Semi-supervised Conditional Diffusion/Score-based Generative Models: Exploring Generative Potentials with Limited Supervision

Paul (Kuo-Ming) Huang (b08902072@ntu.edu.tw or paulhuang102701@gmail.com)

Three months (Feb. - Apr. 2024, Tentative)

Score-based generative models (SGMs) are the most popular family of generative models in recent years. Despite the success of conditional SGMs, we have uncovered disadvantages and potential room for improvements of such models under semi-supervised scenarios. The goal of this project is to understand the current baselines and explore several untouched ideas within this paradigm.

• Implement the first 2 stages of baseline [1] using a new framework [2] and conduct experiments on:
  • CIFAR-10
  • CIFAR-100
  • Imagenet 32×32 (If needed)
• Assist in conducting experiments on new ideas related to semi-supervised conditional SGMs on Imagenet.
  (Potentially under another framework & codebase)

[1] Diffusion Models and Semi-Supervised Learners Benefit Mutually with Few Labels, https://arxiv.org/abs/2302.10586, NeurIPS 2023
[2] Score-Based Generative Modeling through Stochastic Differential Equations, https://arxiv.org/abs/2011.13456, ICLR 2021

• Send an email to Hsuan-Tien Lin (htlin@csie.ntu.edu.tw) and cc the potential mentor above.
• Tell us about your background, why you are interested in the project, your expected workload, and your requested salary.
• If the position is still available, we'll contact you for an interview.