AI Paper Atlas

Research map / Learning theory and algorithms / Matrix and tensor methods

Learning Representation from Neural Fisher Kernel with Low-rank Approximation

Ruixiang ZHANG, Shuangfei Zhai, Etai Littwin, Joshua M. Susskind

ICLR 2022 · Learning theory and algorithms › Matrix and tensor methods · approach: Geometry of low-rank and SPD structure · 1 citations (OpenAlex)

Read the paper

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
Use a 128-dimensional NFK embedding instead of GAN activationsGAN representation for linear evaluation · CIFAR-10accuracy: 65.3 -> 89.8Improved
Compress GAN Fisher vectors into a 128-dimensional NFK embeddingGAN representation for linear evaluation · CIFAR-10accuracy: 89.8 -> 89.8No clear effect
Increase VAE NFK embedding from 128 to 256 dimensions (ablation)VAE-NFK · CIFAR-10accuracy: 63.2 -> 68.7Improved
Increase GAN NFK embedding from 128 to 256 dimensions (ablation)GAN-NFK · CIFAR-10accuracy: 89.8 -> 89.8No clear effect
Use a pretrained GAN's NFK-128d representation with labeled-only linear probingNFK-128d · CIFAR-10 semi-supervised classificationerror rate: 20.68 vs 9.65 (500 labels); 14.77 vs 7.75 (1000 labels); 13.82 vs 7.03 (2000 labels); 12.95 vs 6.24 (4000 labels)Mixed
Use a pretrained GAN's NFK-128d representation with labeled-only linear probingNFK-128d · SVHN semi-supervised classificationerror rate: 8.74 vs 3.64 (500 labels); 4.47 vs 3.27 (1000 labels); 3.82 vs 3.04 (2000 labels); 3.19 vs 2.89 (4000 labels)Mixed
Distill with the teacher's NFK representation rather than layerwise targetsWRN-40-2 teacher → WRN-16-1 student · CIFAR-10 knowledge distillationaccuracy: 91.85 -> 92.42Improved

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