An all-MLP Architecture for Vision
Convolutional Neural Networks have dominated computer vision for nearly 10 years, and that might finally come to an end. First, Vision Transformers (ViT) have shown remarkable performance, and now even simple MLP-based models reach competitive accuracy, as long as sufficient data is used for pre-training. This paper presents MLP-Mixer, using MLPs in a particular weight-sharing arrangement to achieve a competitive, high-throughput model and it raises some interesting questions about the nature of learning and inductive biases and their interaction with scale for future research.
Convolutional Neural Networks (CNNs) are the go-to model for computer vision. Recently, attention-based networks, such as the Vision Transformer, have also become popular. In this paper we show that while convolutions and attention are both sufficient for good performance, neither of them are necessary. We present MLP-Mixer, an architecture based exclusively on multi-layer perceptrons (MLPs). MLP-Mixer contains two types of layers: one with MLPs applied independently to image patches (i.e. “mixing” the per-location features), and one with MLPs applied across patches (i.e. “mixing” spatial information). When trained on large datasets, or with modern regularization schemes, MLP-Mixer attains competitive scores on image classification benchmarks, with pre-training and inference cost comparable to state-of-the-art models. We hope that these results spark further research beyond the realms of well established CNNs and Transformers.