AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Tensor Regression Networks

Jean Kossaifi, Zachary C. Lipton, Arinbjorn Kolbeinsson, Aran Khanna, Tommaso Furlanello, Anima Anandkumar

JMLR 2020 · Architectures and efficiency › CNN design · approach: Theory of convolutional inductive bias

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 fully connected layer with a tensor regression layer (256,3,4,3,1)3D-ResNet · UK Biobank MRI ageMAE (years): 2.96 -> 2.70Improved
replace the fully connected layer with a TRL for gender classification3D-ResNet · UK Biobank MRI genderclassification error (%): 0.79 -> 0.53Improved
replace the fully connected layer with a TRL for BMI regression3D-ResNet · UK Biobank MRI BMIMAE (kg/m2): 2.37 -> 2.26Improved
use the full-rank 6x7x6 activation tensor in the TRL instead of low rank (ablation)TRL on UK Biobank age · UK Biobank MRI ageMAE (years): 2.71No clear effect
remove average pooling without a TRL (keep spatial dims with FC) (ablation)3D-ResNet · UK Biobank MRItraining stabilityHurt
reduce TRL rank from (200,1,1,200) to (100,1,1,100) (ablation)TRL on ResNet-101 · ImageNet-1k valTop-1 accuracy (%): 77.1 -> 74.6Hurt
reduce TRL rank to (50,1,1,50) (ablation)TRL on ResNet-101 · ImageNet-1k valTop-1 accuracy (%): 77.1 -> 73.6Hurt
replace the FC layer with a TRL of rank (300,1,1,700)ResNet-101 · ImageNet-1k valTop-1 accuracy (%): 77.1 -> 77.2Improved
add a TCL (512,7,7) before the fully connected layersVGG-19 · ImageNet-1k valTop-1 accuracy (%): 68.7 -> 69.4Improved
shrink the TCL to (384,5,5) with 3072 hidden unitsVGG-19 · ImageNet-1k valTop-1 accuracy (%): 68.7 -> 68.3No clear effect
replace pooling and FC with an overcomplete TRL (1000,2048,7,7)ResNet-50 · ImageNet-1k valTop-1 accuracy (%): 74.58 -> 73.6Hurt
add a TCL (1024,3,3) before the TRL (ablation)ResNet-50 with TRL (1000,1024,3,3) · ImageNet-1k valTop-1 accuracy (%): 72.16 -> 73.43Mixed

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