Joel Veness

Joel Veness is a computer scientist and DeepMind research scientist whose work runs from universal artificial intelligence to deep reinforcement learning and compression. He wrote the chess engine Bodo and the self-play program Meep, and in 2011 he published A Monte-Carlo AIXI Approximation and completed a PhD at the University of New South Wales. A postdoc at the University of Alberta under Michael Bowling led to the 2012 Arcade Learning Environment paper, and at DeepMind he coauthored the 2015 Nature paper on the deep Q-network, the 2017 paper on overcoming catastrophic forgetting and the 2021 Gated Linear Networks paper. His Language Modeling Is Compression work, posted in 2023 and presented at ICLR 2024, showed that Chinchilla 70B compresses ImageNet patches and LibriSpeech samples better than PNG and FLAC. He is a research scientist at DeepMind in the UK, now Google DeepMind, a role he has held since 2012.

Now at: Google DeepMind
Badges: Bridge

Timeline

By 2011

Chess engines and a UNSW PhD

He writes the chess engine Bodo and builds Meep on top of it, one of the first master-level programs with an evaluation function learned entirely from self-play. He completes a PhD at the University of New South Wales in 2011 with Marcus Hutter as advisor; the thesis is Approximate Universal Artificial Intelligence and Self-Play Learning for Games.
2011

A Monte-Carlo AIXI approximation

He is the first author of A Monte-Carlo AIXI Approximation, published in the Journal of Artificial Intelligence Research in January 2011 with Kee Siong Ng, Marcus Hutter, William Uther and David Silver. It gives the first computationally feasible approximation to the AIXI agent, combining a new Monte-Carlo Tree Search algorithm with an agent-specific extension of Context Tree Weighting.
2011–2012

Postdoc and the Arcade Learning Environment

He works as a postdoctoral researcher at the University of Alberta under Michael Bowling from 2011 to 2012. In July 2012, he and Marc Bellemare, Yavar Naddaf and Bowling post The Arcade Learning Environment, which offers an interface to hundreds of Atari 2600 game environments for testing general, domain-independent AI.
2015

Deep Q-network in Nature

He is one of the Google DeepMind authors of Human-level control through deep reinforcement learning, published in Nature in February 2015 with Volodymyr Mnih, Koray Kavukcuoglu, David Silver and Demis Hassabis among the coauthors. The deep Q-network agent, given only pixels and the game score, reaches a level comparable to a professional human games tester across 49 Atari 2600 games.
2017

Overcoming catastrophic forgetting

He coauthors Overcoming catastrophic forgetting in neural networks, published in the Proceedings of the National Academy of Sciences in March 2017 by James Kirkpatrick and a team at DeepMind in London. The method remembers old tasks by selectively slowing down learning on the weights important for those tasks, and it learns several Atari 2600 games in sequence.
2021

Gated Linear Networks at AAAI

He is the first author of Gated Linear Networks, published at AAAI-21 in May 2021 with Tor Lattimore, Marcus Hutter and other DeepMind colleagues. The paper presents a family of backpropagation-free neural architectures whose credit assignment is distributed and local.
2023–2024

Language modeling is compression

He coauthors Language Modeling Is Compression, posted in September 2023 with Grégoire Delétang, Anian Ruoss, Marcus Hutter and others, and presented at ICLR 2024. The paper shows that Chinchilla 70B compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating PNG and FLAC.
Since 2012

Research scientist at DeepMind

From 2012, he works in the UK as a research scientist at DeepMind, now Google DeepMind. In February 2025, he coauthors Partition Tree Weighting for Non-Stationary Stochastic Bandits with Marcus Hutter, Andras Gyorgy and Jordi Grau-Moya, which extends universal source coding for interaction data to the control setting.