AI Paper Atlas

Research map / Learning theory and algorithms / Kernels and regression theory

Autoencoders that don't overfit towards the Identity

Harald Steck

NeurIPS 2020 · Learning theory and algorithms › Kernels and regression theory · approach: Double descent and overfitting · 23 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 weight decay on U and V with L2 norm on the product UV^T (ablation)linear autoencoder · MSDnDCG@100: 0.249 -> 0.266Improved
replace weight decay with dropout-style L2 on UV^T (ablation)linear autoencoder · ML-20MnDCG@100: 0.376 -> 0.407Improved
use item-specific Lambda instead of constant lambda in L2 regularization (ablation)linear autoencoder · MSDnDCG@100: 0.266 -> 0.314Improved
use item-specific Lambda instead of constant lambda (ablation)linear autoencoder · ML-20MnDCG@100: 0.407 -> 0.417Improved
add full emphasis (emphasized denoising) to DLAE (ablation)DLAE · ML-20MnDCG@100: 0.413 -> 0.420Improved
add emphasized denoising to dropout training (ablation)DLAE · NetflixnDCG@100: 0.383 -> 0.394Improved
add emphasized denoising to dropout training (ablation)DLAE · MSDnDCG@100: 0.309 -> 0.320Improved
train with sampled dropout vs analytic dropout-induced L2 (ablation)LAE · NetflixnDCG@100: 0.384 vs 0.383No clear effect
enforce (near-)orthogonality constraint diag(UV^T)=0 instead of unconstrained diagonal (ablation)DLAE low-rank · NetflixnDCG@100Improved

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