{"id":"arxiv:2606.02572","title":"VISReg: Variance-Invariance-Sketching Regularization for JEPA training","date":"2026-06-01","section":"Self-supervised learning and vision","quality":0.33919,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/","source_url":"https://arxiv.org/abs/2606.02572","text_available":true,"authors":["Haiyu Wu","Randall Balestriero","Morgan Levine"],"outgoing":[{"source":"arxiv:2606.02572","target":"arxiv:2006.09882","evidence":"title","id":"arxiv:2006.09882","title":"Unsupervised Learning of Visual Features by Contrasting Cluster Assignments","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6601018611a96dff6f5324f2e4f39f62484917cf4d8f8a8d5baa7f4be1deb433/"},{"source":"arxiv:2606.02572","target":"arxiv:2104.14294","evidence":"title","id":"arxiv:2104.14294","title":"Emerging Properties in Self-Supervised Vision Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4372186d141ca514b08e93bb3b1e26fdbcba9ce62187fb772c14bf223160c06d/"},{"source":"arxiv:2606.02572","target":"arxiv:2105.04906","evidence":"title","id":"arxiv:2105.04906","title":"VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/822b6414fb02bd6c746e0d559fd3dcc60766476142ebae7795b02bc6c45223ce/"},{"source":"arxiv:2606.02572","target":"arxiv:2111.07832","evidence":"title","id":"arxiv:2111.07832","title":"iBOT: Image BERT Pre-Training with Online Tokenizer","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2bd0269c0872fc518db65c2ca98dd30b5d38701f79b0a72fec2ed49648ba589f/"},{"source":"arxiv:2606.02572","target":"arxiv:2301.08243","evidence":"title","id":"arxiv:2301.08243","title":"Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/475bac0a3b29edd42642bbe10d1a41bfb18a3022ac95b9b1a3db9e5d5d0b95fe/"},{"source":"arxiv:2606.02572","target":"arxiv:2304.07193","evidence":"title","id":"arxiv:2304.07193","title":"DINOv2: Learning Robust Visual Features without Supervision","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b15933596df35497024556ae88bfa696f3b0e8372b678194f04f1bfb9f3be553/"},{"source":"arxiv:2606.02572","target":"arxiv:2511.08544","evidence":"id","id":"arxiv:2511.08544","title":"LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e2c5835a26dabc321e6364380a37c47c10425ee830ec522564b0604733025a94/"}],"incoming":[],"text_parts":[{"start":0,"end":15807,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html"},{"start":15807,"end":31011,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html"},{"start":31011,"end":46784,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html"},{"start":46784,"end":49376,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-4.html"}],"contents":[{"title":"VISReg: Variance-Invariance-Sketching Regularization for JEPA training","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"Abstract","offset":74,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"Keywords:","offset":1540,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"1 Introduction","offset":1558,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"2 Related Work","offset":6803,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"3 VISReg: Variance-Invariance-Sketching Regularization","offset":11281,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"3.1 Regularization Loss","offset":12732,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"Lemma 3.1 (Cramér-Wold Theorem) .","offset":14718,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"Lemma 3.2 (1D Wasserstein Closed-Form) .","offset":15758,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-1.html#text"},{"title":"Proposition 3.3 (VISReg Regularization Objective) .","offset":17095,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html#text"},{"title":"3.2 VISReg Is Friendly to Scale Up","offset":18401,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html#text"},{"title":"Definition 3.4 .","offset":18707,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html#text"},{"title":"3.3 VISReg Is Robust to Low-Quality Datasets","offset":22475,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html#text"},{"title":"4 Experiment","offset":24560,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html#text"},{"title":"4.1 Ablation study","offset":24864,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html#text"},{"title":"4.2 General comparison","offset":28136,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-2.html#text"},{"title":"5 Conclusion","offset":33370,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"6 Acknowledgment","offset":34068,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"Appendix A Implementation details","offset":34201,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"A.1 Training details","offset":34239,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"A.2 Testing details","offset":36918,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"A.3 Datasets","offset":39652,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"Appendix B Additional ablations","offset":43393,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"B.1 Effect of decoupled components in training.","offset":44173,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"B.2 Effect of decoupled components in different training set scenarios.","offset":45154,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"Appendix C Visualizations","offset":46218,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"C.1 VISReg loss indicates the performance","offset":46248,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"C.2 Further comparison with DINOv1 on image and video.","offset":46691,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"PCA Feature Visualization.","offset":46751,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-3.html#text"},{"title":"Instructions for reporting errors","offset":48412,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/part-4.html#text"}],"text_code_points":49376,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/262c1938984761f99c55d9ae1237e686b2d6ab94b69d7181d11dceab2a7210fb/source.md"}