Biologically inspired deep residual networks for computer vision applications

TL;DR

This work proposes Hex-ResNet, a biologically inspired deep residual network that integrates hexagonal convolutions along projection shortcuts while preserving square convolutions in the main path to enhance image classification. Implemented via the Hexagdly PyTorch library, the approach uses efficient two-rectangular-kernel representations to realize hex convolutions with minimal overhead. Empirical results on CIFAR-10 and ImageNet-2012 subsets show consistent Top-1 and Top-5 improvements over baseline ResNets and faster convergence, supporting better generalization. The findings suggest that blending square and hex tessellations in residual architectures can meaningfully improve discriminative power with limited computational cost, motivating broader exploration of hex-based components in CNNs.

Abstract

Deep neural network has been ensured as a key technology in the field of many challenging and vigorously researched computer vision tasks. Furthermore, classical ResNet is thought to be a state-of-the-art convolutional neural network (CNN) and was observed to capture features which can have good generalization ability. In this work, we propose a biologically inspired deep residual neural network where the hexagonal convolutions are introduced along the skip connections. The performance of different ResNet variants using square and hexagonal convolution are evaluated with the competitive training strategy mentioned by [1]. We show that the proposed approach advances the baseline image classification accuracy of vanilla ResNet architectures on CIFAR-10 and the same was observed over multiple subsets of the ImageNet 2012 dataset. We observed an average improvement by 1.35% and 0.48% on baseline top-1 accuracies for ImageNet 2012 and CIFAR-10, respectively. The proposed biologically inspired deep residual networks were observed to have improved generalized performance and this could be a potential research direction to improve the discriminative ability of state-of-the-art image classification networks.

Figures (12)

• Figure 1: Natural hexagonal structures heximages_online. (a) Football, and (b) Honeycomb.
• Figure 2: Optical section arrangement of the foveal cone mosaic curcio1990human. (a) Fovea central and (b) Fovea slope.
• Figure 3: Fundamental block of ResNet architecture with skip connections. (a) Identity shortcut heresnet and (b) Projection shortcut chen2020deep.
• Figure 4: Size of hexagonal kernels. (a) Size 1 and (b) Size 2.
• Figure 5: Split of size one hex kernel into equivalent rectangular kernels.