Scholarlink

<p>Unsupervised anomaly detection has achieved progress with continuous memory representations, which store normal patterns in learnable grids via differentiable interpolation. However, existing methods (e.g., CRAD) employ static update rates, leading to overfitting and poor generalization in few-shot scenarios where available normal samples are scarce. To address this, we propose CRAD-HOPE, a brain-inspired nested learning framework. Our motivation is to address the <i>stability-plasticity dilemma</i>: effectively adapting to new few-shot samples without forgetting the general manifold of normality. Inspired by the brain’s Complementary Learning Systems (CLS), we design a Multi-Frequency Nested Grid architecture. It consists of a “Fast Grid” (simulating the hippocampus) for rapid adaptation to new patterns and a “Slow Grid” (simulating the neocortex) for stable long-term retention. A Grid Knowledge Distillation mechanism further consolidates information from fast to slow grids to help prevent catastrophic forgetting. Extensive experiments on MVTec AD and OCT2017 demonstrate that CRAD-HOPE achieves state-of-the-art performance, improving 1-shot detection by 4.1% I-AUROC over the baseline and achieving superior pixel-level localization (94.7% P-AUROC) with 9.5<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> fewer parameters than VLM-based approaches. Code is available at <a href="https://github.com/tiankongkongkongruye/CRAD_HOPE">https://github.com/tiankongkongkongruye/CRAD_HOPE</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

CRAD-HOPE: brain-inspired nested learning framework for few-shot anomaly detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月7日

查看全文
  • Wenxin Cao,
  • Juanhua Cao,
  • Weijun Wu

摘要

Unsupervised anomaly detection has achieved progress with continuous memory representations, which store normal patterns in learnable grids via differentiable interpolation. However, existing methods (e.g., CRAD) employ static update rates, leading to overfitting and poor generalization in few-shot scenarios where available normal samples are scarce. To address this, we propose CRAD-HOPE, a brain-inspired nested learning framework. Our motivation is to address the stability-plasticity dilemma: effectively adapting to new few-shot samples without forgetting the general manifold of normality. Inspired by the brain’s Complementary Learning Systems (CLS), we design a Multi-Frequency Nested Grid architecture. It consists of a “Fast Grid” (simulating the hippocampus) for rapid adaptation to new patterns and a “Slow Grid” (simulating the neocortex) for stable long-term retention. A Grid Knowledge Distillation mechanism further consolidates information from fast to slow grids to help prevent catastrophic forgetting. Extensive experiments on MVTec AD and OCT2017 demonstrate that CRAD-HOPE achieves state-of-the-art performance, improving 1-shot detection by 4.1% I-AUROC over the baseline and achieving superior pixel-level localization (94.7% P-AUROC) with 9.5 \(\times \) × fewer parameters than VLM-based approaches. Code is available at https://github.com/tiankongkongkongruye/CRAD_HOPE.