G -Softmax: Improving Intraclass Compactness and Interclass Separability of Features.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Intraclass compactness and interclass separability are key metrics for evaluating feature discriminability in machine learning models.
• Convolutional networks learn features, but optimizing these metrics remains a challenge for improving model performance.

Purpose of the Study:

• To investigate intraclass compactness and interclass separability of features learned by convolutional networks.
• To propose a novel Gaussian-based softmax (G-softmax) function to enhance these crucial feature properties.

Main Methods:

• Developed and implemented a Gaussian-based softmax (G-softmax) function as a replacement for the standard softmax function.
• Evaluated the G-softmax function on diverse datasets including CIFAR-10, CIFAR-100, Tiny ImageNet, MS COCO, and NUS-WIDE.

Main Results:

• The G-softmax function significantly improved state-of-the-art model performance across all tested classification and multi-label classification datasets.
• Analysis confirmed that G-softmax enhances both intraclass compactness and interclass separability compared to the standard softmax.
• A strong linear correlation was observed between high intraclass compactness, interclass separability, and improved average precision on MS COCO and NUS-WIDE datasets.

Conclusions:

• The proposed G-softmax function effectively improves feature discriminability by enhancing intraclass compactness and interclass separability.
• The improvements in feature discriminability directly translate to superior performance in image classification tasks.
• Optimizing intraclass compactness and interclass separability is a viable strategy for boosting model accuracy, particularly in multi-label classification scenarios.