy it, which is a canonical way In this case, all the learning happens ahead of time, before the deployment. However, in real world cases, what you learn from past data might not work in the future. There is a necessity to learn during deployment, which is a lifelong learning spirit. This brings us to Continuous Adaptation. Can we train an agent to be good at non stationary environments? We need to find whether at the time of meta training the agent is able to adapt to a new/changing task. We can try changing the dynamics since it’s hard to do ML training in the real world. At the same time, we can also use competitor environments; which means you’re in an environment with other agents who are trying to beat your agent. The only way to succeed is to continuously adapt more quickly than the others. Leverage Simulation Simulation is very helpful and it’s not that expensive. It’s fast and scalable and lets you label more easily. However, the challenge is how to get useful things out of the simulator. One approach is to build realistic simulators. This is quite expensive. Another way is to use a close enough simulator that uses real world data through domain confusion or adaptation. It allows to learn from a small amount of real world data and is quite successful. Further, another approach to look at is Domain Randomisation, which is also working well in the real world. If the model sees enough simulated variations, the real world might appear like just the next simulator. This has worked in the context of using simulator data to train a quadcopter to avoid collision. Moreover, when pre trained from imagenet or just training in simulation, both performances were similar, after around 8000 examples. To conclude, the beauty of meta learning is that it enables the discovery of algorithms that are data driven, as against those that are created from pure human ingenuity. This requires more compute power, but several companies like Nvidia and Intel are working hard to overcome this challenge. This will surely power meta-learning to great heights to be implemented in robotics. While we figure out these above mentioned technical challenges of incorporating AI in robotics, some significant other challenges that we must focus on in parallel are safe learning, and value alignment among others.    

Stack Overflow recently published their annual developer survey in which over 100,000 developers and professionals participated. The survey shed light on some very interesting insights - from the developers’ preferred language for programming, to the development platform they hate the most. As the survey is quite detailed and comprehensive, we thought why not present the most important takeaways and findings for you to go through very quickly? If you are short of time and want to scan through the results of the survey quickly, read on.. Developer Profile Young developers form the majority: Half the developer population falls in the age group of 25-34 years while almost all respondents (90%) fall within the 18 - 44 year age group. Limited professional coding experience: Majority of the developers have been coding from the last 2 to 10 years. That said, almost half of the respondents have a professional coding experience of less than 5 years. Continuous learning is key to surviving as a developer: Almost 75% of the developers have a bachelor’s degree, or higher. In addition, almost 90% of the respondents say they have learnt a new language, framework or a tool without taking any formal course, but with the help of the official documentation and/or Stack Overflow Back-end developers form the majority: Among the top developer roles, more than half the developers identify themselves as back-end developers, while the percentage of data scientists and analysts is quite low. About 20% of the respondents identify themselves as mobile developers Working full-time: More than 75% of the developers responded that they work a full-time job. Close to 10% are freelancers, or self-employed. Popularly used languages and frameworks The Javascript family continue their reign: For the sixth year running, JavaScript has continued to be the most popular programming language, and is the choice of language for more than 70% of the respondents. In terms of frameworks, Node.js and Angular continue to be the most popular choice of the developers. Desktop development ain’t dead yet: When it comes to the platforms, developers prefer Linux and Windows Desktop or Server for their development work. Cloud platforms have not gained that much adoption, as yet, but there is a slow but steady rise. What about Data Science? Machine Learning and DevOps rule the roost: Machine Learning and DevOps are two trends which are trending highly due to the vast applications and research that is being done on these fronts. Tensorflow rises, Hadoop falls: About 75% of the respondents love the Tensorflow framework, and say they would love to continue using it for their machine learning/deep learning tasks. Hadoop’s popularity seems to be going down, on the other hand, as other Big Data frameworks like Apache Spark gain more traction and popularity. Python - the next big programming language: Popular data science languages like R and Python are on the rise in terms of popularity. Python, which surpassed PHP last year, has surpassed C# this year, indicating its continuing rise in popularity. Python based Frameworks like Tensorflow and pyTorch are gaining a lot of adoption. Learn F# for more moolah: Languages like F#, Clojure and Rust are associated with high global salaries, with median salaries above $70,000. The likes of R and Python are associated with median salaries of up to $57,000. PostgreSQL growing rapidly, Redis most loved database: MySQL and SQL Server are the two most widely used databases as per the survey, while the usage of PostgreSQL has surpassed that of the traditionally popular databases like MongoDB and Redis. In terms of popularity, Redis is the most loved database while the developers dread (read looking to switch from) databases like IBM DB2 and Oracle. Job-hunt for data scientists: Approximately 18% of the 76,000+ respondents who are actively looking for jobs are data scientists or work as academicians and researchers. AI more exciting than worrying: Close to 75% of the 69,000+ respondents are excited about the future possibilities with AI than worried about the dangers posed by AI. Some of the major concerns include AI making important business decisions. The big surprise was that most developers find automation of jobs as the most exciting part of a future enabled by AI. So that’s it then! What do you think about the Stack Overflow Developer survey results? Do you agree with the developers’ responses? We would love to know your thoughts. In the coming days, watch out for more fine grained analysis of the Stack Overflow survey data.

