Towards Trustworthy Models in Machine Learning
Xiaoyu Liu
Tuesday, April 30, 2024, 12:30-3:00 pm 
Abstract
Abstract:
In recent years, the proliferation of deep learning techniques, exemplified by large language models (LLMs), has fueled remarkable advancements across various domains. However, the complexity and opacity inherent in these models often impede our ability to comprehend and trust their decision-making processes. In response, this proposal aims to tackle the critical challenges of interpretability and trustworthiness within neural networks.
Our study is structured around two primary objectives. First, we seek to deepen our understanding of neural network structures from a tensor perspective. Second, we aim to optimize feature representations in the latent space using causal inference methods.
To achieve the first objective, we employ tensor methods to scrutinize and refine the design of multi-head self-attention mechanisms—the cornerstone of transformer architectures and LLMs. Our analysis reveals that the current attention mechanisms reside within a vast design space, as depicted by tensor diagrams, thereby unveiling opportunities for more efficient transformer designs.
In addressing the second objective, we leverage causal inference techniques to quantify and improve latent feature representations. Specifically, we (a) establish causal relationships between latent variables and predictive objectives to reduce spurious correlations and (b) incorporate external knowledge to obtain causal disentanglement in the latent space by modeling the confounders.
Our study is structured around two primary objectives. First, we seek to deepen our understanding of neural network structures from a tensor perspective. Second, we aim to optimize feature representations in the latent space using causal inference methods.
To achieve the first objective, we employ tensor methods to scrutinize and refine the design of multi-head self-attention mechanisms—the cornerstone of transformer architectures and LLMs. Our analysis reveals that the current attention mechanisms reside within a vast design space, as depicted by tensor diagrams, thereby unveiling opportunities for more efficient transformer designs.
In addressing the second objective, we leverage causal inference techniques to quantify and improve latent feature representations. Specifically, we (a) establish causal relationships between latent variables and predictive objectives to reduce spurious correlations and (b) incorporate external knowledge to obtain causal disentanglement in the latent space by modeling the confounders.
Examining Committee
Bio
Xiaoyu Liu is a Ph.D. student in the Department of Computer Science at the University of Maryland, College Park. She is advised by Dr. Furong Huang. Her research broadly encompasses model understanding, interpretability and trustworthy machine learning, especially from the causal perspective.
This talk is organized by Migo Gui