Data Augmentation with Diffusion Model

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Data Augmentation with Diffusion Model"

by

XIAO Hanyu

Abstract:

Due to its complexity, heterogeneity, and inter-temporal nature, detecting 
anomalies in multivariate time series data is never an easy task. I propose 
using the diffusion model to extract the normal pattern of data. Memory modules 
are introduced as a means of over-generalization prevention, which serves to 
help detect minor deviations from normal patterns. Loss from diffusion and 
memory modules are used to determine whether a data point is anomalous.


Date            : 2 May 2024 (Thursday)

Time            : 14:00 - 14:40

Venue           : Room 5506 (near lifts 25/26), HKUST

Advisor         : Prof. KWOK James Tin-Yau

2nd Reader      : Dr. XU Dan