LeCun团队推出的自监督学习新方法VISReg通过将表征正则项解耦为尺度与形状两个独立目标，成功攻克了JEPA世界模型的表征坍塌核心问题[1]。该方法无需采用任何启发式训练技巧，在15个数据集上的综合表现超越MoCoV3、DINO、iBOT、I-JEPA、MAE、data2vec等主流自监督学习方法[1]。
在具体性能指标上，VISReg展现了显著优势[1]。ViT-B/16模型在域内线性探测精度达到75.7%，超过MAE的75.1%；ViT-L/14进一步提升至77.0%，超过LeJEPA的75.6%[1]。在分布外泛化能力方面，ViT-B/16的VISReg在6个分布外数据集上平均精度为70.19%，ViT-L/14为70.63%[1]。更引人瞩目的是，VISReg在ImageNet-22K上预训练后，分布外平均精度达72.94%，与用10倍数据规模（LVD-142M）训练的DINOv2基本持平[1]。在ADE20K线性语义分割任务上，ViT-B/16的mIoU达到30.16，也高于DINO的29.40和MAE的23.60[1]。
Yann LeCun对该研究成果进行了转发评价[1]。从算法层面看，VISReg的正则部分计算复杂度为O(NBK)，对所有扩展因子都是线性的[1]。
A new self-supervised learning method called VISReg, developed by LeCun's team, addresses a fundamental challenge in JEPA world models by decoupling the representation regularization term into two independent objectives: scale and shape [1]. The approach achieves superior performance across multiple benchmarks without relying on any heuristic training techniques.
VISReg outperforms seven mainstream self-supervised learning methods—including MoCoV3, DINO, iBOT, I-JEPA, MAE, and data2vec—across 15 datasets [1]. On ImageNet, ViT-B/16 achieves an in-distribution linear probing accuracy of 75.7%, surpassing MAE's 75.1%, while ViT-L/14 reaches 77.0%, exceeding LeJEPA's 75.6% [1]. Notably, VISReg reaches out-of-distribution performance comparable to DINOv2 using approximately one-tenth of the training data, achieving an average OOD accuracy of 72.94% on ImageNet-22K—nearly matching DINOv2's 72.93% trained on the 142-million-image LVD dataset [1]. Across six out-of-distribution datasets, ViT-B/16 VISReg attains an average OOD accuracy of 70.19%, with ViT-L/14 reaching 70.63% [1].
The method also demonstrates strong performance in downstream tasks, achieving a mean intersection-over-union (mIoU) of 30.16 on ADE20K linear semantic segmentation with ViT-B/16, outperforming both DINO (29.40) and MAE (23.60) [1]. The regularization component maintains linear computational complexity of O(NBK) across all scaling factors [1]. LeCun endorsed the work with a post stating, "VICReg begat SIGReg which begat VISReg" [1].
