AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

Quantifying the Knowledge in a DNN to Explain Knowledge Distillation for Classification

Quanshi Zhang, Cheng Xu, Yilan Chen, Zhefan Rao

TPAMI 2022 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives · 40 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train by feature distillation instead of from scratchVGG-16 FC1 · CUB200-2011N_point^fg: 22.50 -> 43.77Improved
train by feature distillation instead of from scratchVGG-16 FC1 · CUB200-2011foreground ratio lambda: 0.68 -> 0.84Improved
train by feature distillation instead of from scratchVGG-16 FC1 · CUB200-2011rho: 0.50 -> 0.66Improved
train by feature distillation instead of from scratchVGG-16 FC1 · CUB200-2011D_var: 4.98 -> 0.06Improved
train by feature distillation instead of from scratchBERT · SQuADN_point^fg: 3.58 -> 8.46Improved
train by feature distillation instead of from scratchBERT · SST-2rho: 0.51 -> 0.80Improved
train by feature distillation instead of from scratchPointNet++ FC1 · ModelNet40 with background pointsN_point^fg: 16.04 -> 51.15Improved
train by feature distillation instead of from scratchDGCNN FC1 · ModelNet40 with background pointsN_point^fg: 102.53 -> 84.44Hurt
train by feature distillation instead of from scratchVGG-16 FC1 · ILSVRC-2013 DETN_point^fg: 36.47 -> 35.98No clear effect
distill from the output layer FC3 instead of scratchVGG-16 FC3 · CUB200-2011lambda: 0.71 -> 0.63Hurt
use 8x8 grid instead of 16x16 for knowledge-point measurement (ablation)VGG-16 FC1 student · CUB200-2011N_point^fg: student still above baseline 10.06 vs 6.47No clear effect
lower threshold b to 0.15 for knowledge-point measurement (ablation)VGG-16 FC1 student · CUB200-2011N_point^fg: student still above baseline 54.53 vs 26.43No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy