Exploring SSL Enhancements Through Triplet-Based Representation Learning

In the research area of machine learning, self-supervised learning (SSL) strategies have demonstrated substantial promise in harnessing the vast pools of unlabeled data. This study presents the Barlow Triplets method, an SSL strategy that builds on the Barlow Twins framework to address its limitations in complex and diverse datasets. Our method enhances representation learning by introducing an additional augmented view through a triplet-based mechanism, which refines feature interactions and balances redundancy reduction with feature invariance. We employed an adapted loss function and validated our approach through experiments on the CIFAR10 dataset using a KNN classifier. Results show that Barlow Triplets surpasses traditional SSL methods, including Barlow Twins predecessor, across accuracy, precision, recall, and F1-scores, demonstrating its potential for practical applications and scalability. We compare two key formulations of Barlow Triplets, one of which is reported here for the first time. Our method requires fewer computational resources while achieving commendable performance. This research aims to contribute to SSL by introducing a method that seeks to optimize performance while potentially maintaining computational efficiency.