Alphabet’s self-driving car company, Waymo, is introducing truly driverless cars to public roads for the first time as part of a limited experiment in Phoenix. Check out this YouTube video for a demo.

Pieter Abbeel, a former Berkeley professor and researcher at OpenAI, and three other researchers left OpenAI to start their own robotics company, Embodied Intelligence. The company is backed by $7 million in funding from the Silicon Valley venture capital firm Amplify Partners and other investors.

While feature visualization is a powerful tool, actually getting it to work involves a number of details. In this article, the authors examine the major issues and explore common approaches to solving them.

Learn how to apply Evolution Strategies (ES) to Reinforcement Learning problems and how to find policies that are stable and robust. The code for the experiments is also available on Github.

The recent successes of deep learning are mostly empirical. This literature course reviews work seeking to build theoretical frameworks deriving deep networks as consequences. The lecture videos are published here.

A neural machine translation system that can produce an entire sentence at a time in a fully parallel way, overcoming a limitation of current neural MT models.

Tangent is an open-source Python library for automatic differentiation. In contrast to existing machine learning libraries, Tangent is a source-to-source system, consuming and generating Python code.

Implementation of various Evolution Strategies, such as GA, PEPG, CMA-ES and OpenAI’s ES using a common interface. Also, check out the related blog post.

Neural networks, 13 new models for 7+ languages, better training, custom pipelines, Pickle & lots of API improvements.

A Real-time Mario Kart AI using CNNs, Offline Search, and DAGGER.

A simple yet powerful generative model that learns discrete representations. The model, the Vector Quantised-Variational AutoEncoder (VQ-VAE), differs from VAEs in two key ways: the encoder network outputs discrete, rather than continuous, codes; and the prior is learned rather than static.

Expert Iteration (ExIt) is a novel reinforcement learning algorithm which decomposes the problem into separate planning and generalization tasks. Planning new policies is performed by tree search, while a deep neural network generalizes those plans. Subsequently, tree search is improved by using the neural network policy to guide search.

Two different approaches to induce block sparsity in RNNs: pruning blocks of weights in a layer and using group lasso regularization to create blocks of weights with zeros. Using these techniques, the authors demonstrate that one can create block-sparse RNNs with sparsity ranging from 80% to 90% with a small loss in accuracy, reducing model size by roughly 10x.