Label-efficient Learning by Exploiting Unlabeled Data with Higher-quality Supervision and Wider Applicability

PhD Thesis Proposal Defence


Title: "Label-efficient Learning by Exploiting Unlabeled Data with 
Higher-quality Supervision and Wider Applicability"

by

Miss Huimin WU


Abstract:

This thesis seeks to explore label-efficient learning techniques that aim to 
mitigate the reliance on large-scale human labels in training deep learning 
models. Our focus is on developing strategies that leverage unlabeled data, 
which are often redundant, and gradually decrease the number of human labels 
required for training predictive models. Specifically, our investigation covers 
semi- supervised learning (with 20% labeled data), barely-supervised learning 
(with only a few labeled data points), and self-supervised learning that 
requires no human labels at all. However, the quality of the constructed 
supervision for the purpose of exploiting unlabeled data is the primary 
challenge we face. As such, our primary objective is to construct 
higher-quality supervision that is more expressive, accurate, and generic. 
Moreover, the current state-of-the-art techniques are limited in terms of their 
scope of use. Thus, we endeavor to enhance the range of applicability of the 
constructed method.

Firstly, we enhance semi-supervised medical image segmentation by seeking more 
expressive forms of supervision without the aid of modality or task priors. 
Traditional methods of supervising segmentation are often limited to one-hot 
vectors or their soft variants. Alternatively, we propose utilizing contrastive 
loss to train a more compact and better-separated feature space. This approach 
is more robust to noise and exhibits better generalization abilities for test 
data. Furthermore, our proposed method is not reliant on specific modalities or 
tasks, making it more adaptable to diverse application scenarios.

Then, we reduce the number of human labels and explore the barely supervised 
learning setting, which has received limited attention. This setting is 
characterized by the presence of only a few labeled data points, making it 
challenging to achieve high-quality supervision. Standard semi-supervised 
methods constructed with fewer human annotations often produce unsatisfactory 
results. Additionally, prior research on barely supervised learning has limited 
applicability due to the assumption of structural similarity between the data 
as the supervision. To address these challenges, we propose an online 
confidence thresholding technique that can generate more accurate pseudo 
labels. Without resorting to ground truth similarity, this algorithm can be 
applied to a broader range of realistic segmentation problems.

Thirdly, we explore the topic of self-supervised learning, a type of learning 
that does not rely on human labels. Instead, it utilizes data as a form of 
supervision. Presently, pre-training strategies lack data generality. 
Therefore, the objective of this work is centered around finding data-generic 
supervision capable of being applied to any data modality. To achieve this 
objective, we propose randomized quantization as contrastive learning 
augmentation. Our method has demonstrated better performance than prior 
data-agnostic self-supervised techniques. We have validated its effectiveness 
over a vast range of data modalities, including vision, audio, 3D point clouds, 
and DABS, a public benchmark for data-agnostic self-supervised learning.

In our last work, we enhance the downstream application generality of 
self-supervised learning techniques. Current self-supervised learning 
techniques are mainly designed for semantic tasks. In order to expand the 
application scenarios of these methods, we propose to adapt general-purpose 
large-scale pre-trained models on natural videos to multi-view geometrical 
tasks with an empirical study on optical flow estimation. Unlike previous flow 
estimation methods that rely on complicated architectural components 
specialized to geometry tasks, our overall architecture does not have any 
task-specific inductive bias, which significantly simplifies the architectural 
design. The strong performance validates its effectiveness.


Date:                   Thursday, 30 May 2024

Time:                   2:00pm - 4:00pm

Venue:                  Room 5510
                        Lifts 25/26

Committee Members:      Prof. Tim Cheng (Supervisor)
                        Dr. Xiaomeng Li (Supervisor)
                        Dr. Xiaojuan Ma (Chairperson)
                        Dr. Shuai Wang