学术报告：邵昊 武汉大学

主讲人：邵昊 武汉大学

题 目：Linearity-Enhanced Logits for Image Classification with Softmax Optimization

时间：2025年10月18日 15:50-17:20

地点：VSport体育官网新校园 B513

Recently, a prominent research direction in image classification has been the integration of softmax loss with well-established convolutional neural networks (CNNs) to achieve remarkable performance. However, the traditional softmax loss fails to provide sufficient discriminability for multi-classification tasks, which in turn hinders the target network from effectively learning the embedded features. To address this limitation, we propose Orthogonal-Softmax to effectively learn the embedded features during the training phase. The key innovation lies in replacing the Taylor-approximated logit with a more robust linear relationship derived from orthogonal polynomial expansion. Furthermore, we introduce Orthogonal-M, a margin-based extension of Orthogonal-Softmax, for maximizing inter-class Sepa-rability while minimizing intra-class variance. Experimental results on several benchmark datasets validate the superiority of our approach over existing methods.