When Royal Bank of Scotland introduced an online chatbot to the public in 2016, executives were keen to present the technology as an aid for hard-pressed call-centre agents. But it was clear the chatbot's powers would grow and perhaps, in time, could come to rival those of its human colleagues. Artificial intelligence, RBS said, lets the chatbot -- first named Luvo but now called Assist -- handle routine web chat requests from clients, such as queries about address changes or asking for a bank card to be authorised. This would free human advisers to help customers with more complex questions. The initial trial that began in December has been limited to 10 per cent of RBS customers in Scotland who use web chat to communicate with the bank.

Roman Mugriyev was driving his long-haul 18-wheeler down a two-lane Texas highway when he saw an oncoming car drift into his lane just a few hundred feet ahead. There was a ditch to his right and more oncoming cars to his left, so there was little for him to do but hit his horn and brake. "I could hear the man who taught me to drive telling me what he always said was rule number one: 'Don't hurt anybody,'" Mugriyev recalls. But it wasn't going to work out that way. It shattered his front axle, and he struggled to keep his truck and the wrecked car now fused to it from hitting anyone else as it barreled down the road. After Mugriyev finally came to a stop, he learned that the woman driving the car had been killed in the collision.

Much of the artificial intelligence talk at Mobile World Congress is at the device level -- services based on voice recognition, in particular. But AI has a role deeper in the infrastructure, of course, and Ericsson touched on that to kick off the conference Monday morning. In Ericsson's traditional MWC opener for press and analysts, Ericsson CEO Börje Ekholm understandably focused on 5G. But AI and machine learning were the second topic of his overview, an indication of how deeply Ericsson expects these technologies to change the network. Nokia is in the process of acquiring Deepfield, a deep learning startup, to help with traffic flows in software-defined networking (SDN).

Putting exhibition text into your phone is nothing new. Big institutions, from the National Gallery to the Netherlands' Rijksmuseum, have had dedicated apps for years, combining audio tours with factoids on collections, artists' lives and specific compositions. The Museo Nacional del Prado's app, for example, goes so far as to let viewers zoom into 14 masterpieces in Ultra HD, and see X-rays that reveal sketches beneath famous works of art. What arguably holds these apps back, however, is the fact they're tied to specific collections in specific buildings. New app on the block Smartify wants to address that, by creating a single platform for gallery-goers to scan artworks to access information and expert commentary across a range of different galleries and museums.

Summary: Count yourself lucky if you're not in one of the regulated industries where regulation requires you to value interpretability over accuracy. This has been a serious financial weight on the economy but innovations in Deep Learning point a way out. As Data Scientists we tend to take as gospel that more accuracy is better. There are some practical limits to this. It may not be profitable to continue to work a model for many days or weeks when the improvement to be had is minor.

This paper shows that a perturbed form of gradient descent converges to a second-order stationary point in a number iterations which depends only poly-logarithmically on dimension (i.e., it is almost "dimension-free"). The convergence rate of this procedure matches the well-known convergence rate of gradient descent to first-order stationary points, up to log factors. When all saddle points are non-degenerate, all second-order stationary points are local minima, and our result thus shows that perturbed gradient descent can escape saddle points almost for free. Our results can be directly applied to many machine learning applications, including deep learning. As a particular concrete example of such an application, we show that our results can be used directly to establish sharp global convergence rates for matrix factorization. Our results rely on a novel characterization of the geometry around saddle points, which may be of independent interest to the non-convex optimization community.

We draw a formal connection between using synthetic training data to optimize neural network parameters and approximate, Bayesian, model-based reasoning. In particular, training a neural network using synthetic data can be viewed as learning a proposal distribution generator for approximate inference in the synthetic-data generative model. We demonstrate this connection in a recognition task where we develop a novel Captcha-breaking architecture and train it using synthetic data, demonstrating both state-of-the-art performance and a way of computing task-specific posterior uncertainty. Using a neural network trained this way, we also demonstrate successful breaking of real-world Captchas currently used by Facebook and Wikipedia. Reasoning from these empirical results and drawing connections with Bayesian modeling, we discuss the robustness of synthetic data results and suggest important considerations for ensuring good neural network generalization when training with synthetic data.

Information that is stored in an encrypted format is, by definition, usually not amenable to statistical analysis or machine learning methods. In this paper we present detailed analysis of coordinate and accelerated gradient descent algorithms which are capable of fitting least squares and penalised ridge regression models, using data encrypted under a fully homomorphic encryption scheme. Gradient descent is shown to dominate in terms of encrypted computational speed, and theoretical results are proven to give parameter bounds which ensure correctness of decryption. The characteristics of encrypted computation are empirically shown to favour a non-standard acceleration technique. This demonstrates the possibility of approximating conventional statistical regression methods using encrypted data without compromising privacy.

As machine learning systems become ubiquitous, there has been a surge of interest in interpretable machine learning: systems that provide explanation for their outputs. These explanations are often used to qualitatively assess other criteria such as safety or non-discrimination. However, despite the interest in interpretability, there is very little consensus on what interpretable machine learning is and how it should be measured. In this position paper, we first define interpretability and describe when interpretability is needed (and when it is not). Next, we suggest a taxonomy for rigorous evaluation and expose open questions towards a more rigorous science of interpretable machine learning.

We introduce a method for using deep neural networks to amortize the cost of inference in models from the family induced by universal probabilistic programming languages, establishing a framework that combines the strengths of probabilistic programming and deep learning methods. We call what we do "compilation of inference" because our method transforms a denotational specification of an inference problem in the form of a probabilistic program written in a universal programming language into a trained neural network denoted in a neural network specification language. When at test time this neural network is fed observational data and executed, it performs approximate inference in the original model specified by the probabilistic program. Our training objective and learning procedure are designed to allow the trained neural network to be used as a proposal distribution in a sequential importance sampling inference engine. We illustrate our method on mixture models and Captcha solving and show significant speedups in the efficiency of inference.