michael (misha) laskin

I'm a postdoc at UC Berkeley in the BAIR lab, where I work on deep unsupervised and reinforcement learning with Pieter Abbeel. I received my PhD in theoretical physics from the University of Chicago, where I worked on statisical quantum physics with Paul Wiegmann. During my PhD I received the Bloomenthal Fellowship, awarded to the best graduate student in theoretical physics. In between my PhD and postdoc I started a company - Claire AI - which made AI products for the retail sector and went through Y Combinator (W17). My co-founder and I made the Forbes 30 Under 30 list in 2017 for our work in retail and e-commerce. Before that, I received a BS from Yale University where I studied physics and literature.

Email / Twitter / GitHub / Google Scholar / LinkedIn

Follow @MishaLaskin

AI Research

	Decoupling Representation Learning from Reinforcement Learning Adam Stooke, Kimin Lee, Pieter Abbeel, Michael Laskin In Submission, 2020 paper / code / twitter First algorithm that decouples unsupervised learning from reinforcement learning while matching or outperforming state-of-the-art end-to-end RL.
	A Simple Unified Framework for Ensemble Learning in Deep Reinforcement Learning Kimin Lee, Michael Laskin, Aravind Srinivas, Pieter Abbeel In Submission, 2020 paper / code Developed a unified framework for utilizing ensembles to greatly stabilize training of both state-based and pixel-based RL algorithms.
	Reinforcement Learning with Augmented Data Michael Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, Aravind Srinivas NeurIPS Spotlight* (top 3% of submissions)*, 2020 paper / code / twitter / press First extensive study of image augmentation in the RL setting. Showed that simple RL algorithms with augmented data achieve SOTA results on many common RL benchmarks.
	CURL: Contrastive Unsupervised Representations for Reinforcement Learning Michael Laskin, Aravind Srinivas, Pieter Abbeel ICML, 2020 paper / code / twitter Showed for the first time that RL from pixels can be as efficient as RL from state by leveraging unsupervised contrastive representations.
	Sparse Graphical Memory for Robust Planning Michael Laskin, Scott Emmons, Ajay Jain, Thanard Kurutach, Pieter Abbeel, Deepak Pathak NeurIPS, 2020 paper / video / code Introduced novel state aggregation criterium - two-way consistency (TWC)* - and utilized it to make any semi-parametric graphical method more robust. Proved theoretically that TWC enables near-optimal search.
	Discrete Representation Learning for Goal-Conditioned Visual Reinforcement Learning Michael Laskin, Thanard Kurutach, Pieter Abbeel NeurIPS Deep Reinforcement Learning Workshop, 2019 paper Discrete representations help reduce the size of the encoded observation space. We showed how utilizing a discrete bottleneck can improve goal-conditioned RL from pixels.

Physics Research

During my physics PhD I studied many-body quantum systems such as the Fractional Quantum Hall Effect, and discovered a universal topological characteristic of such states.

Note: in theoretical physics alphabetical order is the usual convention for authorship.

Emergent conformal symmetry and geometric transport properties of quantum Hall states on singular surfaces

T. Can, Y.H. Chiu, M. Laskin, P. Wiegmann
Physical review letters 117 (26), 266803, 2016

Population of the giant pairing vibration

M. Laskin, R.F. Casten, A.O. Macchiavelli, R.M. Clark, D. Bucurescu
Physical Review C 93 (3), 034321, 2016

Collective field theory for quantum Hall states

T. Can, M. Laskin, P. Wiegmann
Physical Review B 92 (23), 235141 , 2015

Geometry of quantum Hall states: Gravitational anomaly and transport coefficients

T. Can, M. Laskin, P. Wiegmann
Annals of Physics 362, 752-794, 2015

Fractional quantum Hall effect in a curved space: gravitational anomaly and electromagnetic response

T. Can, M. Laskin, P. Wiegmann
Physical review letters 113 (4), 046803, 2014

Field Theory for Fractional Quantum Hall States

T. Can, M. Laskin, P. Wiegmann
arXiv preprint arXiv:1412.8716, 2014

Some Aspects of the Giant-Pairing Vibration

A.O. Macchiavelli, R.M. Clark, M. Laskin, R.F. Casten
Bulletin of the American Physical Society, 2012

Design from this, this, and this

michael (misha) laskin

AI Research

Decoupling Representation Learning from Reinforcement Learning

A Simple Unified Framework for Ensemble Learning in Deep Reinforcement Learning

Reinforcement Learning with Augmented Data

CURL: Contrastive Unsupervised Representations for Reinforcement Learning

Sparse Graphical Memory for Robust Planning

Discrete Representation Learning for Goal-Conditioned Visual Reinforcement Learning