Evgenii Nikishin

I am a PhD student at Mila, University of Montreal, where I have the good fortune of being advised by Pierre-Luc Bacon and Aaron Courville. I am broadly interested in Reinforcement Learning and using capabilities of Deep Learning for decision making.

Previously, I spent 1.5 years at Cornell University. I received an MS in Computer Science from Higher School of Economics and Skoltech where I worked with Dmitry Vetrov at the Samsung Lab. I did my undergrad in CS at Lomonosov Moscow State University supervised by Alexander Dyakonov.

During the 2nd half of 2022, I interned in the RL team at DeepMind London working closely with André Barreto, Junhyuk Oh, and Will Dabney. I spent summer 2020 remotely interning at Mila with Pierre-Luc Bacon and Yoshua Bengio. In summer 2018, I interned at ETH Zürich with Gunnar Rätsch applying RL for personal treatment.

Email  /  CV  /  Google Scholar  /  Twitter  /  GitHub

We show that neuron saturation, which traditionally was viewed as undesirable, could instead lead to sparse yet accurate networks.

We demonstrate that individual ensemble members in RL exhibit surprisingly low performance (whilst aggregate returns are adequate) and propose a remedy.

We propose an intervention for diagnosing the loss of plasticity phenomenon in RL and dynamically growing neural networks in RL for increasing computational efficiency.

An analysis of plasticity loss showing its relation to pathological loss landscapes and demonstrating efficiency of layer normalization to mitigate it.

Resets unlock increasing sample efficiency by scaling the number of updates per environment step.

We identify a damaging tendency of deep RL agents to overfit early experiences and propose a simple yet powerful remedy based on periodic resetting of a part of the agent.

A model learning method for RL that directly optimizes the sum of rewards instead of likelihood, a proxy to the agent's objective.

We propose an algorithm to construct discrete token adversarial examples based on the notion of synonyms.

Domain adaptation via learning shared dynamics in a latent space with adversarial matching of latent states.

Averaging weights during training of an RL agent stabilizes the achieved performance.