Contrastive Learning for Small Datasets Without Data Augmentation
MSc Thesis Proposal by: Tasmia Islam
Date: May 28, 2026
Time: 12 pm
Location: Online via Microsoft Teams
Abstract:
Deep learning models typically require large amounts of labeled data to achieve strong performance. However, in many real-world domains, such as medical imaging and manufacturing defect detection, only limited data samples are available. Existing approaches, including transfer learning, few-shot learning, and supervised contrastive learning, often rely heavily on large datasets or data augmentation techniques. Although data augmentation can improve model generalization, it may also alter the semantic meaning of the original data in certain applications. To address this limitation, this research proposes an augmentation-free contrastive learning approach that utilizes out-of-domain (OOD) data as a source of negative samples. Instead of generating augmented views of the same data, the proposed method introduces naturally distinct OOD samples to create a stronger and semantically consistent contrastive signal. In this study, a subset of the MNIST dataset is used as the in-domain (IND) dataset, while Fashion-MNIST is used as the OOD dataset. The proposed framework employs a ResNet-18 architecture within a two-stage training pipeline. In the first stage, feature representations are learned through contrastive learning, and in the second stage, the learned representations are evaluated using a downstream classifier. The primary objective of this work is to demonstrate that meaningful and well-structured feature representations can be learned from small datasets without relying on data augmentation or large-scale pretraining.
Keyword: Contrastive Learning, Small Dataset, Out-of-Domain (OOD) Data, Representation Learning, Image Classification
Thesis Committee:
Internal Reader: Dr. Muhammad Asaduzzaman
External Reader: Dr. Mehdi Sangani Monfared
Advisor: Dr. Dan Wu