Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN
This work addresses the limitation of MIM being restricted to Transformers, making it applicable to CNNs as well, which is incremental but broadens the scope of self-supervised pre-training in vision.
The paper tackled the problem that masked image modeling (MIM) was thought to be incompatible with CNNs, showing that MIM teaches models to learn better middle-order interactions among patches. They proposed an Architecture-Agnostic MIM framework (A^2MIM) that works with both Transformers and CNNs, achieving improved representation learning and stronger transfer capabilities on various downstream tasks.
Masked image modeling, an emerging self-supervised pre-training method, has shown impressive success across numerous downstream vision tasks with Vision transformers. Its underlying idea is simple: a portion of the input image is masked out and then reconstructed via a pre-text task. However, the working principle behind MIM is not well explained, and previous studies insist that MIM primarily works for the Transformer family but is incompatible with CNNs. In this work, we observe that MIM essentially teaches the model to learn better middle-order interactions among patches for more generalized feature extraction. We then propose an Architecture-Agnostic Masked Image Modeling framework (A$^2$MIM), which is compatible with both Transformers and CNNs in a unified way. Extensive experiments on popular benchmarks show that A$^2$MIM learns better representations without explicit design and endows the backbone model with the stronger capability to transfer to various downstream tasks.