Aprendizado autossupervisionado contrastivo orientado pela estrutura geométrica do espaço latente

Abstract

Self-Supervised Learning (SSL) has emerged as a powerful paradigm in computer vision, enabling models to learn meaningful feature representations directly from unlabeled data. Among SSL approaches, contrastive learning has gained particular prominence for its ability to induce discriminative embeddings by pulling together positive pairs and pushing apart negatives. However, random sampling of such pairs often disregards the underlying geometric structure of the latent space, leading to suboptimal representation quality and inconsistent class separation. To address this limitation, this work introduces Distance-Guided Contrastive Learning (DGCL), a self-supervised approach that systematically selects informative sample pairs based on their geometric configuration in the latent manifold. For each anchor sample, DGCL identifies the farthest intra-class examples (hard positives) and the nearest inter-class examples (hard negatives) through t-Distributed Stochastic Neighbor Embedding (t-SNE) projections. By iteratively refining these relationships across training cycles, the method progressively enhances intra-class compactness and inter-class separability. Experiments conducted on the CIFAR-10, FER-13, KDEF, and RAF-DB datasets demonstrate substantial improvements over conventional models trained without contrastive learning. The results reveal that DGCL yields geometrically consistent latent representations, characterized by reduced intra-class variance and well-structured semantic clusters.