AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Augmenting Deep Classifiers with Polynomial Neural Networks

Grigorios G. Chrysos, Markos Georgopoulos, Jiankang Deng, Jean Kossaifi, Yannis Panagakis, Anima Anandkumar

ECCV 2022 · Architectures and efficiency › CNN design · approach: Filter design and trained-filter analysis · 16 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 ResNet18 blocks with PDC keeping channelsResNet18 · CIFAR100accuracy: 0.756 -> 0.773Improved
shrink PDC channels to compress the modelPDC · CIFAR100accuracy: 0.756 -> 0.760Improved
increase polynomial degree of PDC blocks from two to four (ablation)PDC · CIFAR100accuracy: 0.765 -> 0.771Improved
add first-degree term and unshared factors to disentangled non-local block (ablation)DNL · CIFAR100 (non-local blocks)accuracy: 0.771 -> 0.773Improved
add fourth-degree squeeze-excitation-like term to PDC-NL[3] (ablation)PDC-NL[3] · CIFAR100 (non-local blocks)accuracy: 0.773 -> 0.779Improved
replace non-local block with PDC-NL[4]ResNet18 · ImageNet (non-local blocks)top-1 accuracy: 0.702 -> 0.716Improved
use PDC instead of ResNet18 or SENet on CIFAR10ResNet18 · CIFAR10accuracy: 0.945 -> 0.946No clear effect
use PDC instead of ResNet18 on speech classificationResNet18 · Speech Commandsaccuracy: 0.977 -> 0.978No clear effect
use PDC-param instead of ResNet18 under strong class imbalanceResNet18 · CIFAR10-LT (IF 200)accuracy: 0.645 -> 0.665Improved
use PDC-comp instead of ResNet18 with very limited dataResNet18 · CIFAR10 (50 samples per class)accuracy: 0.347 -> 0.426Improved
use PDC polynomial blocks instead of plain ResNet18 when activations are removed (ablation)ResNet18 · CIFAR100 (no activation functions)accuracy: 0.168 -> 0.689Improved
raise degree of Pi-net blocks from two to four without activations (ablation)Pi-net-ResNet · CIFAR100 (no activation functions)accuracy: 0.667 -> 0.626Hurt

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