Obtaining enough labeled data to robustly train complex discriminative models is a major bottleneck in the machine learning pipeline. A popular solution is combining multiple sources of weak supervision using generative models. The structure of these models affects the quality of the training labels, but is difficult to learn without any ground truth labels. We instead rely on weak supervision sources having some structure by virtue of being encoded programmatically. We present Coral, a paradigm that infers generative model structure by statically analyzing the code for these heuristics, thus significantly reducing the amount of data required to learn structure. We prove that Coral's sample complexity scales quasilinearly with the number of heuristics and number of relations identified, improving over the standard sample complexity, which is exponential in n for learning n-th degree relations. Empirically, Coral matches or outperforms traditional structure learning approaches by up to 3.81 F1 points. Using Coral to model dependencies instead of assuming independence results in better performance than a fully supervised model by 3.07 accuracy points when heuristics are used to label radiology data without ground truth labels.

The adoption of automated, data-driven decision making in an ever expanding range of applications has raised concerns about its potential unfairness towards certain social groups. In this context, a number of recent studies have focused on defining, detecting, and removing unfairness from data-driven decision systems. However, the existing notions of fairness, based on parity (equality) in treatment or outcomes for different social groups, tend to be quite stringent, limiting the overall decision making accuracy. In this paper, we draw inspiration from the fair-division and envy-freeness literature in economics and game theory and propose preference-based notions of fairness -- given the choice between various sets of decision treatments or outcomes, any group of users would collectively prefer its treatment or outcomes, regardless of the (dis)parity as compared to the other groups. Then, we introduce tractable proxies to design margin-based classifiers that satisfy these preference-based notions of fairness. Finally, we experiment with a variety of synthetic and real-world datasets and show that preference-based fairness allows for greater decision accuracy than parity-based fairness.

This blogpost is a round up of the various sets of principles of robotics and AI that have been proposed to date, ordered by date of first publication. The principles are presented here (in full or abridged) with notes and references but without commentary. If there any (prominent) ones I've missed please let me know. Asimov's three laws of Robotics (1950) A robot may not injure a human being or, through inaction, allow a human being to come to harm. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law. A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws. I have included these to explicitly acknowledge, firstly, that Asimov undoubtedly established the principle that robots (and by extension AIs) should be governed by principles, and secondly that many subsequent principles have been drafted as a direct response. The three laws first appeared in Asimov's short story Runaround [1]. This wikipedia article provides a very good account of the three laws and their many (fictional) extensions. A robot must respond to humans as appropriate for their roles.

One of the most important issues that Congress will face in 2018 is how and when to regulate our growing dependence on artificial intelligence (AI). During the U.S. National Governors Association summer meetings, Elon Musk urged the group to push forward with regulation "before it's too late," stating that AI was an "existential threat to humanity." Hyperbole aside, there are legitimate concerns about the technology and its use. But a rush to regulation could exacerbate current issues, or create new issues that we're not prepared to deal with along the way. To begin with, one of the biggest issues in the world of AI is the lack of clear definition for what the technology is -- and is not.

A common theme among some of the most notable advances and new devices was the integration of artificial intelligence in smart and innovative ways. Despite a handful of flubs, AI-powered technologies still helped make the world a little smarter, kinder, and more innovative this year. Here are some of the moments when AI really shone in 2017. Earlier this month, NASA announced it used machine learning to discover two new planets. Researchers used old data from the Kepler space telescope to locate the two new additions to our galaxy. This wasn't the first time researchers applied AI to sift through the massive amount of data NASA's telescopes collect, but it is a promising example of how neural networks can find even some of the weakest signs of distant worlds. Thanks to AI, we have now discovered a planetary system that ties our solar system in the number of planets it has, which brings us one step closer to discovering more of the mysteries the giant void around us contains.

In the age of Artificial Intelligence (AI), data is power. And in a country like India, the scope for AI is colossal, especially in healthcare. From precision medicine and data management to remote diagnosis and predictive analysis, technology in healthcare is making giant strides. However, these innovations are also posing new-fangled security challenges, particularly in terms of personal health records. According to a recent global study, more than one-third (35.6 per cent) of surveyed professionals in the Internet of Things-connected medical device ecosystem said that they had experienced an incident related to cybersecurity in the past year.

WE ALL know how it ends: the machines rise up to enslave their puny masters. Robots and artificial intelligences may so far have confined themselves to blameless pursuits such as vacuum cleaning, beating us at board games and recommending products we might also like. But as they continue their inexorable rise, entering a "singularity" of runaway self-improvement, they will inevitably turn their attention to robopocalypse. Stephen Hawking says AI could spell the end for humanity. Elon Musk thinks it could lead to world war three. Vladimir Putin says whoever controls AI will control the world.

We are poised at the beginning of a new industrial revolution, this one powered by artificial intelligence (AI) and machine learning. The past several years have seen rapid advances in AI technology, driven in large part by deep neural networks. Machines have bested humans in a variety of games, including chess, Jeopardy, Go, and poker, and are now being applied to help solve a wide variety of practical problems, such as health care, finance, and transportation. AI is already having a significant impact on national security. Automation is used heavily in cyber security and defense applications.

I must not have got the memo, because as a young lecturer in computer science at the University of Southampton in 1985 I was unaware that "women didn't do computing". Southampton had always recruited a healthy number of women to study computing in our fledgling department, and a quarter of the staff were women, but the student lists for the new academic year showed that quite suddenly, or so it appeared, we'd achieved the unenviable record of having no female students in that year's intake. Many women made important contributions to computing in its early decades, figures such as Karen Spärck Jones in Britain or Grace Hopper in the US, among many others who worked in the vital field of cryptography during the Second World War or, later, on the enormous challenges of the space race. But it had become clear that by the mid-1980s something fundamental had changed. We found that UK university admission figures revealed that the number of girls studying computing had fallen dramatically compared to the number of boys: from 25% percent in 1978 to just 10% in 1985.

Deep neural networks are able to solve tasks across a variety of domains and modalities of data. Despite many empirical successes, we lack the ability to clearly understand and interpret the learned internal mechanisms that contribute to such effective behaviors or, more critically, failure modes. In this work, we present a general method for visualizing an arbitrary neural network's inner mechanisms and their power and limitations. Our dataset-centric method produces visualizations of how a trained network attends to components of its inputs. The computed "attention masks" support improved interpretability by highlighting which input attributes are critical in determining output. We demonstrate the effectiveness of our framework on a variety of deep neural network architectures in domains from computer vision, natural language processing, and reinforcement learning. The primary contribution of our approach is an interpretable visualization of attention that provides unique insights into the network's underlying decision-making process irrespective of the data modality.