UC San Diego

In recent years, neural networks have demonstrated remarkable breakthroughs in perceptual tasks such as computer vision and natural language processing, which achieve exceptional classification accuracy and robust generalization capabilities. My research primarily resides within the domain of deep learning, including two major dimensions:

Interpretability: Tabular data plays a crucial role as a primary source of structured information, serving as the foundation of decision-making across various fields, ranging from marketing and healthcare to government policy evaluation. In recognition of its significance, researchers have recently turned their attention to applying neural network models to tabular data due to their generally superior performance in comparison to rule-based methods. While neural networks excel in performance, they often act as "black-box" models, posing challenges for applications requiring human interpretability, including medical diagnosis and loan analysis. To address this issue, we propose a series of innovative paradigms aimed at generating human-readable predictions in tabular classification problems through novel neural network training approaches. This allows our interpretable models to leverage the high generalization capacity of neural networks.

Efficiency and Scalability: The continuous improvement in neural network performance has come at the cost of increased complexity, including higher storage requirements and computational demands. Despite the substantial processing power of modern hardware for training these models, real-time inference and energy consumption remain significant obstacles, particularly in mobile and wearable applications. To tackle this challenge, we propose to encode deep neural networks using a low-precision number representation, such that the models could achieve accuracy levels comparable to their full-precision counterparts. In addition, we introduce an approach that combines certain steps during the feed-forward phase by pre-computing various intermediate results, allowing the trained neural network to primarily operate in the low-precision domain with fewer floating-point operations.

By addressing these two critical aspects, my work contributes to the advancement of neural network applications in both high-stakes interpretability-sensitive domains and resource-constrained deployment scenarios.