📄 中文摘要
在资源受限的分布式多媒体学习场景中，协作知识蒸馏的最新进展展现了前沿性能。然而，实现这种竞争力需要解决一个根本性的不匹配问题：高维教师知识的复杂性与异构客户端学习能力之间的差异，这在当前阻碍了边缘视觉分析系统的部署。受到课程学习原则的启发，提出了联邦自适应渐进蒸馏（FAPD），这是一个共识驱动的框架，旨在协调自适应知识转移。FAPD通过基于主成分分析（PCA）的结构对教师特征进行层次分解，提取按方差排序的主成分，从而实现高效的知识传递。
渐进知识蒸馏的协作自适应课程
出处: Collaborative Adaptive Curriculum for Progressive Knowledge Distillation
发布: 2026年3月24日
📄 English Summary
Collaborative Adaptive Curriculum for Progressive Knowledge Distillation
Recent advancements in collaborative knowledge distillation have showcased state-of-the-art performance in resource-constrained distributed multimedia learning environments. However, achieving such competitiveness necessitates addressing a fundamental mismatch between the complexity of high-dimensional teacher knowledge and the heterogeneous learning capacities of clients, which currently hinders deployment in edge-based visual analytics systems. Inspired by principles of curriculum learning, the Federated Adaptive Progressive Distillation (FAPD) framework is proposed as a consensus-driven approach to orchestrate adaptive knowledge transfer. FAPD hierarchically decomposes teacher features using PCA-based structuring, extracting principal components ordered by variance to facilitate efficient knowledge transfer.