Gary Wu

We explore interpretable machine learning techniques to help stakeholders understand ML models and their decision-making processes. This work focuses on capturing the underlying reasoning of black box methods through various interpretability approaches.

We address the challenge of creating stable compute marketplaces for machine learning, where users can rent computational resources for training large models. Our work implements pricing and matching algorithms that handle multi-dimensional preference reporting in practice.

We present PRIMES, a novel incentive-aligned approach for federated learning that addresses the misreporting and free-riding problem. By selecting clients based on next-step loss and paying them based on global performance, we incentivize truthful reporting and high-quality contributions.

We implement a homomorphic encryption-based neural network using polynomial activation approximation to enable secure inference on encrypted data. We compare the security posture against standard unencrypted networks, focusing on membership inference, model extraction, and gradient-based input recovery attacks.

We investigate how fairness breaks down in online learning settings with strategic gaming, where agents may modify their reported features for favorable outcomes. We introduce algorithmic changes to ensure fairness across multiple metrics, accounting for different cost functions between advantaged and disadvantaged agents.

We explore the feasibility of Large Language Model routers that can intelligently recommend the best LLM for a given task. Using KMeans clustering on prompts, we build a model to predict which LLM will provide the best response.