Michael Zhang

I’m Michael Zhang, a final-year PhD student at the University of Toronto and Vector Institute, where I am fortunate to be supervised by Jimmy Ba and supported by the NSERC CGS Fellowship and the Schwartz Reisman Graduate Fellowship.

I am interested in building safe, general-purpose machine learning systems. I did my undergrad and Master's at the University of California, Berkeley and have previously done internships at Tesla, Google Research, and LinkedIn.

You can contact me at [first name] @cs.toronto.edu.

CV  /  Google Scholar  /  Twitter  /  Substack

profile photo
  • Dec 2023: We released a new preprint on using language models for hyperparameter tuning; we'll present this at the Neurips FMDM workshop. Arxiv link.
  • Dec 2023: I gave an interview about AI Safety and related research directions at Toronto, which was featured in an U of T news article. Link
  • July 2023: Presenting work on language model+prompting-based approach for answering course questions at ITiCSE in Finland.
  • May 2023: I'm grateful to receive the Schwartz Reisman Institute graduate fellowship!
  • Apr 2023: We released a new preprint on a prompt boosting algorithm with large language models: arxiv.
  • Feb 2023: Multi-Rate VAE received an oral presentation (top-5% of accepted papers) at ICLR.
  • Jan 2023: I am an instructor for CSC311: Intro to Machine Learning this semester!

Some topics I am currently interested in:
-How can we develop current and future AI models that are more likely to be socially positive?
-Improving our understanding of deep neural network optimization
-Can we make hyperparameter tuning easier (e.g. more efficient and automated)?
-AI Safety and interdisciplinary thinking about technology (e.g. through SRI).

Please feel free to reach out if you have an idea you'd like to discuss. A full list of publications is on my Google Scholar.

Large Language Models for Hyperparameter Optimization
Michael R. Zhang , Nishkrit Desai, Juhan Bae, Jonathan Lorraine, Jimmy Ba
NeurIPS Foundation Models for Decision Making Workshop
Paper / Code

We develop a methodology where LLMs suggest hyperparameters and show it can match or outperform traditional HPO methods like Bayesian optimization across different models on standard benchmarks.

Unlearnable Algorithms for In-context Learning
Andrei Muresanu, Anvith Thudi, Michael R. Zhang, Nicolas Papernot

We analyze unlearning in the in-context learning setting and propose an algorithm that is amenable to unlearning and efficiently selects examples for in-context learning.

Decomposed Prompting to Answer Questions on a Course Discussion Board
Brandon Jaipersaud, Paul Zhang, Jimmy Ba, Andrew Petersen, Lisa Zhang, Michael R. Zhang
AIED 2023

We propose and evaluate a question-answering system that uses decomposed prompting to classify and answer student questions on a course discussion board.

Multi-Rate VAE: Train Once, Get the Full Rate-Distortion Curve
Juhan Bae, Michael R. Zhang, Michael Ruan, Eric Wang, So Hasegawa, Jimmy Ba, Roger Grosse
ICLR 2023 (top-5% of accepted papers)

We propose Multi-Rate VAE (MR-VAE), a hypernetwork which is capable of learning multiple VAEs with different rates in a single training run.

Autoregressive Models for Offline Policy Evaluation and Optimization
Michael R. Zhang, Tom Le Paine, Ofir Nachum, Cosmin Paduraru, George Tucker, Ziyu Wang, Mohammad Norouzi
ICLR 2021
Paper / Video

Autoregressive models are better for learning dynamics models than standard feedforward models in the fixed dataset setting.

Analyzing Monotonic Linear Interpolation in Neural Network Loss Landscapes
James Lucas, Juhan Bae, Michael R. Zhang , Stanislav Fort, Richard Zemel, Roger Grosse
ICML 2021

We investigate and analyze why many neural networks have the loss decreasing monotonically in weight space from initialization to final solution.

Objective Social Choice: Using Auxiliary Information to Improve Voting Outcomes
Silviu Pitis, Michael R. Zhang
AAMAS 2020
Paper / Code / Video

Framework and aggregation rules for combining the preferences of multiple agents with noisy views of some ground truth.

Lookahead Optimizer: k steps forward, 1 step back
Michael R. Zhang, James Lucas, Geoffrey Hinton, Jimmy Ba
NeurIPS 2019
Paper / Code / Video

Optimization algorithm that speeds up training by using a search direction generated by multiple steps of an inner optimizer.

Reverse Curriculum Generation for Reinforcement Learning
Carlos Florensa, David Held, Markus Wulfmeier, Michael Zhang, Pieter Abbeel
CoRL 2017
Webpage / Paper / Blog post

Approach for tackling sparse reward tasks by generating goals of intermediate difficulty.

Probabilistically Safe Policy Transfer
David Held, Zoe McCarthy, Michael Zhang, Fred Shentu, Pieter Abbeel
ICRA 2017
Paper / Video

Framework for safely transferring policies learning in simulation to the real world.


I enjoy teaching and am grateful to have had the opportunity to teach a variety of different subjects in computer science.


Toronto CSC311: Introduction to Machine Learning: Winter 2023

Teaching Assistant

Toronto CSC311: Introduction to Machine Learning : Fall 2021 (head TA)
Toronto CSC2547: Introduction to Reinforcement Learning : Winter 2021
Toronto CSC421/2516: Neural Networks and Deep Learning : Winter 2019
Toronto CSC2541: Deep Reinforcement Learning : Fall 2018
UC Berkeley CS170: Introduction to Algorithms - Spring 2018
UC Berkeley CS189: Introduction to Machine Learning - Fall 2017
UC Berkeley CS189: Introduction to Machine Learning - Spring 2017
UC Berkeley CS70: Discrete Math and Probability - Fall 2016
UC Berkeley CS70: Discrete Math and Probability - Spring 2016

At Berkeley, I was awarded the Outstanding Graduate Student Instructor award.


I try to adhere to the principle "journey before destination" (Brandon Sanderson). Some activities I enjoy: basketball, running, exploring new places, reading, improv.

2020 Book Recs

Template from here!