Dendritic Learning-Incorporated Vision Transformer for Image Recognition

Dendritic Learning-Incorporated Vision Transformer for Image Recognition

February 2024 | Zhiming Zhang, Zhenyu Lei, Masaaki Omura, Hideyuki Hasegawa, Shangce Gao
This letter proposes a novel neural network architecture, Dendritic Learning-Incorporated Vision Transformer (DVT), which integrates dendritic learnable network architecture with Vision Transformer to enhance image recognition accuracy. Inspired by dendritic neurons in neuroscience, the DVT combines a Vision Transformer for feature extraction and a dendritic network for feature classification. The Vision Transformer processes images into multiple patches using self-attention mechanisms, while the dendritic network, with its synapse, dendrite, and soma layers, ensures accurate feature classification. Extensive experiments on various benchmarks, including CIFAR10, SVHN, CIFAR100, and Tiny-ImageNet, demonstrate that DVT outperforms state-of-the-art methods, particularly in handling complex nonlinear classification problems. The study also highlights the importance of biological interpretability in improving both performance and understanding of visual perception mechanisms.This letter proposes a novel neural network architecture, Dendritic Learning-Incorporated Vision Transformer (DVT), which integrates dendritic learnable network architecture with Vision Transformer to enhance image recognition accuracy. Inspired by dendritic neurons in neuroscience, the DVT combines a Vision Transformer for feature extraction and a dendritic network for feature classification. The Vision Transformer processes images into multiple patches using self-attention mechanisms, while the dendritic network, with its synapse, dendrite, and soma layers, ensures accurate feature classification. Extensive experiments on various benchmarks, including CIFAR10, SVHN, CIFAR100, and Tiny-ImageNet, demonstrate that DVT outperforms state-of-the-art methods, particularly in handling complex nonlinear classification problems. The study also highlights the importance of biological interpretability in improving both performance and understanding of visual perception mechanisms.
Reach us at info@study.space
Understanding Dendritic Learning-Incorporated Vision Transformer for Image Recognition