Masked Image Residual Learning for Scaling Deeper Vision Transformers

NeurIPS 2023  ·  Guoxi Huang, Hongtao Fu, Adrian G. Bors ·

Deeper Vision Transformers (ViTs) are more challenging to train. We expose a degradation problem in deeper layers of ViT when using masked image modeling (MIM) for pre-training. To ease the training of deeper ViTs, we introduce a self-supervised learning framework called Masked Image Residual Learning (MIRL), which significantly alleviates the degradation problem, making scaling ViT along depth a promising direction for performance upgrade. We reformulate the pre-training objective for deeper layers of ViT as learning to recover the residual of the masked image. We provide extensive empirical evidence showing that deeper ViTs can be effectively optimized using MIRL and easily gain accuracy from increased depth. With the same level of computational complexity as ViT-Base and ViT-Large, we instantiate 4.5$\times$ and 2$\times$ deeper ViTs, dubbed ViT-S-54 and ViT-B-48. The deeper ViT-S-54, costing 3$\times$ less than ViT-Large, achieves performance on par with ViT-Large. ViT-B-48 achieves 86.2% top-1 accuracy on ImageNet. On one hand, deeper ViTs pre-trained with MIRL exhibit excellent generalization capabilities on downstream tasks, such as object detection and semantic segmentation. On the other hand, MIRL demonstrates high pre-training efficiency. With less pre-training time, MIRL yields competitive performance compared to other approaches.

PDF Abstract NeurIPS 2023 PDF NeurIPS 2023 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Image Classification ImageNet MIRL(ViT-S-54) Top 1 Accuracy 84.8% # 270
Number of params 96M # 856
GFLOPs 18.8 # 361
Image Classification ImageNet MIRL (ViT-B-48) Top 1 Accuracy 86.2% # 164
Number of params 341M # 923
GFLOPs 67.0 # 438
Self-Supervised Image Classification ImageNet (finetuned) MIRL (ViT-B-48) Number of Params 341M # 12
Top 1 Accuracy 86.2% # 16
Self-Supervised Image Classification ImageNet (finetuned) MIRL (ViT-S-54) Number of Params 96M # 31
Top 1 Accuracy 84.8% # 26

Methods


No methods listed for this paper. Add relevant methods here