In a paper published on February 4, Google engineers drafted out plans to forward federated learning at a scale. It showcases the high-level plans, challenges, solutions, and applications. Federated learning was first introduced in 2017 by Google. The idea is to use data from a number of computing devices like smartphones instead of a centralized data source. Federated learning can help with privacy Federated learning can be beneficial as it addresses the privacy concern. Android phones are used for the system where the data is only used but never uploaded to any server. A deep neural network is trained by using TensorFlow on the data stored in the Android phone. The Federated averaging algorithm by Brendan McMahan uses a similar approach as synchronous training. The weights of the neural network are combined in the cloud using Federated Averaging. This creates a global model which is then pushed back to the phones as results/desirable actions. To enhance privacy approaches like differential privacy and Secure aggregation are taken. The paper addresses challenges like time zone differences, connectivity issues, interrupted execution etc,. Their work is mature enough to deploy the system in production for tens of millions of devices. They are working towards supporting billions of devices now. The training protocol The system involves devices and the Federated Learning server communicating availability and the server selecting devices to run a task. A subset of the available devices are selected for a task. The Federated Learning server instructs the devices what computing task to run with a plan. A plan would consist a TensorFlow graph and instructions to execute it. There are three phases for the training to take place: Selection of the devices that meet eligibility criteria Configuring the server with simple or Secure Aggregation Reporting from the devices where reaching a certain number would get the training round started Source: Towards Federated Learning at Scale: System Design The devices are supposed to maintain a repository of the collected data and the applications are responsible to provide data to the Federated Learning runtime as an example store. The Federated Learning server is designed to operate on orders of many magnitudes. Each round can mean updates from devices in the range of KBs to tens of MBs coming going the server. Data collection To avoid harming the phone’s battery life and performance, various analytics are collected in the cloud. The logs don’t contain any personally identifiable information. Secure aggregation Secure aggregation uses encryption to make individual device updates uninspectable. They plant to use it for protection against threats in data centers. Secure aggregation would ensure data encryption even when it is in-memory. Challenges of federated learning Compared to a centralized dataset, federated learning poses a number of challenges. The training data is not inspectable, tooling is required to work with proxy data. Models cannot be run interactively and must be compiled to be deployed in the Federated Learning server. Model resource consumption and runtime compatibility also come into the picture when working with many devices in real-time. Applications of Federated Learning It is best for cases where the data on devices is more relevant than data on servers. Ranking items for better navigation, suggestions for on-device keyboard, and next word prediction. This has already been implemented on Google pixel and Gboard. Future work is to eliminate bias caused be restrictions in device selection, algorithms to support better parallelism (more devices in one round), avoiding retraining already trained tasks on devices, and compression to save bandwidth. Federated computation, not federated learning The authors do no mention machine learning explicitly anywhere in the paper. They believe that the applications of such a model are not limited to machine learning. Federated Computation is the term they want to use for this concept. Federated computation and edge computing Federated learning and edge computing are very similar, there are but subtle differences in the purpose of these two. Federated learning is used to solve problems with specific tasks assigned to endpoint smartphones. Edge computing is for predefined tasks to be processed at end nodes, for example, IoT cameras. Federated learning decentralizes the data used while edge computing decentralizes the task computation to various devices. For more details on the architecture and its working, you can check out the research paper. Technical and hidden debts in machine learning – Google engineers’ give their perspective Researchers introduce a machine learning model where the learning cannot be proved What if AIs could collaborate using human-like values? DeepMind researchers propose a Hanabi platform.

