复现研究质疑LICO模型的可解释性提升效果
Reproducibility study of "LICO: Explainable Models with Language-Image Consistency"
- 用ResNet和经典方法复现LICO，验证其语言图像一致性设计
- 未复现原文宣称的分类性能提升与可解释性改进结果
- 提醒可解释性研究需加强实验透明度与严谨评估
机器学习领域的可复现性危机凸显了对研究成果审慎检验的必要性。本文针对Lei等人（2023）提出的LICO方法展开全面复现研究，该方法旨在通过视觉-语言模型的语言监督增强特征表示并指导学习过程，以提升后处理可解释性技术并改善图像分类性能。我们采用(Wide) ResNets及Grad-CAM、RISE等成熟可解释性方法进行复现，但未能一致复现原作者报告的结果。具体而言，未观察到LICO在分类性能上的持续提升，也未发现其在定量与定性可解释性指标上的显著改善。研究结果强调了在可解释性研究中严谨评估与透明报告的重要性。
原文摘要 · Abstract (English)
The growing reproducibility crisis in machine learning has brought forward a need for careful examination of research findings. This paper investigates the claims made by Lei et al. (2023) regarding their proposed method, LICO, for enhancing post-hoc interpretability techniques and improving image classification performance. LICO leverages natural language supervision from a vision-language model to enrich feature representations and guide the learning process. We conduct a comprehensive reproducibility study, employing (Wide) ResNets and established interpretability methods like Grad-CAM and RISE. We were mostly unable to reproduce the authors' results. In particular, we did not find that LICO consistently led to improved classification performance or improvements in quantitative and qualitative measures of interpretability. Thus, our findings highlight the importance of rigorous evaluation and transparent reporting in interpretability research.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。