MLP-Mixer is a novel vision architecture that replaces convolutional and self-attention mechanisms with multi-layer perceptrons (MLPs). It consists of two types of MLP layers: channel-mixing MLPs and token-mixing MLPs. Channel-mixing MLPs process features across different channels, while token-mixing MLPs process spatial information across patches. MLP-Mixer is trained on large datasets and achieves competitive performance on image classification benchmarks, with pre-training and inference costs comparable to state-of-the-art models. It is simpler than CNNs and Transformers, and can be seen as a special case of CNNs with 1×1 convolutions for channel mixing and full receptive field convolutions for token mixing. However, typical CNNs are not special cases of MLP-Mixer. Despite its simplicity, MLP-Mixer achieves competitive results, reaching near state-of-the-art performance on ImageNet with pre-training on 100M images. When pre-trained on smaller datasets and combined with modern regularization techniques, it also achieves strong performance. It is slightly less effective than specialized CNN architectures but performs well in terms of accuracy and computational cost. MLP-Mixer is compared to other models, including CNNs, Transformers, and attention-based models, and is found to be competitive in terms of accuracy and computational efficiency. It is also invariant to input permutations, which is a key property of CNNs. The architecture is evaluated on various downstream tasks and is found to be effective in terms of accuracy and computational cost. The results show that MLP-Mixer is a promising alternative to CNNs and Transformers for vision tasks.MLP-Mixer is a novel vision architecture that replaces convolutional and self-attention mechanisms with multi-layer perceptrons (MLPs). It consists of two types of MLP layers: channel-mixing MLPs and token-mixing MLPs. Channel-mixing MLPs process features across different channels, while token-mixing MLPs process spatial information across patches. MLP-Mixer is trained on large datasets and achieves competitive performance on image classification benchmarks, with pre-training and inference costs comparable to state-of-the-art models. It is simpler than CNNs and Transformers, and can be seen as a special case of CNNs with 1×1 convolutions for channel mixing and full receptive field convolutions for token mixing. However, typical CNNs are not special cases of MLP-Mixer. Despite its simplicity, MLP-Mixer achieves competitive results, reaching near state-of-the-art performance on ImageNet with pre-training on 100M images. When pre-trained on smaller datasets and combined with modern regularization techniques, it also achieves strong performance. It is slightly less effective than specialized CNN architectures but performs well in terms of accuracy and computational cost. MLP-Mixer is compared to other models, including CNNs, Transformers, and attention-based models, and is found to be competitive in terms of accuracy and computational efficiency. It is also invariant to input permutations, which is a key property of CNNs. The architecture is evaluated on various downstream tasks and is found to be effective in terms of accuracy and computational cost. The results show that MLP-Mixer is a promising alternative to CNNs and Transformers for vision tasks.