Research  ·  2022–2024
‍


A Comprehensive Analysis of Fine-Tuning Strategies for BERT · CS 224N Final Project
Co-authors: Adam Chun, Emily Nguyen

In recent years, BERT has achieved breakthrough results in the field of language-modeling. Although it is pre-trained on a large corpus of data, fine-tuning BERT on a specific downstream task may be time-consuming and require significant experimentation. Our work aims to expedite this process by providing holistic insights on best practices for fine-tuning BERT. This study examines three major fine-tuning methods: (1) smoothness-inducing adversarial regularization, (2), cosine-similarity fine-tuning, and (3) multitask fine-tuning with gradient surgery.


Multi-Agent Path-Finding as a Markov Game · CS 238 Final Project
Co-authors: Enok Choe

In the Multi-Agent Path Finding (MAPF) problem, multiple agents each with different start states must navigate to their respective end states (destinations) without colliding with each other. In this paper, we frame MAPF as a Markov Game (MG) to model multiple agents with their own reward functions. Under this framework, transition probabilities depend on the joint action and each agent is trying to maximize their own reward. We implement fictitious play and gradient ascent as the algorithms to update the policies of the agents of the MG.
‍