Facebook is playing a huge role in artificial intelligence research. It’s not only a core part of the Facebook platform, it’s central to how the organization works. The company launched its AI research lab - FAIR - back in 2013. Today, led by some of the best minds in the field, it's not only helping Facebook to leverage artificial intelligence, it's also making it more accessible to researchers and engineers around the world. Let’s take a look at some of the tools built by Facebook that are doing just that. PyTorch: Facebook's leading artificial intelligence tool PyTorch is a hugely popular deep learning framework (rivalling Google's TensorFlow) that, by combining flexiblity and dynamism with stability, bridges the gap between research and production. Using a tape-based auto-differentiation system, PyTorch can be modified and changed by engineers without losing speed. That’s good news for everyone. Although PyTorch steals the headlines, there are a range of supporting tools that are making artificial intelligence and deep learning more accessible and achievable for other engineers. Read next: Is PyTorch better than Google’s TensorFlow? Find PyTorch eBooks and videos on the Packt website.  Facebook's computer vision tools Another field that Facebook has revolutionized is computer vision and image processing. Detectron, Facebook’s state-of-the-art object detection software system, has powered many research projects including Mask R-CNN - a simple and flexible way of developing Convolution Neural Networks for image processing. Mask R-CNN has also helped to power DensePose, a tool that map all human pixels of an RGB image to a 3D surface-based representation of the human body. Facebook has also heavily contributed to research in detecting and recognizing Human-Object interactions as well. Their contribution to the field of generative modeling is equally very important, with tasks such as minimizing variations in the quality of images, JPEG compression as well as image quantization now becoming easier and more accessible. Facebook, language and artificial intelligence We share updates, we send messages - language is a cornerstone of Facebook. This is why it's such an important area for Facebook’s AI researchers. There are a whole host of libraries and tools that are built for language problems. FastText is a library for text representation and classification, while ParlAI is a platform pushing the boundaries of dialog research. The platform is focused on tackling 5 key AI tasks: question answering, sentence completion, goal-oriented dialog, chit-chat dialog, and visual dialog. The ultimate aim for ParlAI is to develop a general dialog AI. There are also a few more language tools in Facebook’s AI toolkit - Fairseq and Translate are helping with translation and text generation, while Wav2Letter is an Automatic Speech Recognition system that can be used for transcription tasks. Rational artificial intelligence for gaming and smart decision making Although Facebook isn’t known for gaming, its interest in developing artificial intelligence that can reason could have an impact on the way games are built in the future. ELF is a tool developed by Facebook that allows game developers to train and test AI algorithms in a gaming environment. ELF was used by Facebook researchers to recreate DeepMind’s AlphaGo Zero, the AI bot that has defeated Go champions. Running on a single GPU, the ELF OpenGo bot defeated four professional Go players 14-0. Impressive, right? There are other tools built by Facebook that aim to build AI into game reasoning. Torchcraft is probably the most notable example - its a library that’s making AI research on Starcraft - a strategy game - accessible to game developers and AI specialists alike. Facebook is defining the future of artificial intelligence As you can see, Facebook is doing a lot to push the boundaries of artificial intelligence. However, it’s not just keeping these tools for itself - all these tools are open source, which means they can be used by anyone.

Yee Whye Teh is a professor at the department of Statistics of the University of Oxford and also a research scientist at DeepMind. He works on statistical machine learning, focussing on Bayesian nonparametrics, probabilistic learning, and deep learning. The motive of this article aims to bring our readers to Yee’s keynote speech at the NIPS 2017. Yee’s keynote ponders deeply on the interface between two perspectives on machine learning: Bayesian learning and Deep learning by exploring questions like: How can probabilistic thinking help us understand deep learning methods or lead us to interesting new methods? Conversely, how can deep learning technologies help us develop advanced probabilistic methods? For a more comprehensive and in-depth understanding of this novel approach, be sure to watch the complete keynote address by Yee Whye Teh on  NIPS facebook page. All images in this article come from Yee’s presentation slides and do not belong to us. The history of machine learning has shown a growth in both model complexity and in model flexibility. The theory led models have started to lose their shine. This is because machine learning is at the forefront of a revolution that could be called as data led models or the data revolution. As opposed to theory led models, data-led models try not to impose too many assumptions on the processes that have to be modeled and are rather superflexible non-parametric models that can capture the complexities but they require large amount of data to operate.   On the model flexibility side, we have various approaches that have been explored over the years. We have kernel methods, Gaussian processes, Bayesian nonparametrics and now we have deep learning as well. The community has also developed evermore complex frameworks both graphical and programmatic to compose large complex models from simpler building blocks. In the 90’s we had graphical models, later we had probabilistic programming systems, followed by deep learning systems like TensorFlow, Theano, and Torch. A recent addition is probabilistic Torch, which brings together ideas from both the probabilistic Bayesian learning and deep learning. On one hand we have Bayesian learning, which deals with learning as inference in some probabilistic models. On the other hand we have deep learning models, which view learning as optimization functions parametrized by neural networks. In recent years there has been an explosion of exciting research at this interface of these two popular approaches resulting in increasingly complex and exciting models. What is Bayesian theory of learning Bayesian learning describes an ideal learner as one who interacts with the world in order to know its state, which is given by θ. He/she makes some observations about the world by deducing a model in Bayesian context. This model is a joint distribution of both the unknown state of the world θ and the observation about the world x. The model consists of prior distribution and marginal distribution, combining which gives a reverse conditional distribution