A recent experiment ended up with an AI agent that beat ‘80s classic game Q*bert in a way no one has ever seen before. It discovered a bug that allowed it to rack up near infinite points. Training agents to solve games could be an interesting way to discover unintended bugs. You can read the paper here.

The researchers gathered training data from electrodes implanted in the patients’ brains while they were solving a memory task. The same electrodes are then used to stimulate neural activity. The system improved patients’ ability to recall words by an average of 15 percent.

The company calls the product a “smart camera,” a reference to the machine learning happening on the device in order to make the experience as simple as possible. It “captures the moments that happen in between posed pictures by using on-device machine learning to look for great facial expressions from the people—and pets—in your life”, according to Google.

Google’s 15+ hours Machine Learning Crash Course features a series of lessons with video lectures, real-world case studies, and hands-on practice exercises. Using Tensorflow.

This post provides inspiration and ideas for research directions to junior researchers and those trying to get into research. It gathers a collection of research topics that are interesting with a focus on NLP and transfer learning.

A dive into Variational Autoencoders, a class of neural networks that can learn to compress data in a completely unsupervised way.

If life plays by the rules, image recognition works well. But as soon as people - or sheep - do something unexpected, the algorithms show their weaknesses.

“How does depth help?” is a fundamental question in the theory of deep learning.  Conventional wisdom backed by theoretical studies is that adding layers increases expressive power, but optimization becomes harder. However, it turns out that increasing depth can sometimes accelerate optimization.

OpenAI open-sourced eight simulated robotics environments and a baseline implementation of the Hindsight Experience Replay algorithm. The environments are now part of OpenAI gym and the algorithm implementation can be found in the baselines repository.

Lore is a Python framework to make machine learning approachable for engineers and maintainable for Machine Learning researchers. This post walks you through an example of how to design and deploy a model with Lore.

The release includes prebuilt binaries are now built against CUDA 9.0 and cuDNN 7, a new optimizer internal API for non-slot variables, improvements to SavedModels exports, FFT support added to XLA CPU/GPU, and more.

Keras implementations of many Generative Adversarial Networks (GANs) from various research papers. Some of these are simplified versions of what’s found in the papers, but they serve as an excellent learning resource.

Infants are experts at playing, with an amazing ability to generate novel structured behaviors in unstructured environments that lack clear extrinsic reward signals. The authors investigate curiosity-driven intrinsic motivation and propose a “world-model” network that learns to predict the dynamic consequences of the agent’s actions. Simultaneously, they train a separate explicit “self-model” that allows the agent to track the error map of its own world-model, and then uses the self-model to adversarially challenge the developing world-model.

The authors demonstrate qualitatively that ES algorithms have very different performance characteristics than traditional RL algorithms: on some games, they learn to exploit the environment and perform much better while on others they can get stuck in suboptimal local minima. Combining their strengths with those of traditional RL algorithms is therefore likely to lead to new advances in the state of the art.

The authors propose a new simulated benchmark for robotic grasping that emphasizes off-policy learning and generalization to unseen objects. They evaluate the benchmark tasks against a variety of Q-function estimation methods, a method previously proposed for robotic grasping with deep neural network models, and a novel approach based on a combination of Monte Carlo return estimation and an off-policy correction.

Incremental training of an increasingly general problem solver, continually learning to solve new tasks without forgetting previous skills. The problem solver is a single recurrent neural network called ONE and can be trained in various ways, e.g. black-box optimization, reinforcement learning, artificial evolution as well as supervised and unsupervised learning.