Improving LLM planning and reasoning in cross-domain benchmarks
Amazon AGI Autonomy team
- Robustly improved LLM planning and reasoning in cross-domain benchmarks including science/math, day-to-day planning and games, with flexible tree-search and self-play methods.
Thesis projects:
Learning How Humans Play Board Games with GPT-4IAR [AAAI workshop]
Wei Ji Ma lab + Luigi Acerbi lab, Department of Computer Science, University of Helsinki+ NYU (2023.6 — Present)
- Trained a GPT model on a large mobile game dataset to understand and predict different characteristics of human gameplay.
Compare planning between AI and humans [Talk]
Wei Ji Ma lab, Center for Neural Science, NYU (2021.5 — Present)
- Trained Deep Reinforcement learning models (AlphaZero type agents) to play a planning task of intermediate complexity.
- Analyzed features learned by the trained Deep Reinforcement learning networks.
- Studied the learning and planning mechanisms of AlphaZero agents and compared those to a human planning model
Improve the efficiency of an unbiased log-likelihood estimation method
Luigi Acerbi lab, Department of Computer Science, University of Helsinki (2021.3 — present)
- Compared the efficiencies of log-likelihood estimations in different models using Inverse Binomial Sampling with different
allocation methods.
- Develop a toolbox for a more efficient Inverse Binomial Sampling method that can estimate the log-likelihood unbiasedly.
The latent factors of complex planning decisions
Wei Ji Ma lab, Center for Neural Science, NYU (2020.10 — Present)
- Coded a battery of 9 planning tasks and cognitive tasks to run a large web-based online study.
- Investigated the individual differences and latent structure of planning decisions and modelled how planning is related to basic cognitive abilities.
Other projects:
Using Artificail Neural Networks to Approximate Bayesian Inference [Paper]
Computational Cognitive Modelling final project, Center for Data Science, NYU (2021.2 — 2021.5)
- Trained artificial neural networks (ANNs) on a task that requires inductive reasoning and found that ANNs can perform these tasks using Bayesian-like strategies without the need for an explicit computation of the log likelihood
“Double Descent” and Deep Gaussian Processes
Bayesian Machine Learning final project, Center for Data Science, NYU (2020.10 — 202012)
- Investigated the “Double Descent” phenomenon in Deep Gaussian Processes by looking into performance of testing error as a function of model depth/width/kernel complexity
Hidden Markov Models
Eero Simoncelli Lab, Center for Neural Science, NYU (2020.2 - 2020.9)
- Used Hidden Markov Model to model the context-dependent representations of Visual Cortex.
Automatic behavioral tracking
Christine Constantinople lab, Center for Neural Science, NYU (2019.9 - 2019.12)
- Developed behavioral analysis pipeline by training a Convolutional Neural Network to automatically track head angles of rats
Large neural population analysis
Takaki Komiyama Lab, UCSD (2017.1 — 2019.6)
- Analyzed large-scale activities of over 8000 neurons and investigated the source of information segregation for motor control