FP-nets as novel deep networks inspired by vision

Area of Science:

• Computer Vision
• Computational Neuroscience
• Deep Learning

Background:

• State-of-the-art deep networks like ResNet and MobileNet are foundational in image recognition.
• End-stopped cortical cells in neuroscience exhibit feature selectivity, inspiring novel computational models.
• Understanding neuronal properties can lead to more robust and efficient artificial intelligence.

Purpose of the Study:

• To enhance existing deep neural networks by incorporating feature-product units (FP-units).
• To evaluate the performance of these enhanced networks on standard image recognition benchmarks.
• To analyze the emergent properties of neurons in the enhanced networks, specifically hyperselectivity and its impact on robustness.

Main Methods:

• Integration of feature-product units (FP-units) into ResNet and MobileNet architectures.
• Training and evaluation of the enhanced Feature-Product Networks (FP-nets) on Cifar-10 and ImageNet datasets.
• Analysis of neuronal hyperselectivity, end-stopped properties, and representational sparsity within the trained FP-nets.

Main Results:

• FP-nets demonstrated superior performance compared to baseline ResNet and MobileNet on Cifar-10 and ImageNet.
• Neuronal hyperselectivity in FP-nets correlated with increased resistance to adversarial attacks and JPEG compression artifacts.
• Learned neurons exhibited varying degrees of end-stopped properties, leading to sparser representations with decreasing entropy as hyperselectivity increased.

Conclusions:

• Feature-product networks offer a promising approach to improve deep learning model performance and robustness.
• The biological inspiration of end-stopped cells provides a valuable framework for designing more resilient AI systems.
• Hyperselectivity emerges as a key property for enhancing network defense against common image degradations and attacks.