Ever thought of making CNNs for image classification smaller, faster, and more accurate, by simply hacking the the JPEG codec? Also included are surprising insights about frequency space and color information as they relate to network architecture design.

Generative Adversarial Networks (GANs) have achieved impressive feats in realistic image generation and image repair. At Uber, GANs have myriad potential applications, including strengthening our machine learning (ML) models against adversarial attacks, learning simulators for traffic, ride requests, or demand patterns over time.

In deep reinforcement learning (RL), solving the Atari games Montezuma’s Revenge and Pitfall has been a grand challenge. Today we introduce Go-Explore, a new family of algorithms capable of achieving scores over 2,000,000 on Montezuma’s Revenge and scoring over 400,000 on average!

Uber held its first open source summit on November 15, 2018, inviting members of the open source community for presentations given by experts on some of the projects we have contributed in the fields of big data, visualization, machine learning, and front-end web application building.

We are delighted to welcome Jan Pedersen to Uber AI. Jan has a tremendous amount of experience in technology and science leadership roles, having served as Chief Scientist for Core Search at Microsoft, VP for Data Science at Twitter, and most recently, Chief Scientist at eBay. He joins Uber as a Distinguished Scientist and will help us navigate and grow our AI and machine learning efforts.

Interested in accelerating your career by tackling some of Uber’s most challenging AI problems? Apply for Uber residency, a research fellowship dedicated to fostering the next generation of AI talent.

Motion-sensor cameras can cheaply & unobtrusively gather vast amounts of data on wild animals. We show that deep learning can automate animal identification for 99.3% of the Snapshot Serengeti dataset while performing at the same 96.6% accuracy of humans.

Visual Inspector for Neuroevolution (VINE) is an interactive data visualization tool that offers fresh insight into the underlying dynamics of evolution to help those who are interested better understand and explore this family of algorithms.

We have open sourced much faster code for evolving deep neural networks. Such neuroevolution approaches can solve challenging deep reinforcement learning tasks. This code allows training Atari in a few hours on a single modern desktop.