also known as posterior, which describes the totality of the agent's knowledge about the world after he/she sees x. This posterior can also be used for predicting future observations and act accordingly. Issues associated with Bayesian learning Rigidity Learning can be wrong if model is wrong Not all prior knowledge can be encoded as joint distribution Simple analytic forms are limiting for conditional distributions 2. Scalability: Intractable to compute this posterior and approximations have to be made, which then introduces trade offs between efficiency and accuracy. As a result, it is often assumed that Bayesian techniques are not scalable. To address these issues, the speaker highlights some of his recent projects which showcase scenarios where deep learning ideas are applied to Bayesian models (Deep Bayesian learning) or in the reverse applying Bayesian ideas to Neural Networks ( i.e. Bayesian Deep learning) Deep Bayesian learning: Deep learning assists Bayesian learning Deep learning can improve Bayesian learning in the following ways: Improve the modeling flexibility by using neural networks in the construction of Bayesian models Improve the inference and scalability of these methods by parameterizing the posterior way of using neural networks Empathizing inference over multiple runs These can be seen in the following projects showcased by Yee: Concrete VAEs(Variational Autoencoders) FIVO: Filtered Variational Objectives Concrete VAEs What are VAEs? All the qualities mentioned above, i.e. improving modeling flexibility, improving inference and scalability, and empathizing inference over multiple runs by using neural networks can be seen in a class of deep generative models known as VAE (Variational Autoencoders). Fig: Variational Autoencoders VAEs include latent variables that describe the contents of a scene i.e objects, pose. The relationship between these latent variables and the pixels have to be highly complex and nonlinear. So, in short, VAEs are used to parameterize generative and variable posterior distribution that allows for greater scope flexible modeling. The key that makes VAEs work is the reparameterization trick Fig: Adding reparameterization to VAEs The reparameterization trick is crucial to the continuous latent variables in the VAEs. But many models naturally include discrete latent variables. Yee suggests application of the reparameterization on the discrete latent variables as a work around. This brings us to the concept of Concrete VAEs.. CONtinuous relaxation of disCRETE distributions.Also, the density can be further calculated: This concrete distribution is the reparameterization trick for discrete variables which helps in calculating the KL divergence that is needed for variational inference. FIVO: Filtered Variational Objectives FIVO extends VAEs towards models for sequential and time series data. It is built upon another extension of VAEs known as Importance Weighted Autoencoder, a generative model with a similar as that of the VAE, but which uses a strictly tighter log-likelihood lower bound. Variational lower bound: Rederivation from importance sampling: Better to use multiple samples: Using Importance Weighted Autoencoders we can use multiple sampling, with which we can get a tighter lower bound and optimizing this lower bound should lead to better learning. Let’s have a look at the FIVO objectives: We can use any unbiased estimator p(X) of marginal probabilityTightness of bound related to variance of estimatorFor sequential models, we can use particle filters which produce unbiased estimator of marginal probability. They can also have much lower variance than importance samplers. Bayesian Deep learning: Bayesian approach for deep learning gives us counterintuitive and surprising ways to make deep learning scalable. In order to explore the potential of Bayesian learning with deep neural networks, Yee introduced a project named, The posterior server. The Posterior server The posterior server is a distributed server for deep learning. It makes use of the Bayesian approach in order to make neural networks highly scalable. This project focuses on Distributed learning, where both the data and the computations can be spread across the network. The figure above shows that there are a bunch of workers and each communicates with the parameter server, which effectively maintains the authoritative copy of the parameters of the network. At each iteration, each worker obtains the latest copy of the parameter from the server, computes the gradient update based on its data and sends it back to the server which then updates it to the authoritative copy. So, communications on the network tend to be slower than the computations that can be done on the network. Hence, one might consider multiple gradient steps on each iteration before it sends the accumulated update back to the parameter server. The problem is that the parameter and the worker quickly get out of sync with the authoritative copy on the parameter server. As a result, this leads to stale updates which allow noise into the system and we often need frequent synchronizations across the network for the algorithm to learn in a stable fashion. The main idea here in Bayesian context is that we don't just want a single parameter, we want a whole distribution over them. This will then relax the need for frequent synchronizations across the network and hopefully lead to algorithms that are robust to last frequent communication. Each worker is simply going to construct its own tractable approximation to his own likelihood function and send this information to the posterior server which then combines these approximations together to form the full posterior or an approximation of it. Further, the approximations that are constructed would be based on the statistics of some sampling algorithms that happens locally on that worker. The actual algorithm includes a combination of the variational algorithms, Stochastic Gradient EP and the Markov chain Monte Carlo on the workers themselves. So the variational part in the algorithm handles the communication part in the network whereas the MCMC part handles the sampling part that is posterior to construct the statistics that the variational part needs. For scalability, a stochastic gradient Langevin algorithm which is a simple generalization of the SGT, which includes additional injected noise, to sample from posterior noise. To experiment with this server, it was trained densely connected neural networks with 500 reLU units on MNIST dataset. You can have a detailed understanding of these examples in the keynote video. This interface between Bayesian learning and deep learning is a very exciting frontier. Researchers have brought management of uncertainties within deep learning. Also, flexibility and scalability in Bayesian modeling. Yee concludes with two questions for the audience to think about. Does being Bayesian in the space of functions makes more sense than being Bayesian in the sense of parameters? How to deal with uncertainties under model misspecification?    

