Rafał Józefowicz

Also known as Rafal Jozefowicz

Rafał Józefowicz is a deep learning researcher whose papers cover recurrent networks, language models and reinforcement learning. After a computer science degree at Jagiellonian University and jobs at Microsoft and Two Sigma Investments, they joined Google in 2014 and worked as a researcher at Google Brain. In 2015, with Wojciech Zaremba and Ilya Sutskever, they tested over ten thousand recurrent network architectures; in 2016, they coauthored Exploring the Limits of Language Modeling and were among the authors of the TensorFlow paper. They joined OpenAI in June 2016 and coauthored the 2017 sentiment-unit paper with Alec Radford and Ilya Sutskever, followed by the 2018 dexterous in-hand manipulation paper. The 2019 Dota 2 paper, which lists them among its authors, reports that OpenAI Five became the first AI system to defeat the world champions at an esports game.

Now at: OpenAI

Timeline

By 2014

Quant engineer at Two Sigma

They graduate in computer science from Jagiellonian University and are hired by Microsoft as a software development engineer. Before joining Google in 2014, they work as a quantitative software engineer at Two Sigma Investments on equity derivatives and global macro strategies, using artificial intelligence, machine learning and distributed computing.
2014–2016

Researcher at Google Brain

They join Google in April 2014 and leave in June 2016. As a researcher at Google Brain focused on natural language processing with neural networks, they give an overview of TensorFlow at the TensorFlow New York meetup in March 2016. The arXiv paper describing TensorFlow, posted in March 2016, lists them among its authors; the system was built at Google and released as open source in November 2015.
2015

Searching ten thousand RNN architectures

With Wojciech Zaremba and Ilya Sutskever, they publish "An Empirical Exploration of Recurrent Network Architectures" in the proceedings of the 32nd International Conference on Machine Learning in 2015. The study evaluates over ten thousand RNN architectures and shows that adding a bias of 1 to the LSTM's forget gate closes the gap between the LSTM and the GRU.
2016

Exploring the limits of language modeling

In February 2016, they coauthor "Exploring the Limits of Language Modeling" with Oriol Vinyals, Mike Schuster, Noam Shazeer and Yonghui Wu. The best single model lowers perplexity on the One Billion Word Benchmark from 51.3 to 30.0, and an ensemble sets a record by lowering it from 41.0 to 23.7.
2016–2017

A sentiment unit at OpenAI

They join OpenAI in June 2016. In 2017, they coauthor "Learning to Generate Reviews and Discovering Sentiment" with Alec Radford and Ilya Sutskever, which finds a single unit in a byte-level recurrent language model that performs sentiment analysis and reaches state of the art on the binary subset of the Stanford Sentiment Treebank.
2018

Dexterous in-hand manipulation

In 2018, they are among the authors of "Learning Dexterous In-Hand Manipulation", which uses reinforcement learning to train vision-based object reorientation policies for a Shadow Dexterous Hand entirely in simulation. The policies transfer to the physical robot, and the results come from the same distributed RL system that trained OpenAI Five.
2019

OpenAI Five beats Dota 2 champions

In April 2019, OpenAI Five becomes the first AI system to defeat the world champions at an esports game, beating the Dota 2 champion Team OG. They are among the authors of "Dota 2 with Large Scale Deep Reinforcement Learning", posted in December 2019, which describes the distributed training system behind OpenAI Five, a run of ten months that learns from batches of about 2 million frames every 2 seconds.