AI Paper Atlas

Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

Improving Compositionality of Neural Networks by Decoding Representations to Inputs

Mike Wu, Noah Goodman, Stefano Ermon

NeurIPS 2021 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Priors and functional variational uncertainty · 0 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
Add a jointly trained decoder that reconstructs inputs from every classifier activationStandard neural network · MNIST, FashionMNIST, CelebAaccuracy: 97.1 -> 97.8; 87.7 -> 88.8; 90.8 -> 90.8Improved
Recursively compose the decodable classifier with itselfDecNN · MNIST, FashionMNIST, CelebAaccuracy: 97.1 -> 97.5; 87.7 -> 88.1; 90.8 -> 90.8No clear effect
Use recursive-path disagreement as an uncertainty scoreReDecNN · CelebA correctly vs. incorrectly classifiedROC-AUC: 0.657 -> 0.692Improved
Use recursive-path disagreement as an uncertainty scoreReDecNN · MNIST and FashionMNIST correctly vs. incorrectly classifiedROC-AUC: 0.878 -> 0.869; 0.818 -> 0.795Hurt
Use recursive-path disagreement for OOD detectionReDecNN · MNIST trained; FashionMNIST OODROC-AUC: 0.888 -> 0.812Hurt
Use recursive-path disagreement for OOD detectionReDecNN · CelebA trained; held-out wearing-a-hat attributeROC-AUC: 0.519 -> 0.604Improved
Use recursive compositionStandard neural network · MNIST, FashionMNIST, CelebA testECE: 1.18 -> 0.33; 0.56 -> 0.21; 2.75 -> 1.92Improved
Increase recursive depth from 2 through 8 (ablation)ReDecNN · MNIST trained; FashionMNIST OODROC-AUC: 0.665 -> 0.788 -> 0.805 -> 0.812Improved
Increase recursive depth from 2 through 8 (ablation)ReDecNN · MNIST testaccuracy: 97.7; 97.6; 97.7; 97.5No clear effect
Compose with a pretrained linear classifierReDecNN · MNIST trained; FashionMNIST OODROC-AUC: 0.812 -> 0.845Improved
Compose with a pretrained ResNet-18ReDecNN · MNIST trained; FashionMNIST OODROC-AUC: 0.812 -> 0.650Hurt
Regularize predictions on reconstructions rather than on the original input (ablation)ReDecNN with pretrained linear model · MNIST; FashionMNIST OODROC-AUC: 0.699 -> 0.845Improved

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