arXiv cs.CV· Shuaibu Shehu (SR University), Elijah M. Flomo (SR University), Mohammad Safiya (GITAM University), Anurodh Kumar (SR University), Subodh Kumar (Dr. KNS Memorial Institute of Medical Sciences)·· 4 小时前AI 评分33
跨架构知识蒸馏用于轻量医疗视觉 Transformer（LiteMed-ViT）的潜力与局限
Promise and Limits of Cross-Architecture Knowledge Distillation for Lightweight Medical Vision Transformers (LiteMed-ViT)
AI 导读
LiteMed-ViT 用双 token 蒸馏把 ResNet18 教师迁移到轻量 ViT，仅 0.234M 参数，比教师少 47.8 倍。它在 BloodMNIST 与 PathMNIST 上分别达 95.05% 和 89.15%，MedMNIST-C 损坏下不弱于紧凑 CNN 基线。但 DermaMNIST 平衡准确率仍落后最强紧凑 CNN 至多 1.9 个点，更强教师也未带来更准的学生。
来源：arXiv cs.CV · arxiv.org