Your article (Hi-tech brain drain threatens British university research, 2 November) contains one particularly chilling revelation: that Google DeepMind now runs artificial intelligence courses at UCL and Oxford. Having met the DeepMind people in my role with the MIT Media Lab, I know that their definition of "intelligence" is so impoverished that it doesn't extend beyond the abstract calculations that an algorithm can achieve, and completely fails to understand that human intelligence is embodied and distributed throughout our physical selves – and indeed between them, in the mirror neurons that fire in sympathy when we watch a dancer or help an injured friend. Artificial intelligence of the kind Google promotes can play Go and even – at a pinch – recognise Bach or Picasso. It can never produce Bach or Picasso, still less understand the complexity of social forms and culture that made their lives possible. If we entrust the education of those who will determine the future relationship of people and machines to a company whose core belief is that all human experience can be replicated by algorithms, all we can hope is that global warming wipes us out before the machines do.

A possibly lethal exception could be the error that leads a self-driving car's AI to make the wrong decision at the wrong time. That is why researchers developed a bug-hunting method that can systematically expose bad decision-making by the deep learning algorithms deployed in online services and autonomous vehicles. The new DeepXplore method uses at least three neural networks--the basic architecture of deep learning algorithms--to act as "cross-referencing oracles" in checking each other's accuracy. Researchers at Columbia University and Lehigh University designed DeepXplore to solve an optimization problem in which they looked to strike the best balance between two objectives: maximizing the number of neurons activated within neural networks, and triggering as many conflicting decisions as possible among different neural networks. By assuming that the majority of neural networks will generally make the right decision, DeepXplore automatically retrains the neural network that made the lone dissenting decision to follow the example of the majority in a given scenario.

Farfield speech recognition has become a popular research area in the past few years, from more research focused activities such as the CHiME Challenges, to the launches of Amazon Echo and Google Home. This talk will describe the research efforts around Google Home. Most multichannel ASR systems commonly separate speech enhancement, including localization, beamforming and postfiltering, from acoustic modeling. In this talk, we will introduce a framework to do multichannel enhancement jointly with acoustic modeling using deep neural networks. Inspired by beamforming, which leverages differences in the fine time structure of the signal at different microphones to filter energy arriving from different directions, we explore modeling the raw time-domain waveform directly.

Jon Bruner oversees O'Reilly's publications on hardware, the Internet of Things, manufacturing, and electronics, and has been program chair along with Joi Ito of the O'Reilly Solid conference, focused on the intersection between software and the physical world. Before coming to O'Reilly, he was data editor at Forbes Magazine, where he combined writing and programming to approach a broad variety of subjects, from the operation of the Columbia River's dams to migration within the United States. He studied mathematics and economics at the Univers..

Computers can now beat humans at chess and Go, but it may be a while before people trust their driving. The danger of self-driving cars was highlighted last year when Tesla's autonomous car collided with a truck it mistook for a cloud, killing its passenger. Self-driving cars depend on a form of machine learning called deep learning. Modeled after the human brain, layers of artificial neurons process and consolidate information, developing a set of rules to solve complex problems, from recognizing friends' faces online to translating email written in Chinese. The technology has achieved impressive feats of intelligence, but as more tasks become automated this way, concerns about safety, security, and ethics, are growing.

New software developed at the University of Waterloo could make it easier to adopt and trust powerful artificial intelligence (AI) systems that generate stock market predictions, assess who qualifies for mortgages and set insurance premiums. The software is designed to analyze and explain decisions made by deep-learning AI algorithms, providing key insights needed to satisfy regulatory authorities and give analysts confidence in their recommendations. "The potential impact, especially in regulatory settings, is massive," said Devinder Kumar, lead researcher and a PhD candidate in systems design engineering at Waterloo. "If you can't provide reasons for their decisions, you can't use those state-of-the-art systems right now." Deep-learning AI algorithms essentially teach themselves by processing and detecting patterns in vast quantities of data. As a result, even their creators don't know why they come to their conclusions.

We interpret HyperNetworks within the framework of variational inference within implicit distributions. Our method, Bayes by Hypernet, is able to model a richer variational distribution than previous methods. Experiments show that it achieves comparable predictive performance on the MNIST classification task while providing higher predictive uncertainties compared to MC-Dropout and regular maximum likelihood training.

Generative Adversarial Networks (GANs) are powerful models for learning complex distributions. Stable training of GANs has been addressed in many recent works which explore different metrics between distributions. In this paper we introduce Fisher GAN which fits within the Integral Probability Metrics (IPM) framework for training GANs. Fisher GAN defines a critic with a data dependent constraint on its second order moments. We show in this paper that Fisher GAN allows for stable and time efficient training that does not compromise the capacity of the critic, and does not need data independent constraints such as weight clipping. We analyze our Fisher IPM theoretically and provide an algorithm based on Augmented Lagrangian for Fisher GAN.

Modern neural networks are often augmented with an attention mechanism, which tells the network where to focus within the input. We propose in this paper a new framework for sparse and structured attention, building upon a smoothed max operator. We show that the gradient of this operator defines a mapping from real values to probabilities, suitable as an attention mechanism. Our framework includes softmax and a slight generalization of the recently-proposed sparsemax as special cases. However, we also show how our framework can incorporate modern structured penalties, resulting in more interpretable attention mechanisms, that focus on entire segments or groups of an input. We derive efficient algorithms to compute the forward and backward passes of our attention mechanisms, enabling their use in a neural network trained with backpropagation. To showcase their potential as a drop-in replacement for existing ones, we evaluate our attention mechanisms on three large-scale tasks: textual entailment, machine translation, and sentence summarization. Our attention mechanisms improve interpretability without sacrificing performance; notably, on textual entailment and summarization, we outperform the standard attention mechanisms based on softmax and sparsemax.

Deep neural networks (NNs) are powerful black box predictors that have recently achieved impressive performance on a wide spectrum of tasks. Quantifying predictive uncertainty in NNs is a challenging and yet unsolved problem. Bayesian NNs, which learn a distribution over weights, are currently the state-of-the-art for estimating predictive uncertainty; however these require significant modifications to the training procedure and are computationally expensive compared to standard (non-Bayesian) NNs. We propose an alternative to Bayesian NNs that is simple to implement, readily parallelizable, requires very little hyperparameter tuning, and yields high quality predictive uncertainty estimates. Through a series of experiments on classification and regression benchmarks, we demonstrate that our method produces well-calibrated uncertainty estimates which are as good or better than approximate Bayesian NNs. To assess robustness to dataset shift, we evaluate the predictive uncertainty on test examples from known and unknown distributions, and show that our method is able to express higher uncertainty on out-of-distribution examples. We demonstrate the scalability of our method by evaluating predictive uncertainty estimates on ImageNet.