DJ Strouse

I am a Senior Research Scientist at DeepMind in London.

I did my PhD in Physics at Princeton University, advised by David Schwab and Bill Bialek, and funded by a Hertz Fellowship and DOE Computational Science Graduate Fellowship. Before that, I did a master's at the University of Cambridge with Mate Lengyel as a Churchill Scholar and studied physics and mathematics at the University of Southern California, where I worked with Bartlett Mel and Paolo Zanardi and had a blog. Throughout my studies, I interned at DeepMind with Matt Botvinick, Stanford University with Kwabena Boahen, the Institute for Quantum Computing with Andrew Childs, and Spotify NYC with their machine learning team. I also enjoy a good puzzle.

Email  |  Twitter  |  CV  |  Scholar  |  GitHub

Research

I'm broadly interested in reinforcement and deep learning. Lately, I've been especially interested in the efficient training of agents that can meaningfully interact and collaborate with humans. My PhD thesis ("Optimization of MILES") focused on applications of the information bottleneck (IB) across supervised, unsupervised, and reinforcement learning, and definitely not on collecting airline miles. In past lives, I've also worked on quantum information theory and computational neuroscience.

Learning more skills through optimistic exploration
DJ Strouse*, Kate Baumli, David Warde-Farley, Vlad Mnih, Steven Hansen*
International Conference on Learning Representations (ICLR), 2022 (Spotlight)
arxiv | openreview | github | tweet | show bibtex

We highlight the inherent pessmism towards exploration in a popular family of variational unsupervised skill learning methods. To curb this pessimism, we propose an ensemble uncertainty based exploration bonus that we call discriminator disagreement intrinsic reward, or DISDAIN. We show that DISDAIN improves skill learning in both a gridworld and the Atari57 suite. Thus, we encourage researchers to treat pessimism with DISDAIN.

@inproceedings{strouse2022disdain,
  title = {Learning more skills through optimistic exploration},
  author = {Strouse, DJ and Baumli, Kate and Warde-Farley, David and Mnih, Vlad and Hansen, Steven},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year = {2022},
}

Learning Truthful, Efficient, and Welfare Maximizing Auction Rules
Andrea Tacchetti, DJ Strouse, Marta Garnelo, Thore Graepel, Yoram Bachrach
ICLR Gamification and Multiagent Solutions Workshop, 2022
arxiv | openreview | show bibtex

We present a deep learning approach to auction design that guarantees truthfulness (bidders are incentivized to be honest) and efficiency (whoever wants the item most gets it). We focus on social utility maximizing auctions, where the goal is to achieve the former constraints while placing as little economic burden on the bidders as possible.

@inproceedings{tacchetti2022auctioncnn,
  title = {Learning Truthful, Efficient, and Welfare Maximizing Auction Rules},
  author = {Tacchetti, Andrea and Strouse, DJ and Garnelo, Marta and Graepel, Thore and Bachrach, Yoram},
  booktitle = {ICLR Gamification and Multiagent Solutions Workshop},
  year = {2022},
}

Collaborating with Humans without Human Data
DJ Strouse*, Kevin R. McKee, Matt Botvinick, Edward Hughes, Richard Everett*
Neural Information Processing Systems (NeurIPS), 2021 (Spotlight)
arxiv | neurips | openreview | tweet | alignment newsletter | show bibtex

We introduce Fictitious Co-Play (FCP), a simple and intuitive training method for producing agents capable of zero-shot coordination with humans in Overcooked. FCP works by training an agent as the best response to a frozen pool of self-play agents and their past checkpoints. Notably, FCP exhibits robust generalization to humans, despite not using any human data during training.

@inproceedings{strouse2021fcp,
  title = {Collaborating with Humans without Human Data},
  author = {Strouse, DJ and McKee, Kevin R. and Botvinick, Matt and Hughes, Edward and Everett, Richard},
  booktitle = {Neural Information Processing Systems (NeurIPS)},
  year = {2021},
}

Social Influence as Intrinsic Motivation for Multi-Agent Deep Reinforcement Learning
Natasha Jaques, Angeliki Lazaridou, Edward Hughes, Caglar Gulcehre, Pedro A. Ortega, DJ Strouse, Joel Z. Leibo, Nando de Freitas
International Conference on Machine Learning (ICML), 2019 (Best Paper Honorable Mention)
arxiv | icml | openreview | show bibtex

We reward agents for influencing the actions of other agents, and show that this gives rise to better cooperation and more meaningful emergent communication protocols.

