TY - UNPB
T1 - PushPull-Net
T2 - Inhibition-driven ResNet robust to image corruptions
AU - Bennabhaktula, Guru Swaroop
AU - Alegre, Enrique
AU - Strisciuglio, Nicola
AU - Azzopardi, George
PY - 2024/8/7
Y1 - 2024/8/7
N2 - We introduce a novel computational unit, termedPushPull-Conv, in the first layer of a ResNet architecture, inspired bythe anti-phase inhibition phenomenon observed in the primary visualcortex. This unit redefines the traditional convolutional layer byimplementing a pair of complementary filters: a trainable push kerneland its counterpart, the pull kernel. The push kernel (analogous totraditional convolution) learns to respond to specific stimuli, while thepull kernel reacts to the same stimuli but of opposite contrast. Thisconfiguration enhances stimulus selectivity and effectively inhibitsresponse in regions lacking preferred stimuli. This effect is attributed tothe push and pull kernels, which produce responses of comparablemagnitude in such regions, thereby neutralizing each other. Theincorporation of the PushPull-Conv into ResNets significantly increasestheir robustness to image corruption. Our experiments with benchmarkcorruption datasets show that the PushPull-Conv can be combinedwith other data augmentation techniques to further improve modelrobustness. We set a new robustness benchmark on ResNet50 achievingan mCE of 49.95% on ImageNet-C when combining PRIMEaugmentation with PushPull inhibition.
AB - We introduce a novel computational unit, termedPushPull-Conv, in the first layer of a ResNet architecture, inspired bythe anti-phase inhibition phenomenon observed in the primary visualcortex. This unit redefines the traditional convolutional layer byimplementing a pair of complementary filters: a trainable push kerneland its counterpart, the pull kernel. The push kernel (analogous totraditional convolution) learns to respond to specific stimuli, while thepull kernel reacts to the same stimuli but of opposite contrast. Thisconfiguration enhances stimulus selectivity and effectively inhibitsresponse in regions lacking preferred stimuli. This effect is attributed tothe push and pull kernels, which produce responses of comparablemagnitude in such regions, thereby neutralizing each other. Theincorporation of the PushPull-Conv into ResNets significantly increasestheir robustness to image corruption. Our experiments with benchmarkcorruption datasets show that the PushPull-Conv can be combinedwith other data augmentation techniques to further improve modelrobustness. We set a new robustness benchmark on ResNet50 achievingan mCE of 49.95% on ImageNet-C when combining PRIMEaugmentation with PushPull inhibition.
U2 - 10.48550/arXiv.2408.04077
DO - 10.48550/arXiv.2408.04077
M3 - Preprint
BT - PushPull-Net
PB - arXiv
CY - Kolkata
ER -