About Me


My name is Stephen Chung, who graduated from the University of Massachusetts Amherst with a master's degree in 2021. My primary research interest includes reinforcement learning (RL), biologically-inspired machine learning, and deep learning.

During my master's years, I was fortunate to be guided by Professor Andrew G. Barto, a pioneer in RL. Under his guidance, I studied methods to train a deep neural network without backpropagation efficiently. Despite being used in almost all deep learning methods, backpropagation is generally regarded as being biologically implausible. A more biologically plausible way of training a deep network is through treating each unit in the network as an RL agent that receives the same reward signal, but the learning speed of this method is very slow. Therefore, I investigated methods that can speed up learning while retaining the biological plausibility of this learning method. To this end, I proposed two novel algorithms: MAP Propagation (Chung, 2021) and Weight Maximization (Chung, 2022). I argue that both algorithms, built on RL methods, are more biologically plausible than backpropagation while maintaining a similar learning speed. As such, these algorithms may shed light on biological learning and substitute backpropagation in training deep learning models. My paper on MAP Propagation was recently accepted in NeurIPS 2021, and my paper on Weight Maximization was recently accepted in AAAI 2022. You can read more about this exciting research here!

In addition, I also worked with Professor Hava Siegelmann on the theoretical capability of recurrent neural networks (RNNs). My research on the theoretical capability RNNs led to our paper on the Turing-completeness of RNNs, which was recently accepted in NeurIPS 2021 (Chung & Sieglemann, 2021). In this paper, we proved the sufficient conditions for an RNN to be Turing-complete and demonstrate how to simulate a Turing machine with an RNN. This work thus allows the construction of an RNN that runs any algorithms without prior training and extends the fundamental theories on the computational power of RNNs. I also studied the training methods of spiking neural networks under the guidance of Professor Robert Kozma (Chung & Kozma, 2020).

As for my interest, I love reading Western and Chinese philosophy books, such as Zhuangzi and Nietzche. I enjoy thinking about the world and philosophical questions. I also like playing tennis and hiking!

Research Interest

  • Reinforcement learning (RL): Multi-agent RL, model-based RL, and hierarchical RL;
  • Biologically-inspired machine learning: Spiking neural networks, and biologically-inspired learning rules;
  • Deep learning: Alternatives to backpropagation, theories of deep neural networks, neural ODEs.


Research Experience

Research on Reinforcement Learning with Professor Andrew Barto

Jan 2020 - Present
University of Massachusetts Amherst
  • Investigated methods to train a deep artificial neural network (ANN) without backpropagation based on RL methods;
  • Proposed two novel algorithms called MAP Propagation (Chung, 2021) and Weight Maximization (Chung, 2022) that can train a deep ANN to solve standard RL tasks efficiently without any backpropagation;
  • Proved the theoretical properties of both algorithms and conducted various experiments;
  • Successfully published a sole-author paper on MAP Propagation in NeurIPS 2021;
  • Successfully published a sole-author paper on Weight Maximization in AAAI 2022.

Research on Biologically-Plausible ML with Professor Hava Siegelmann

Aug 2020 - Aug 2021
BINDS Lab at The University of Massachusetts Amherst
  • Prepared part of the proposal for a grant from DARPA on applying AI in intelligent hardware simulation, which was recently accepted;
  • Proposed new theorems and constructed proof on the Turing completeness of recurrent neural networks;
  • Successfully published a paper as the co-first author on the Turing completeness of recurrent neural networks in NeurIPS 2021 (Chung & Sieglemann, 2021).

Research on Spiking Neural Networks with Professor Robert Kozma

Jan 2020 - Jun 2020
University of Massachusetts Amherst
  • Investigated methods to train a spiking neuron network to solve RL tasks using STDP-based learning rules;
  • Proposed a novel learning rule called feedback-modulated TD-STDP that can train a spiking neuron network to solve RL tasks efficiently (Chung & Kozma, 2020).


M.S. in Computer Science

Aug 2019 - Jun 2021
University of Massachusetts Amherst
  • Cumulative GPA: 4.0/4.0
  • Artificial Intelligence: Machine Learning, Reinforcement Learning, Computer Vision, Natural Language Processing
  • Computational Neuroscience: Neurodynamic, Spiking Neural Network
  • Miscellaneous: Quantum Information Systems, Advanced Algorithms

B.S. in Quantitative Finance

Sep 2012-Jun 2016
The University of Hong Kong
Hong Kong
  • Mathematics & Statistics: Real Analysis, Linear Algebra, Multivariable Calculus, Probability Modeling, Statistical Inference, Stochastic Calculus, Markov Chain
  • Computer Science: Computer Programming, Algorithms, Spreadsheet Finance Modeling
  • Finance: Corporate Finance, Investment, Derivatives, Financial Statement Analysis, Mathematical Finance

Exchange Student

London School of Economics and Political Science
United Kingdom