@inproceedings{jaques2019influence,
  title = {Social Influence as Intrinsic Motivation for Multi-Agent Deep Reinforcement Learning},
  author = {Jaques, Natasha and Lazaridou, Angeliki and Hughes, Edward and Gulcehre, Caglar and Ortega, Pedro and Strouse, DJ and Leibo, Joel Z. and De Freitas, Nando},
  booktitle = {International Conference on Machine Learning (ICML)},
  year = {2019},
}

The information bottleneck and geometric clustering
DJ Strouse, David Schwab
Neural Computation (NECO), 2019
pdf | neco | arxiv | code | show bibtex

We show how to use the (deterministic) information bottleneck to perform geometric clustering, introducing a novel information-theoretic model selection criterion. We show how this relates to and generalizes k-means and gaussian mixture models (GMMs).

@article{strouse2019clustering,
  title = {Geometric Clustering with the Information Bottleneck},
  author = {Strouse, DJ and Schwab, David J.},
  journal = {Neural Computation},
  year = {2019},
  volume = {31},
  number = {3},
  pages = {596-612},
}

InfoBot: Transfer and Exploration via the Information Bottleneck
Anirudh Goyal, Riashat Islam, DJ Strouse, Zafarali Ahmed, Hugo Larochelle, Matt Botvinick, Sergey Levine, Yoshua Bengio
International Conference on Learning Representations (ICLR), 2019
arxiv | openreview | show bibtex

We train agents in multi-goal environments with an information bottleneck between their goal and policy. This encourages agents to develop useful "habits" that generalize across goals. We identify the states where agents must deviate from their habits to solve a task as "decision states" and show that they are useful targets for an exploration bonus.

@inproceedings{goyal2019infobot,
  title={Transfer and Exploration via the Information Bottleneck},
  author={Anirudh Goyal and Riashat Islam and DJ Strouse and Zafarali Ahmed and Matthew Botvinick and Hugo Larochelle and Yoshua Bengio and Sergey Levine},
  booktitle={International Conference on Learning Representations (ICLR)},
  year = {2019},
}

Learning to share and hide intentions using information regularization
DJ Strouse, Max Kleiman-Weiner, Josh Tenenbaum, Matt Botvinick, David Schwab
Neural Information Processing Systems (NIPS), 2018
arxiv | nips | code | show bibtex

We train agents to cooperate / compete by regularizing the reward-relevant information they share with other agents, enabling agents trained alone to nevertheless perform well in a multi-agent setting.

@inproceedings{strouse2018intentions,
  title={Learning to share and hide intentions using information regularization},
  author = {Strouse, DJ and Kleiman-Weiner, Max and Tenenbaum, Josh and Botvinick, Matt and Schwab, David J},
  booktitle = {Neural Information Processing Systems (NeurIPS)},
  year = {2018},
}

The deterministic information bottleneck
DJ Strouse, David Schwab
Neural Computation (NECO), 2017 & Uncertainty in Artificial Intelligence (UAI), 2016
pdf | arxiv | code | uai | neco | show bibtex

We introduce the deterministic information bottleneck (DIB), an alternative formulation of the information bottleneck that uses entropy instead of mutual information to measure compression. This results in a hard clustering algorithm with a built-in preference for using fewer clusters.

@article{strouse2017dib,
  title = {The Deterministic Information Bottleneck},
  author = {Strouse, DJ and Schwab, David J.},
  journal = {Neural Computation},
  year = {2017},
  volume = {29},
  number = {6},
  pages = {1611-1630},
}

Neuroscience
How Dendrites Affect Online Recognition Memory
Xundong Wu, Gabriel Mel, DJ Strouse, Bartlett Mel
PLoS Computational Biology, 2019
plos

We study the optimal conditions for online recognition memory in a biologically-inspired neural network with "dendrite-aware" learning rules.

Using neural networks to understand the computational role of dendrites
DJ Strouse, Balazs Ujfalussy, Mate Lengyel
Computational and Systems Neuroscience (Cosyne), 2012 & 2013
2012 poster & abstract | 2013 poster & abstract | master's thesis | why

We fit neural network models to single neuron data to understand the computational role of dendrites in integrating their synaptic input.

Behaviorally-locked structure in a sensory neural code
DJ Strouse, Jakob Macke, Roman Shusterman, Dima Rinberg, Elad Schneidman
Sensory Coding & the Natural Environments (SCNE), 2012
abstract | poster

We study the olfactory neural code in mice and find that much of the information about the stimulus is only decodable when interpreting neural activity relative to the sniff phase, providing evidence for the importance of considering sensory sampling behavior when interpreting neural codes.

Physics
Levinson's theorem for graphs
Andrew Childs, DJ Strouse
Journal of Mathematical Physics (JMP), 2011
arxiv | jmp | talk

We prove an analog of a classic result in quantum scattering theory for the setting of scattering on graphs. The goal is to provide additional tools for designing quantum algorithms in this setting.


Good artists copy.