AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Generalized Shape Metrics on Neural Representations

Alex H Williams, Erin Kunz, Simon Kornblith, Scott Linderman

NeurIPS 2021 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability · 21 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
Replace the mean-canonical-correlation dissimilarity with arccos of the mean correlationLinear heuristic · Allen Brain Observatory and NAS-Bench-101 Stem layertriangle inequality violations: 17/1128; 10128/100000 -> 0 by metric constructionImproved
Use the shape distance instead of the linear heuristicHierarchical clustering of Allen Brain Observatory representations · Allen Brain Observatorydendrogram similarity: ~85.1% similarityMixed
Increase the MDS embedding dimension from 2 to a sufficiently large dimensionMDS Euclidean embedding · Allen Brain Observatory and NAS-Bench-101pairwise-distance distortion: ~5% (ABO); 10% (NAS-Bench-101)Improved
Use the convolution-aware metric rather than a flattened metricConvolutional-layer representations · CIFAR-10 test setconvergence of distance estimate with sample size: m=10^4Improved
Use a looser equivalence relation with more flexible alignmentsGeneralized shape metrics · Deep convolutional network representationssample size needed for a stable distance estimateHurt

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