OptiVAE: A Unified Parallel Gumbel-Softmax VAE Framework with Performance-Based Tuning

OptiVAE: A Unified Parallel Gumbel-Softmax VAE Framework with Performance-Based Tuning

Presenter(s)

Fangshi Zhou

Comments

Presentation: 11:00-11:20, Kennedy Union 222

Description

Classic training algorithms for Gumbel Softmax Variational Autoencoders (GS-VAEs) often rely on an annealing scheme, which reduces the Softmax temperature according to a given function. We find that this leads to suboptimal performance. To improve the design, we propose a novel framework for GS-VAEs, which embraces dual latent layers and a parallel multi-model structure with diverse temperature strategies. By dynamically tuning the temperature in response to the loss difference between each sub-model and the best sub-model with the minimum loss at each training epoch, our model utilizes exploration and exploitation and significantly surpasses a standard GS-VAE in data reconstruction, detection of altered data, and model robustness. In particular, our model can reconstruct data of unfamiliar categories that are never observed during training. Moreover, in the presence of patch attack or white-box adversarial attack, our model greatly outperforms a standard GS-VAE and other existing models studied in this work.

Publication Date

4-17-2024

Project Designation

Graduate Research

Primary Advisor

Luan V. Nguyen, Zhongmei Yao, Tianming Zhao

Primary Advisor's Department

Computer Science

Keywords

Stander Symposium, College of Arts and Sciences

Institutional Learning Goals

Scholarship; Faith; Scholarship

OptiVAE: A Unified Parallel Gumbel-Softmax VAE Framework with Performance-Based Tuning

Share

COinS