Pendar Alirezazadeh : “Angular margin-based softmax losses: toward discriminative deep metric learnig”

Zuzendariak_Directores:  Abdelmalik Moujahid Moujahid/ Fadi Dornaika

2024_04_25; 11: 00  Sala Ada Lovelace aretoa.

Abstract:

"Discriminative deep metric learning aims to construct an embedding space in which instances of the same class can be grouped together while being effectively distinguished from instances belonging to other classes by deeply learned representations. In this context, angular deep metric learning emerges as a specialized subset of discriminative deep metric learning, which is characterized by focusing on the angles between the feature vectors rather than their magnitudes.

Classical methods such as ArcFace and CosFace are considered pioneers in the field of angle-dependent metric learning as they introduce angle-dependent margins into the softmax loss function. This strategic approach aims to promote more coherent clustering within classes while achieving greater angular separation between different classes. These methods have been applied specifically in the context of face recognition.

This thesis presents several research contributions consisting of novel softmax loss functions based on angular margins. The first contribution is to extend the applicability of these loss functions beyond the field of face recognition. The effectiveness of these functions is investigated in challenging contexts with limited labeled data. Topics such as fashion image retrieval, fashion style recognition and classification of histopathologic breast cancer images are covered.

In the area of fashion image retrieval, Discriminative Margin Loss (DML) is proposed to investigate the adjustment of margin penalties for positive and negative classes. The underlying goal is to improve the discriminative power of the learned embeddings specifically for fashion image retrieval.

For the challenges related to fashion style and face recognition, Additive Cosine Margin Loss (ACML) is introduced. ACML simplifies the fine-tuning of margin penalties while strengthening the separation between classes and the cohesion within classes. This approach leads to performance improvements in these special recognition tasks.

The Boosted Additive Angular Margin Loss (BAM) method is proposed for the field of breast cancer diagnosis using histopathological images. BAM not only penalizes the angle between the deep feature and its corresponding weight from the target class, but also considers the angles between deep features and their corresponding weights from non-target classes. This approach aims to facilitate the detection of highly discriminative features for accurate diagnosis while improving the intra-class cohesion and increasing the inter-class discrepancy to take advantage of margin constraints.

Overall, these new loss functions, including DML, ACML and BAM, contribute significantly to the field of softmax losses based on angular margins by extending their application beyond conventional constraints. With this expanded scope, these loss functions are able to effectively address the distinct challenges inherent in diverse domains.

The proposed loss functions have been rigorously tested and validated in various studies addressing data limitations. The robustness of the results was demonstrated through various metrics, feature visualizations and statistical analysis. It is important to mention that these loss functions have improved the performance of numerous deep learning architectures. The superiority of these loss functions over various complex feature-based architectures containing significant parameters has been confirmed through extensive experiments and comparisons with angular margin-based losses on various benchmark datasets. In addition, the performance of the models was evaluated against other methods on large datasets to obtain a comprehensive assessment of their capabilities.

Keywords:
Discriminative Deep Metric Learning, Loss Function, Angular Margin-based Softmax Loss,
Cross-Domain Fashion Retrieval, Deep Fashion Style Recognition, Deep Face Recognition,
Breast Cancer Classification."