One of the most awaited machine learning conference, NeurIPS 2018 is happening throughout this week in Montreal, Canada. It will feature a series of tutorials, invited talks, product releases, demonstrations, presentations, and announcements related to machine learning research. For the first time, NeurIPS invited a diversity and inclusion (D&I) speaker Laura Gomez to talk about the lack of diversity in the tech industry, which leads to biased algorithms, faulty products, and unethical tech. Laura Gomez is the CEO of Atipica that helps tech companies find and hire diverse candidates. Being a Latina woman herself, she had to face oppression when seeking capital and funds for her startup trying to establish herself in Silicon Valley. This experience led to her realization that there is a strong need to talk about why diversity and inclusion matters. Her efforts were not in vain and recently, she raised $2M in seed funding led by True Ventures. “At Atipica, we think of Inclusive AI in terms of data science, algorithms, and their ethical implications. This way you can rest assure our models are not replicating the biases of humans that hinder diversity while getting patent-pending aggregate demographic insights of your talent pool,” reads the website. She talks about her journey as a Latina woman in the tech industry. She reminisced on how she was the only one like her who got an internship with Hewlett Packard and the fact that she hated it. Nevertheless, she still decided to stay, determined not to let the industry turn her into a victim. She believes she made the right choice going forward with tech; now, years later, diversity is dominating the conversation in the industry. After HP, she also worked at Twitter and YouTube, helping them translate and localize their applications for a global audience. She is also a founding advisor of Project Include, which is a non-profit organization run by women, that uses data and advocacy to accelerate diversity and inclusion solutions in the tech industry. She opened her talk by agreeing to a quote from Safiya Noble, who wrote Algorithms of Oppression. “Artificial Intelligence will become a major human rights issue in the twenty-first century.” She believes we need to talk about difficult questions such as where AI is heading? And where should we hold ourselves and each other accountable.” She urges people to evaluate their role in AI, bias, and inclusion, to find the empathy and value in difficult conversations, and to go beyond your immediate surroundings to consider the broader consequences. It is important to build accountable AI in a way that allows humanity to triumph. She touched upon discriminatory moves by tech giants like Amazon and Google. Amazon recently killed off its AI recruitment tool because it couldn’t stop discriminating against women. She also criticized upon Facebook’s Myanmar operation where Facebook data scientists were building algorithms for hate speech. They didn’t understand the importance of localization or language or actually internationalize their own algorithms to be inclusive towards all the countries. She also talked about algorithmic bias in library discovery systems, as well as how even ‘black robots’ are being impacted by racism. She also condemned Palmer Luckey's work who is helping U.S. immigration agents on the border wall identify Latin refugees. Finally, she urged people to take three major steps to progress towards being inclusive: Be an ally Think of inclusion as an approach, not a feature Work towards an Ethical AI Head over to NeurIPS facebook page for the entire talk and other sessions happening at the conference this week. NeurIPS 2018: Deep learning experts discuss how to build adversarially robust machine learning models NeurIPS 2018 paper: DeepMind researchers explore autoregressive discrete autoencoders (ADAs) to model music in raw audio at scale NeurIPS 2018: A quick look at data visualization for Machine learning by Google PAIR researchers [Tutorial]

This year the Neural Information Processing Systems Conference is asking serious questions to improve diversity, equity, and inclusion at NeurIPS. “Our goal is to make the conference as welcoming as possible to all.” said the heads of the new diversity and inclusion chairs introduced this year. https://twitter.com/InclusionInML/status/1069987079285809152 The Diversity and Inclusion chairs were headed by Hal Daume III, a professor from the University of Maryland and machine learning and fairness groups researcher at Microsoft Research and Katherine Heller, assistant professor at Duke University and research scientist at Google Brain. They opened up the talk by acknowledging the respective privilege that they get as a group of white man and woman and the fact that they don’t reflect the diversity of experience in the conference room, much less the world. They talk about the three major goals with respect to inclusion at NeurIPS: Learn about the challenges that their colleagues have faced. Support those doing the hard work of amplifying the voices of those who have been historically excluded. To begin structural changes that will positively impact the community over the coming years. They urged attendees to start building an environment where everyone can do their best work. They want people to: see other perspectives remember the feeling of being an outsider listen, do research and learn. make an effort and speak up Concrete actions taken by the NeurIPS diversity and inclusion chairs This year they have assembled an advisory board and run a demographics and inclusion survey. They have also conducted events such as WIML (Women in Machine Learning), Black in AI, LatinX in AI, and Queer in AI. They have established childcare subsidies and other activities in collaboration with Google and DeepMind to support all families attending NeurIPS by offering a stipend of up to $100 USD per day. They have revised their Code of Conduct, to provide an experience