BRIDGING CLASSIFICATION AND GENERATION: FROM NEURAL COLLAPSE INIMBALANCED CLASSIFICATION TO DISCRIMINATIVE-DRIVEN IMAGE GENERATION

The Hong Kong University of Science and Technology (Guangzhou)

Data Science and Analytics Thrust

PhD Thesis Proposal Examination

By Ms. Xuantong LIU

Abstract

Deep learning models have demonstrated remarkable success in both classification and generation tasks; meanwhile, their behavior under imbalanced data distributions and their ability to adapt to new tasks remain active areas of research. In this report, we address two
fundamental challenges: improving model performance under long-tailed distributions and
leveraging discriminative models for generative tasks. First, we explore the phenomenon
of Neural Collapse in deep learning and investigate its role in mitigating the challenges of
long-tailed classification. By explicitly inducing Neural Collapse during training, we show
that deep neural networks can achieve significantly improved discriminative performance on
imbalanced datasets. Second, we propose a novel, training-free framework for conditional
image generation by reversing discriminative vision-language models. Our approach demonstrates how these models can be effectively repurposed for generative tasks, bypassing the
need for additional training. These two lines of research offer complementary perspectives
on understanding and enhancing model behavior. Finally, we outline future work exploring
the design space of language models in image generation, motivated by the observation that
language models are optimized with classification objectives, providing a natural bridge to
generative tasks through their learned representations.

TPE Committee

Chair of Committee: Prof. YI, Ke

Prime Supervisor: Prof. YAO, Yuan

Co-Supervisor: Prof. ZHANG, Nevin L

Examiner: Prof. WANG, Wenjia