Stable Remaster: Bridging the Gap Between Old Content and New Displays
arXiv preprint 2023
Nathan Paull, Shuvam Keshari, Yian Wong

Exploring Advanced Neural Network Architectures for Synthetic Well Log Generation
American Rock Mechanics Association Symposium 2022
Yian Wong, Sau-Wai Wong

Using Convolutional Neural Networks and Monte-Carlo Dropout to Generate Synthetic Well Logs with Accurate Uncertainty Estimation
International Symposium for Geotechnical Safety & Risk 2022
Yian Wong, Sau-Wai Wong


Integrated System for Exploring Multiple Perspectives to Mitigate Cognitive Biases [demo]
Generative AI
Part of my Master’s thesis, in a research project sponsored by Microsoft Research. This project enables users to make better decisions about difficult topics. Given a difficult topic, we create various personas, and ask LLMs to roleplay as these personas, each contributing a different perspective about the topic in a debate-style format. The goal is to mitigate user biases.

Interpretable Robot Locomotion by Monte Carlo Tree Search and Self-Supervised Learning [paper]
Computer Vision, Reinforcement Learning
Our team implemented MuZero and applied it to the problem of robot locomotion, highlighting the complexity of this control task and the limitations of model-based RL in this domain. We implement a Variational Auto-Encoder to encode pixel observations into a compact latent state. We also proposed potential solutions to enable MuZero to generalize to continuous control tasks in this domain.

Probabilistic Circuits for Causal Discovery and Model-Based Reinforcement Learning [paper] [slides] [talk]
Reinforcement Learning, Causality, Probabilistic Modeling
For my class in Causal Reinforcement Learning, I explored Sum Product Networks (SPNs), and their ability to generalize while succinctly exposing internalized causal relationships that were discovered in deep reinforcement learning. I showed that SPNs are comparable to neural networks in reinforcement learning settings, while providing us greatly improved interprability from a causal perspective.

TwitchMoji [paper]
Natural Language Processing
We fine-tune the BART language model to accurately detect the sentiment. We do this by training it to predict what Twitch emojis were used in Twitch chat messages. By training on billions of chat instances, our fine-tuned BART model is able to better adapt to other sentiment analysis tasks.

Adversarial Training for Improving Compositional Understanding of Vision-Language Models [paper]
Natural Language Processing, Computer Vision
In this project, we tackle the problem of grounding natural language. One important benchmark in this field is the Winoground Challenge, which tasks vision-language models to correctly match similar sentence-image pairs. We propose an adversarial learning adjustments to help improve compositional understanding of the CLIP model.

ConnectFourRL [code]
Computer Vision, Reinforcement Learning
I developed an off-policy actor critic algorithm and a residual network for the game of ConnectFour, which was later enhanced with Monte-Carlo Tree Search to enable better decision-making through strategic planning. By playing against itself, the algorithm continually improved its performance. Using my implementation, players can test their skills against the AI and observe its capabilities in various situations.

Reinforcement Learning in SuperTuxKart [paper]
Computer Vision, Reinforcement Learning
Our team implemented asynchronous actor critic with a residual network for the game of SuperTuxKart and identified its sample inefficiency. To address this issue, we proposed an imitation learning framework based on DAgger, demonstrating the ability to accurately imitate expert behavior. Additionally, we explored methods for extrapolating beyond the performance of the expert agent.”