for all participants that is free from harassment, bullying, discrimination, and retaliation. They have added inclusion tips on Twitter offering tips and bits of advice related to D&I efforts. The conference also offers pronoun stickers (only them and they), first-time attendee stickers, and information for participant needs. They have also made significant infrastructure improvements for visa handling. They had discussions with people handling visas on location, sent out early invitation letters for visas, and are choosing future locations with visa processing in mind. In the future, they are also looking to establish a legal team for details around Code of Conduct. Further, they are looking to improve institutional structural changes that support the community, and improve the coordination around affinity groups & workshops. For the first time, NeurIPS also invited a diversity and inclusion (D&I) speaker Laura Gomez to talk about the lack of diversity in the tech industry, which leads to biased algorithms, faulty products, and unethical tech. Head over to NeurIPS website for interesting tutorials, invited talks, product releases, demonstrations, presentations, and announcements. NeurIPS 2018: Deep learning experts discuss how to build adversarially robust machine learning models NeurIPS 2018 paper: DeepMind researchers explore autoregressive discrete autoencoders (ADAs) to model music in raw audio at scale NeurIPS 2018: A quick look at data visualization for Machine learning by Google PAIR researchers [Tutorial]

TensorFlow 2.0 was made available in October. One of the major highlights of this release was the integration of Keras into TensorFlow. Keras is an open-source deep-learning library that is designed to enable fast, user-friendly experimentation with deep neural networks. It serves as an interface to several deep learning libraries, most popular of which is TensorFlow, and it was integrated into TensorFlow main codebase in TensorFlow 2.0. In September, Lex Fridman, Research scientist at MIT popularly known for his podcasts, spoke to François Chollet, who is the author of Keras on Keras, Deep Learning, and the Progress of AI. In this post, we have tried to highlight François’ views on the Keras and TensorFlow 2.0 integration, early days of Keras and the importance of design decisions for building deep learning models. We recommend the full podcast that’s available on Fridman’s YouTube channel. Want to build Neural Networks? [box type="shadow" align="" class="" width=""]If you want to build multiple neural network architectures such as CNN, RNN, LSTM in Keras, we recommend you to read Neural Networks with Keras Cookbook by V Kishore Ayyadevara. This book features over 70 recipes such as object detection and classification, building self-driving car applications, understanding data encoding for image, text and recommender systems and more. [/box] Early days of Keras and how it was integrated into TensorFlow I started working on Keras in 2015, says Chollet. At that time Caffe was the popular deep learning library, based on C++ and was popular for building Computer Vision projects. Chollet was interested in Recurrent Neural Networks (RNNs) which was a niche topic at that time. Back then, there was no good solution or reusable open-source implementation of RNNs and LSTMs, so he decided to build his own and that’s how Keras started. “It was going to be mostly around RNNs and LSTMs and the models would be defined by Python code, which was going against mainstream,” he adds. Later, he joined Google’s research team working on image classification. At that time, he was exposed to the early internal version of Tensorflow - which was an improved version of Theano. When Tensorflow was released in 2015, he refactored Keras to run on TensorFlow. Basically he was abstracting away all the backend functionality into one module so that the same codebase could run on top of multiple backends. A year later, the TensorFlow team requested him to integrate the Keras API into TensorFlow more tightly.  They build a temporary TensorFlow-only version of Keras that was in tf.contrib for a while. Then they finally moved to TensorFlow Core in 2017. TensorFlow 2.0 gives both usability and flexibility to Keras Keras has been a very easy-to-use high-level interface to do deep learning. However, it lacked in flexibility - Keras framework was not the optimal way to do things compared to just writing everything from scratch. TensorFlow 2.0 offers both usability and flexibility to Keras. You have the usability of the high-level interface but you have the flexibility of the lower-level interface. You have this spectrum of workflows where you can get more or less usability and flexibility,  the trade-offs depending on your needs. It's very flexible, easy to debug, and powerful but also integrates seamlessly with higher-level features up to classic Keras workflows. “You have the same framework offering the same set of APIs that enable a spectrum of workflows that are more or less high level and are suitable for you know profiles ranging from researchers to data scientists and everything in between,” says Chollet. Design decisions are especially important while integrating Keras with Tensorflow “Making design decisions is as important as writing code”, claims Chollet. A lot of thought and care is taken in coming up with these decisions, taking into account the diverse user base of TensorFlow - small-scale production users, large-scale production users, startups, and researchers. Chollet says, “A lot of the time I spend on Google is actually discussing design. This includes writing design Docs, participating in design review meetings, etc.” Making a design decision is about satisfying a set of constraints but also trying to do so in the simplest way possible because this is what can be maintained and expanded in the future. You want to design APIs that are modular and hierarchical so that they have an API