NOTE: This story has been revised from when it was originally published in October 2015 to reflect the latest information. Yesterday, news emerged that Google was using a machine-learning artificial intelligence system called "RankBrain" to help sort through its search results. Wondering how that works and fits in with Google's overall ranking system? Here's what we know about RankBrain. The information covered below comes from three original sources and has been updated over time, with notes where updates have happened.

SpotMini is a new smaller version of the Spot robot, weighing 55 lbs dripping wet (65 lbs if you include its arm.) SpotMini is all-electric (no hydraulics) and runs for about 90 minutes on a charge, depending on what it is doing. SpotMini is one of the quietest robots we have ever built. It has a variety of sensors, including depth cameras, a solid state gyro (IMU) and proprioception sensors in the limbs. These sensors help with navigation and mobile manipulation.

Would you buy a car that might risk hurting you to protect a pedestrian? It is a question social-science researchers are exploring amid the development of driverless cars. While commercial applications may be years away, any fully autonomous vehicle that eventually takes to the road will need to make decisions--like whether to swerve to miss one pedestrian at the risk of hitting another. Many ethicists argue that a public conversation should be part of the development process. In a study published in Science, researchers found people want the cars to be programmed to minimize casualties while on the road.

It has been observed in a variety of contexts that gradient descent methods have great success in solving low-rank matrix factorization problems, despite the relevant problem formulation being non-convex. We tackle a particular instance of this scenario, where we seek the $d$-dimensional subspace spanned by a streaming data matrix. We apply the natural first order incremental gradient descent method, constraining the gradient method to the Grassmannian. In this paper, we propose an adaptive step size scheme that is greedy for the noiseless case, that maximizes the improvement of our metric of convergence at each data index $t$, and yields an expected improvement for the noisy case. We show that, with noise-free data, this method converges from any random initialization to the global minimum of the problem. For noisy data, we provide the expected convergence rate of the proposed algorithm per iteration.

The problem of probabilistic forecasting and online simulation of real-time electricity market with stochastic generation and demand is considered. By exploiting the parametric structure of the direct current optimal power flow, a new technique based on online dictionary learning (ODL) is proposed. The ODL approach incorporates real-time measurements and historical traces to produce forecasts of joint and marginal probability distributions of future locational marginal prices, power flows, and dispatch levels, conditional on the system state at the time of forecasting. Compared with standard Monte Carlo simulation techniques, the ODL approach offers several orders of magnitude improvement in computation time, making it feasible for online forecasting of market operations. Numerical simulations on large and moderate size power systems illustrate its performance and complexity features and its potential as a tool for system operators.

A common analytical problem in neuroscience is the interpretation of neural activity with respect to sensory input or behavioral output. This is typically achieved by regressing measured neural activity against known stimuli or behavioral variables to produce a "tuning function" for each neuron. Unfortunately, because this approach handles neurons individually, it cannot take advantage of simultaneous measurements from spatially adjacent neurons that often have similar tuning properties. On the other hand, sharing information between adjacent neurons can errantly degrade estimates of tuning functions across space if there are sharp discontinuities in tuning between nearby neurons. In this paper, we develop a computationally efficient block Gibbs sampler that effectively pools information between neurons to de-noise tuning function estimates while simultaneously preserving sharp discontinuities that might exist in the organization of tuning across space. This method is fully Bayesian and its computational cost per iteration scales sub-quadratically with total parameter dimensionality. We demonstrate the robustness and scalability of this approach by applying it to both real and synthetic datasets. In particular, an application to data from the spinal cord illustrates that the proposed methods can dramatically decrease the experimental time required to accurately estimate tuning functions.

In this paper, we consider the problem of modeling discrete social network data and learning the underlying group dynamics. The goal is to develop probabilistic profiles of large collections of data while preserving the essential temporal relationships that provide insights for various applications of interest. For example, in social network analysis, we want to analyze relationships between social agents and their behaviors over time and on various social media sites (i.e., Facebook, Twitter, Instagram, Google, etc.). In web advertising analysis, we want to analyze the relationships between customers and the types of products they buy from different shopping sites to capture customers' buying behaviors and learn the intrinsic factors that effect their buying decision process. In the study of scientific collaboration, using co-authorship networks from multiple journals on related subjects, one can analyze relationships between subjects and authors.

Generalized linear models with nonlinear feature transformations are widely used for large-scale regression and classification problems with sparse inputs. Memorization of feature interactions through a wide set of cross-product feature transformations are effective and interpretable, while generalization requires more feature engineering effort. With less feature engineering, deep neural networks can generalize better to unseen feature combinations through low-dimensional dense embeddings learned for the sparse features. However, deep neural networks with embeddings can over-generalize and recommend less relevant items when the user-item interactions are sparse and high-rank. In this paper, we present Wide & Deep learning---jointly trained wide linear models and deep neural networks---to combine the benefits of memorization and generalization for recommender systems. We productionized and evaluated the system on Google Play, a commercial mobile app store with over one billion active users and over one million apps. Online experiment results show that Wide & Deep significantly increased app acquisitions compared with wide-only and deep-only models. We have also open-sourced our implementation in TensorFlow.

We present three case studies of organizations using a data science competition to answer a pressing question. The first is in education where a nonprofit that creates smart school budgets wanted to automatically tag budget line items. The second is in public health, where a low-cost, nonprofit women's health care provider wanted to understand the effect of demographic and behavioral questions on predicting which services a woman would need. The third and final example is in government innovation: using online restaurant reviews from Yelp, competitors built models to forecast which restaurants were most likely to have hygiene violations when visited by health inspectors. Finally, we reflect on the unique benefits of the open, public competition model.

This paper deals with prediction of anopheles number, the main vector of malaria risk, using environmental and climate variables. The variables selection is based on an automatic machine learning method using regression trees, and random forests combined with stratified two levels cross validation. The minimum threshold of variables importance is accessed using the quadratic distance of variables importance while the optimal subset of selected variables is used to perform predictions. Finally the results revealed to be qualitatively better, at the selection, the prediction , and the CPU time point of view than those obtained by GLM-Lasso method.