The Current State of the Art in Deep Learning for Image Classification: A Review

We present a review of the methods behind the top 40 highest accuracies achieved on the ILSVRC 2012 Imagenet validation set as ranked on Papers with Code. A significant proportion of these methods involve using transformer-based architectures, but it should be noted that none of the methods are naive self-attention transformers, which would be unmanageably large if the tokens were derived on a per-pixel basis. Rather, the works we review here all toil with different methods of combining the global nature of self-attention with the local nature of fine-grained image features, which have historically been the strength of convolutional neural networks. However, it should be noted that 9 out of 22 works reviewed did NOT use transformers.

Automated summary

This article reviews the methods used in the top 40 highest accuracy models on the ILSVRC 2012 Imagenet validation set. Many of these models utilize transformer-based architectures, although none of them are naive self-attention transformers. Instead, the reviewed works explore different ways to combine the global nature of self-attention with the local nature of fine-grained image features, traditionally the strength of convolutional neural networks.