surface that is as small as possible. You want this modular hierarchical architecture to reflect the way that domain experts think about the problem. On the future of Keras and TensorFlow. What’s going to happen in TensorFlow 3.0? Chollet says that he’s really excited about developing even higher-level APIs with Keras. He’s also excited about hyperparameter tuning by automated machine learning. He adds, “The future is not just, you know, defining a model, it's more like an automatic model.” Limits of deep learning wrt function approximators that try to generalize from data Chollet emphasizes that “Neural Networks don't generalize well, humans do.” Deep Learning models are like huge parametric and differentiable models that go from an input space to an output space, trained with gradient descent. They are learning a continuous geometric morphing from an input vector space to an output space. As this is done point by point; a deep neural network can only make sense of points in space that are very close to things that it has already seen in string data. At best it can do the interpolation across points. However, that means in order to train your network you need a dense sampling of the input, almost a point-by-point sampling which can be very expensive if you're dealing with complex real-world problems like autonomous driving or robotics.  In contrast to this, you can look at very simple rules algorithms. If you have a symbolic rule it can actually apply to a very large set of inputs because it is abstract, it is not obtained by doing a point by point mapping. Deep learning is really like point by point geometric morphings. Meanwhile, abstract rules can generalize much better. I think the future is which can combine the two. Chollet also talks about self-improving Artificial General Intelligence, concerns about short-term and long-term threats in AI, Program synthesis, Good test for intelligence and more. The full podcast is available on Lex’s YouTube channel. If you want to implement neural network architectures in Keras for varied real-world applications, you may go through our book Neural Networks with Keras Cookbook. TensorFlow.js contributor Kai Sasaki on how TensorFlow.js eases web-based machine learning application development 10 key announcements from Microsoft Ignite 2019 you should know about What does a data science team look like?

Key takeaways from the discussion To solve problems with machine learning, you must first understand them. Different people or groups of people are going to define a problem in a different way. So, we shouldn't believe that the way we want to frame the problem computationally is the right way. If we allow that our systems include people and society, it is clear that we have to help negotiate values, not simply define them. Last week, at the 32nd NeurIPS 2018 annual conference, Nitin Koli, Joshua Kroll, and Deirdre Mulligan presented the common pitfalls we see when studying the human side of machine learning. Machine learning is being used in high-impact areas like medicine, criminal justice, employment, and education for making decisions. In recent years, we have seen that this use of machine learning and algorithmic decision making have resulted in unintended discrimination.  It’s becoming clear that even models developed with the best of intentions may exhibit discriminatory biases and perpetuate inequality. Although researchers have been analyzing how to put concepts like fairness, accountability, transparency, explanation, and interpretability into practice in machine learning, properly defining these things can prove a challenge. Attempts have been made to define them mathematically, but this can bring new problems. This is because applying mathematical logic to human concepts that have unique and contested political and social dimensions necessarily has blind spots - every point of contestation can’t be integrated into a single formula. In turn, this can cause friction with other disciplines as well as the public. Based on their research on what various terms mean in different contexts, Nitin Koli, Joshua Krill, and Deirdre Mulligan drew out some of the most common misconceptions machine learning researchers and practitioners hold. Sociotechnical problems To find a solution to a particular problem, data scientists need precise definitions. But how can we verify that these definitions are correct? Indeed, many definitions will be contested, depending on who you are and what you want them to mean. A definition that is fair to you will not necessarily be fair to me”, remarks Mr. Kroll. Mr. Kroll explained that while definitions can be unhelpful, they are nevertheless essential from a mathematical perspective.  This means there appears to be an unresolved conflict between concepts and mathematical rigor. But there might be a way forward. Perhaps it’s wrong to simply think in this dichotomy of logical rigor v. the messy reality of human concepts. One of the ways out of this impasse is to get beyond this dichotomy. Although it’s tempting to think of the technical and mathematical dimension on one side, with the social and political aspect on the other, we should instead see them as intricately related. They are, Kroll suggests, socio-technical problems. Kroll goes on to say that we cannot ignore the social consequences of machine learning: “Technologies don’t live in a vacuum and if we pretend that they do we kind of have put our blinders on and decided to ignore any human problems.” Fairness in machine learning In the real world, fairness is a concept directly linked to processes. Think, for example, of the voting system. Citizens cast votes to their preferred candidates and the candidate who receives the most support is elected. Here, we can say that even though the winning candidate was not the one a candidate voted for, but at least he/she got the chance to participate in the process. This type of fairness is called procedural fairness. However, in