Constrained Center Loss for Convolutional Neural Networks.

Area of Science:

• Computer Science
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) transform input data into feature vectors for classification.
• Standard softmax loss in CNNs often neglects the compactness of features within the same class.
• This limitation can hinder the robustness and discriminative power of learned features.

Purpose of the Study:

• To propose a constrained center loss (CCL) algorithm for more robust feature extraction in CNNs.
• To enhance the intra-class compactness of feature representations.
• To improve the overall performance of CNNs in classification tasks.

Main Methods:

• A novel training objective combining softmax loss and CCL is introduced.
• CCL aims to cluster feature vectors from the same class together.
• An alternative learning strategy updates CNN weights and cluster centers iteratively, distinct from standard stochastic gradient descent (SGD).
• A simplified CCL (SCCL) algorithm is also proposed.

Main Results:

• Experiments on six benchmark datasets demonstrate the effectiveness of the proposed methods.
• The CCL-based algorithm significantly improves feature representation by enforcing intra-class compactness.
• Both CCL and SCCL algorithms outperform several existing state-of-the-art approaches.

Conclusions:

• The proposed CCL and SCCL algorithms offer a more effective approach to feature learning in CNNs.
• Enforcing intra-class compactness alongside class separation leads to more robust feature extraction.
• The alternative learning strategy provides an efficient way to train CNNs with CCL.