Humans see low and high spatial frequency components at the same time, and combine the information from both to form a visual scene. Drawing on this neuroscientific inspiration, we propose an altered Vision Transformer architecture where patches from scaled down versions of the input image are added to the input of the first Transformer Encoder layer. We name this model Retina Vision Transformer (RetinaViT) due to its inspiration from the human visual system. Our experiments show that when trained on the ImageNet-1K dataset with a moderate configuration, RetinaViT achieves a 3.3% performance improvement over the original ViT. We hypothesize that this improvement can be attributed to the inclusion of low spatial frequency components in the input, which improves the ability to capture structural features, and to select and forward important features to deeper layers. RetinaViT thereby opens doors to further investigations into vertical pathways and attention patterns.
View on arXiv