the technical world, fairness is often viewed in a subtly different way. When you place it in a mathematical context, fairness centers on outcome rather than process. Kohli highlighted that trade offs between these different concepts can’t be avoided. They’re inevitable. A mathematical definition of fairness places a constraint over the behavior of a system, and this constraint will narrow down the cause of models that can satisfy these conditions. So, if we decide to add too many fairness constraints to the system, some of them will be self-contradictory. One more important point machine learning practitioners should keep in mind is that when we talk about the fairness of a system, that system isn’t a self-contained and coherent thing. It is not a logical construct - it’s a social one. This means there are a whole host of values, ideas, and histories that have an impact on its reality.. In practice, this ultimately means that the complexity of the real world from which we draw and analyze data can have an impact on how a model works. Kohli explained this by saying, “it doesn’t really matter... whether you are building a fair system if the context in which it is developed and deployed in is fundamentally unfair.” Accountability in machine learning Accountability is ultimately about trust. It’s about the extent you can be sure you know what is ‘true’ about a system. It refers to the fact that you know how it works and why it does things in certain ways. In more practical terms, it’s all about invariance and reliability. To ensure accountability inside machine learning models, we need to follow a layered model. The bottom layer is an accounting or recording layer, that keeps track of what a given system is doing and the ways in which it might have been changed.. The next layer is a more analytical layer. This is where those records on the bottom layer are analyzed, with decisions made about performance - whether anything needs to be changed and how they should be changed. The final and top-most layer is about responsibility. It’s where the proverbial buck stops - with those outside of the algorithm, those involved in its construction. “Algorithms are not responsible, somebody is responsible for the algorithm,”  explains Kroll. Transparency Transparency is a concept heavily tied up with accountability. Arguably you have no accountability without transparency. The layered approach discussed above should help with transparency, but it’s also important to remember that transparency is about much more than simply making data and code available. Instead, it demands that the decisions made in the development of the system are made available and clear too. Mr. Kroll emphasizes, “to the person at the ground-level for whom the decisions are being taken by some sort of model, these technical disclosures aren’t really useful or understandable.” Explainability In his paper Explanation in Artificial Intelligence: Insights from the Social Sciences, Tim Miller describes what is explainable artificial intelligence. According to Miller, explanation takes many forms such as causal, contrastive, selective, and social. Causal explanation gives reasons behind why something happened, for example, while contrastive explanations can provide answers to questions like“Why P rather than not-P?". But the most important point here is that explanations are selective. An explanation cannot include all reasons why something happened; explanations are always context-specific, a response to a particular need or situation. Think of it this way: if someone asks you why the toaster isn’t working, you could just say that it’s broken. That might be satisfactory in some situations, but you could, of course, offer a more substantial explanation, outlining what was technically wrong with the toaster, how that technical fault came to be there, how the manufacturing process allowed that to happen, how the business would allow that manufacturing process to make that mistake… you could, of course, go on and on. Data is not the truth Today, there is a huge range of datasets available that will help you develop different machine learning models. These models can be useful, but it’s essential to remember that they are models. A model isn’t the truth - it’s an abstraction, a representation of the world in a very specific way. One way of taking this fact into account is the concept of ‘construct validity’. This sounds complicated, but all it really refers to is the extent to which a test - say a machine learning algorithm - actually measures what it says it’s trying to measure. The concept is widely used in disciplines like psychology, but in machine learning, it simply refers to the way we validate a model based on its historical predictive accuracy. In a nutshell, it’s important to remember that just as data is an abstraction of the world, models are also an abstraction of the data. There’s no way of changing this, but having an awareness that we’re dealing in abstractions ensures that we do not lapse into the mistake of thinking we are in the realm of ‘truth’. To build a fair(er) systems will ultimately require an interdisciplinary approach, involving domain experts working in a variety of fields. If machine learning and artificial intelligence is to make a valuable and positive impact in fields such as justice, education, and medicine, it’s vital that those working in those fields work closely with those with expertise in algorithms. This won’t fix everything, but it will be a more robust foundation from which we can begin to move forward. You can watch the full talk on the Facebook page of NeurIPS. Researchers unveil a new algorithm that allows analyzing high-dimensional data sets more effectively, at NeurIPS conference Accountability and algorithmic bias: Why diversity and inclusion matters [NeurIPS Invited Talk] NeurIPS 2018: A quick look at data visualization for Machine learning by Google PAIR researchers [Tutorial]