Automated Data Augmentation

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

Final Year Thesis Oral Defense

Title: "Automated Data Augmentation"

by

ANG Clyde Wesley Si

Abstract:

Automated Data Augmentation (AutoDA) is a new field that emerged as a way 
to remove researcher bias when choosing what transformations are applied 
to expand a dataset. In this paper, we present different ideas to enhance 
existing models. Image Specific Augmentation (ISA) aims to incorporate the 
information from input images when choosing the appropriate augmentations. 
Uniformly Sampled Augmentation (USA) applies the concept of uniform 
sampling to reduce the complexity of supernet models used in AutoDA. 
Varying Loss Function (VLF) is an ablation study that inspects the optimal 
loss function for the MODALS framework. The results of this paper showed 
that ISA was unsuccessful since choosing appropriate transformations is 
more effective than adding image-specific customization. Meanwhile, the 
USA model was unable to produce competitive results and lacked a clear way 
to produce an ordering of good augmentations. In contrast, VLF showed that 
the contrastive loss function can replace the triplet loss function used 
in the MODALS framework. The contrastive loss function has the most 
flexible nature to allow for better results on general tasks. Future 
research papers can use this project as a reference for the aforementioned 
ideas.


Date            : 5 May 2021 (Wednesday)

Time            : 15:00-15:40

Zoom Link:
https://hkust.zoom.us/j/93527078789?pwd=ZVhCckUwaGhQcVhnSk4vdEJVbjJJQT09

Meeting ID      : 935 2707 8789
Passcode        : 926017

Advisor         : Prof. YEUNG Dit-Yan

2nd Reader      : Dr. XU Dan