Research on Enhancing Classification and Coreset Selection with KNN Based Framework

In the field of machine learning, effective learning strategies play a critical role in enhancing model performance and efficiency. This paper utilizes Shapley entropy values to increase inter-class distances to enhance classification tasks. The proposed framework integrates these values with adversarial techniques and extra-intra object distance minimization strategies inspired by adversarial optimization and object distance refinement techniques. Shapley entropy provides a nuanced way of addressing the limitations of recent OoD(out of distribution) loss functions, which primarily rely on adversarial methods but struggle with complex inter-cluster relationships. By leveraging a knn(k-nearest neighbor) based approach, Shapley entropy effectively defines multiple inter-cluster relationships, improving class separation and intra-class cohesion in a more robust and scalable manner. Moreover, the shapley entropy demonstrates superior efficiency by simultaneously addressing coreset selection and enhancing classification performance, making it applicable across both tasks seamlessly. Experimental results include class-wise accuracy, t-SNE visualizations validated on CIFAR-10, and the application of active learning.