With its powerful visual-language alignment capability, CLIP performs well in zero-shot and few-shot learning tasks. However, we found in experiments that CLIP's logits suffer from serious inter-class confusion problems in downstream tasks, and the ambiguity between categories seriously affects the accuracy. To address this challenge, we propose a novel method called Logits DeConfusion, which effectively learns and eliminates inter-class confusion in logits by combining our Multi-level Adapter Fusion (MAF) module with our Inter-Class Deconfusion (ICD) module. Our MAF extracts features from different levels and fuses them uniformly to enhance feature representation. Our ICD learnably eliminates inter-class confusion in logits with a residual structure. Experimental results show that our method can significantly improve the classification performance and alleviate the inter-class confusion problem. The code is available at https://github.com/LiShuo1001/LDC.


    Access

    Download


    Export, share and cite



    Title :

    Logits DeConfusion with CLIP for Few-Shot Learning


    Contributors:
    Li, Shuo (author) / Liu, Fang (author) / Hao, Zehua (author) / Wang, Xinyi (author) / Li, Lingling (author) / Liu, Xu (author) / Chen, Puhua (author) / Ma, Wenping (author)

    Publication date :

    2025


    Remarks:

    CVPR 2025


    Type of media :

    Preprint


    Type of material :

    Electronic Resource


    Language :

    English




    Knowledge Distillation with Refined Logits

    Sun, Wujie / Chen, Defang / Lyu, Siwei et al. | ArXiv | 2024

    Free access

    Exploiting Features and Logits in Heterogeneous Federated Learning

    Chan, Yun-Hin / Ngai, Edith C. -H. | ArXiv | 2022

    Free access

    NormKD: Normalized Logits for Knowledge Distillation

    Chi, Zhihao / Zheng, Tu / Li, Hengjia et al. | ArXiv | 2023

    Free access

    Principal components analysis of the logits of the survivorship function

    Mitra, S. / Levin, M. L. / International Union for the Scientific Study of Population | British Library Conference Proceedings | 1993


    Adaptive Temperature Based on Logits Correlation in Knowledge Distillation

    Matsuyama, Kazuhiro / Anjum, Usman / Matsuyama, Satoko et al. | ArXiv | 2025

    Free access