Papers › Masking meets Supervision: A Strong Learning Alliance

Masking meets Supervision: A Strong Learning Alliance

20 Jun 2023CVPR 2025 1arXiv:2306.11339archive 2025-07-28

Byeongho Heo, Taekyung Kim, Sangdoo Yun, Dongyoon Han

Pre-training with random masked inputs has emerged as a novel trend in self-supervised training. However, supervised learning still faces a challenge in adopting masking augmentations, primarily due to unstable training. In this paper, we propose a novel way to involve masking augmentations dubbed Masked Sub-branch (MaskSub). MaskSub consists of the main-branch and sub-branch, the latter being a part of the former. The main-branch undergoes conventional training recipes, while the sub-branch merits intensive masking augmentations, during training. MaskSub tackles the challenge by mitigating adverse effects through a relaxed loss function similar to a self-distillation loss. Our analysis shows that MaskSub improves performance, with the training loss converging faster than in standard training, which suggests our method stabilizes the training process. We further validate MaskSub across diverse training scenarios and models, including DeiT-III training, MAE finetuning, CLIP finetuning, BERT training, and hierarchical architectures (ResNet and Swin Transformer). Our results show that MaskSub consistently achieves impressive performance gains across all the cases. MaskSub provides a practical and effective solution for introducing additional regularization under various training recipes. Code available at https://github.com/naver-ai/augsub

PaperPDFConference PDFCode

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

naver-ai/augsub officialmentioned in papermentioned on GitHubpytorchNOASSERTION report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Image ClassificationSelf-Supervised Image ClassificationTransfer Learning

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Image Classification ImageNet ViT-H @224 (DeiT-III + AugSub) Number of params 632M #212 of 1060 Archive leaderboard report
Image Classification ImageNet ViT-H @224 (DeiT-III + AugSub) Top 1 Accuracy 85.7% #212 of 1060 Archive leaderboard report
Image Classification ImageNet ViT-L @224 (DeiT-III + AugSub) Number of params 304M #242 of 1060 Archive leaderboard report
Image Classification ImageNet ViT-L @224 (DeiT-III + AugSub) Top 1 Accuracy 85.3% #242 of 1060 Archive leaderboard report
Image Classification ImageNet ViT-B @224 (DeiT-III + AugSub) Number of params 86.6M #342 of 1060 Archive leaderboard report
Image Classification ImageNet ViT-B @224 (DeiT-III + AugSub) Top 1 Accuracy 84.2% #342 of 1060 Archive leaderboard report
Self-Supervised Image Classification ImageNet (finetuned) MAE + AugSub finetune (ViT-H/14) Number of Params 632M #9 of 65 Archive leaderboard report
Self-Supervised Image Classification ImageNet (finetuned) MAE + AugSub finetune (ViT-H/14) Top 1 Accuracy 87.2% #9 of 65 Archive leaderboard report
Self-Supervised Image Classification ImageNet (finetuned) MAE + AugSub finetune (ViT-L/16) Number of Params 304M #17 of 65 Archive leaderboard report
Self-Supervised Image Classification ImageNet (finetuned) MAE + AugSub finetune (ViT-L/16) Top 1 Accuracy 86.1% #17 of 65 Archive leaderboard report
Self-Supervised Image Classification ImageNet (finetuned) MAE + AugSub finetune (ViT-B/16) Number of Params 87M #41 of 65 Archive leaderboard report
Self-Supervised Image Classification ImageNet (finetuned) MAE + AugSub finetune (ViT-B/16) Top 1 Accuracy 83.9% #41 of 65 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

AttentionCLIPDense ConnectionsDropoutLayer NormalizationMAEMIMMulti-Head AttentionRandom ErasingResidual ConnectionStochastic DepthSwin TransformerTransformerVision Transformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections