Deep Feature Disentanglement for Supervised Contrastive Learning: Application to Image Classification

Résumé fourni par la source

Abstract In machine learning, deep metric learning from original data is essential, with supervised contrastive learning being a notable approach. This method aims to form a deep feature space where similar samples from the same class are clustered together, while dissimilar samples from different classes are separated. However, a common limitation of contrastive learning methods is that they utilize the entire feature space for data embedding and often neglect the within-class variability. To overcome this limitation, we propose a novel supervised contrastive learning method that decomposes deep features into two distinct components: common features, which encapsulate the essential, class-defining characteristics, and style features, which capture the within-class variability and nuanced differences. Additionally, we enhance this approach by introducing an overlapping field that synergistically integrates elements from both feature spaces, enabling a more comprehensive and robust feature representation. Our experiments with different image datasets and deep encoders, including CNNs and transformers, show that our approach outperforms traditional single-feature contrastive methods. On the CIFAR100 and PASCAL VOC databases, traditional supervised contrastive learning achieved accuracy rates of 75.5% and 51.41%, respectively, while our method improved them to 77.81% and 59.38%, respectively. We present an algorithm for deep contrastive learning that utilizes two feature spaces: one for encoding common class features and another for capturing within-class variability. This is achieved by partitioning the features of the last layer of the encoder into (i) a common field and (ii) a style field. Our loss function contrasts the common features while summarizing the style features within the same class so that the style field can capture the intra-class variability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.