AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

Feature segregation by signed weights in artificial vision systems and biological models

Giordano Ramos-Traslosheros, Carlos R Ponce

ICLR 2026 · Vision-language and multimodal › Human and embodied modeling · approach: fMRI visual decoding

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
remove positive weights to class units (ablate positive inputs) (ablation)ImageNet CNN output units · Background challengebackground reliance R_bg: 0.157 -> 0.488Hurt
remove negative weights to class units (ablation)ImageNet CNN output units · Background challengebackground reliance R_bg: 0.157 -> 0.078Improved
ablate positive weights (ablation)ImageNet CNN output units · Background challengeforeground reliance R_fg: 0.903 -> 0.839No clear effect
ablate negative weights (ablation)ImageNet CNN output units · Background challengeforeground reliance R_fg: 0.903 -> 0.711Hurt
use ReLU instead of Tanh activations (ablation)ResNet18 trained 16 epochs · ImageNet-1k valtop-5 accuracy: 0.797 -> 0.870Improved
use ReLU instead of Tanh activations (ablation)ResNet18 · feature segregation by signrepresentational change under ablation: ReLU model showed clear segregation; Tanh similar changes across ablation typesImproved
add adversarial robust training (larger L∞) (ablation)ResNet50 · ablation similarity changeSpearman rho of change vs robustness: -0.47 (6e-5) positive, -0.57 (5e-7) negative at alpha 1.0Improved
train with supervision and more depth (training vs random init)ImageNet CNNs · output layerDale index: 0.6004 vs 0.5125 for resnet50Improved
add batchnorm to VGGVGG16 · ImageNet-1k valDale index: 0.5606 -> 0.5761Improved
use unsupervised SimSiam pretraining with linear headResNet50 · feature segregationablation effect: object features still in positive weights; disappeared at lower ablation strengthsNo clear effect
constrain neuron models to positive input weights only (ablation)PLS neuron models of macaque V1/V4/IT · neural response predictiontrain and test accuracyHurt
clear the background around neuron's preferred featuremacaque ventral stream neurons · in vivofiring responseImproved

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