Shibhansh Dohare
I am a Ph.D. candidate at the University of Alberta,
advised by Dr. Richard Sutton and Dr. Rupam Mahmood.
I completed my B.Tech. at IIT Kanpur in Computer Science and Engineering.
My long-term research goal is to understand the workings of our minds.
Specifically, to help find the computational principles that give rise to the mind.
In pursuit of this goal, I'm working on various aspects of continual learning, deep learning, and reinforcement learning.
Continual learning is starting to be applied in the industry.
I think many applications require continual adaptation, and deep continual learning has big potential in the next 5 years.
See a list of possible applications here.
During my Ph.D., I have contributed to exposing a fundamental problem with deep learning systems,
where these systems can lose the ability to learn new things.
I also developed the continual backpropagation algorithm to overcome this problem.
My Ph.D. research has been published in Nature and featured in some popular media outlets,
such as New Scientist.
If you prefer podcasts, I have also discussed my work on the Nature Podcast and AMII's Approximately Correct Podcast.
Email  / 
CV  / 
Google Scholar  / 
Github
|
|
|
Reinitializing Weights vs Units For Maintaining Plasticity in Neural Networks
J. Fernando Hernandez-Garcia,
Shibhansh Dohare,
Jun Luo,
Richard S. Sutton
CoLLAs, Oral Presentation, 2025
Paper
|
Code
We propose a new algorithm, which we name selective weight reinitialization,
for reinitializing the least useful weights in a network.
We find that selective weight reinitialization is more effective at maintaining plasticity than
reinitializing units (continual backpropagation) when the network includes layer normalization or attention layers.
|
|
Loss of Plasticity in Deep Continual Learning
Shibhansh Dohare,
J. Fernando Hernandez-Garcia,
Qingfeng Lan,
Parash Rahman,
A. Rupam Mahmood,
Richard S. Sutton
Nature 2024
Paper
|
Code
|
Nature Podcast
|
News
We provide direct demonstrations of plasticity loss in deep continual learning.
We propose a new algorithm, continual backpropagation, that fully maintains plasticity.
Continual backpropagation reinitializes a small fraction of less-used units alongside gradient descent at each update.
|
|
Overcoming Policy Collapse in Deep Reinforcement Learning
Shibhansh Dohare,
Qingfeng Lan,
A. Rupam Mahmood
EWRL 2023
Paper
We show that popular deep RL algorithms, like PPO, do not scale with experience. Their performance gets worse over time. We look deeper into this problem and provide simple solutions to reduce performance degradation.
|
|
Automatic Noise Filtering with Dynamic Sparse Training in Deep Reinforcement Learning
Bram Grooten,
Ghada Sokar,
Shibhansh Dohare,
Elena Mocanu,
Matthew E. Taylor,
Mykola Pechenizkiy,
Decebal Constantin Mocanu
AAMAS 2023
Paper
We show that standard Deep RL algorithms fail when the input contains noisy features.
Dynamic sparse training successfully filters through the noisy features and performs well.
|
|
Gamma-Nets: Generalizing Value Estimation over Timescale
Craig Sherstan,
Shibhansh Dohare,
James MacGlashan,
Johannes Günther,
Patrick M. Pilarski,
AAAI, Oral Presentation, 2020
Paper
We present Gamma-nets, a method for generalizing value function estimation over timescale.
|
|
Unsupervised Semantic Abstractive Summarization
Shibhansh Dohare,
Vivek Gupta,
Harish Karnick,
ACL, Student Research Workshop, 2018
Paper
A novel algorithm for abstractive text summarization based on Abstract Meaning Representation.
|
|