Scholarlink

<p>Image classification is a fundamental task in computer vision with widespread applications. However, in real-world scenarios, visual systems often learn in an incremental manner, where category distributions shift over time and space. During long-term learning, models are constrained by storage limitations and data privacy concerns, and tend to suffer from catastrophic forgetting when past data cannot be stored. To tackle these challenges, we propose a calibrating generative replay (CGR) method for exemplar-free class-incremental learning. CGR includes the hybrid-conditional variational replay (HCVR) and Fourier kernel alignment module (FKAM). Specifically, HCVR utilizes a multi-conditional VAE to synthesize discriminative pseudo-features under semantic, geometric, and synthetic constraints for exemplar-free class reconstruction. FKAM further enhances the model’s robustness by aligning feature distributions across tasks using random Fourier features for efficient kernel approximations, thereby reducing representation drift. Extensive evaluations on CIFAR-100 and TinyImageNet across various incremental learning scenarios show that CGR outperforms existing state-of-the-art methods, demonstrating enhanced stability across extended task sequences. The code is publicly available at: <a href="https://github.com/Ethan2186/CGR">https://github.com/Ethan2186/CGR</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CGR: calibrating generative replay for exemplar-free class-incremental learning

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Xingcheng Zhu,
  • Kai Han,
  • Xiaocheng Hu,
  • Chongwen Lyu,
  • Jun Chen,
  • Yi Liu,
  • Zhe Liu

摘要

Image classification is a fundamental task in computer vision with widespread applications. However, in real-world scenarios, visual systems often learn in an incremental manner, where category distributions shift over time and space. During long-term learning, models are constrained by storage limitations and data privacy concerns, and tend to suffer from catastrophic forgetting when past data cannot be stored. To tackle these challenges, we propose a calibrating generative replay (CGR) method for exemplar-free class-incremental learning. CGR includes the hybrid-conditional variational replay (HCVR) and Fourier kernel alignment module (FKAM). Specifically, HCVR utilizes a multi-conditional VAE to synthesize discriminative pseudo-features under semantic, geometric, and synthetic constraints for exemplar-free class reconstruction. FKAM further enhances the model’s robustness by aligning feature distributions across tasks using random Fourier features for efficient kernel approximations, thereby reducing representation drift. Extensive evaluations on CIFAR-100 and TinyImageNet across various incremental learning scenarios show that CGR outperforms existing state-of-the-art methods, demonstrating enhanced stability across extended task sequences. The code is publicly available at: https://github.com/Ethan2186/CGR.