{"id":"arxiv:2607.05609","title":"To Retain or to Adapt? Generalizing Continual Learning","date":"2026-07-06","section":"Post-training","quality":0.077814,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/","source_url":"https://arxiv.org/abs/2607.05609","text_available":true,"authors":["Giulia Lanzillotta","Mandana Samiei","Doina Precup","Razvan Pascanu"],"outgoing":[{"source":"arxiv:2607.05609","target":"arxiv:1801.01290","evidence":"title","id":"arxiv:1801.01290","title":"Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/21e70876ae9603903b2b20dd667a413bdd8b0abf37b9c11e284ea8b5e4124c74/"},{"source":"arxiv:2607.05609","target":"arxiv:2108.06325","evidence":"id","id":"arxiv:2108.06325","title":"Continual Backprop: Stochastic Gradient Descent with Persistent Randomness","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/"},{"source":"arxiv:2607.05609","target":"doi:10.1038/s41586-024-07711-7","evidence":"id","id":"doi:10.1038/s41586-024-07711-7","title":"Maintaining Plasticity in Deep Continual Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/"}],"incoming":[],"text_parts":[{"start":0,"end":15987,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html"},{"start":15987,"end":31622,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html"},{"start":31622,"end":47250,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html"},{"start":47250,"end":62800,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html"},{"start":62800,"end":77920,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html"},{"start":77920,"end":93765,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html"},{"start":93765,"end":109369,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html"},{"start":109369,"end":125351,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html"},{"start":125351,"end":141123,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html"},{"start":141123,"end":154774,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-10.html"},{"start":154774,"end":170703,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html"},{"start":170703,"end":185676,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html"},{"start":185676,"end":201368,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html"},{"start":201368,"end":216996,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html"},{"start":216996,"end":219757,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-15.html"}],"contents":[{"title":"To Retain or to Adapt? Generalizing Continual Learning","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"Abstract","offset":58,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"1 Introduction","offset":1958,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"Contributions.","offset":7388,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"2 Background & Setup","offset":9128,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"2.1 A General Objective for Continual Learning","offset":9154,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"General setting.","offset":9600,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"Continual learning objective.","offset":11212,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"2.2 Analysis Setup and Formalism","offset":13265,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"Definition 1 (Task-based data generating process.) .","offset":13547,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"2.3 The algorithm as a Markov Process","offset":15943,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-1.html#text"},{"title":"2.4 JTL and ITL processes","offset":18375,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"The Joint-Task Learning algorithm ( 𝒜 JTL \\mathcal{A}_{\\text{JTL}} ).","offset":18618,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"The Independent-Task Learning algorithm ( 𝒜 ITL \\mathcal{A}_{\\text{ITL}} ).","offset":19378,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"Equivalent computational costs.","offset":20074,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"3 Transfer Efficiency","offset":21115,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"Definition 2 (Transfer Efficiency) .","offset":21303,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"Definition 3 (Instability) .","offset":23613,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"Definition 4 (Transient Error) .","offset":23927,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"Lemma 5 (TE Decomposition) .","offset":24573,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"3.1 The Critical Task Duration","offset":25721,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"Lemma 6 (The Critical Task Duration) .","offset":26026,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"3.2 Empirical Validation of the Framework","offset":27834,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"(P1) Controlled Instability: Permuted- and Shuffled-CIFAR.","offset":30935,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-2.html#text"},{"title":"(P2) Crossing the Critical Task Duration in CLEAR.","offset":32405,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"4 Quadratic-Loss Analysis","offset":33240,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"4.1 Setup","offset":34789,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"4.2 Closed-form Instability and Transient Error","offset":36321,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"4.3 Asymptotic Conditions for Positive Transfer","offset":40364,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"Definition 7 (Growth Rate) .","offset":40867,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"Theorem 8 (Asymptotic behaviour of Transfer for Quadratic Objectives — Informal) .","offset":41902,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"4.4 The Continuous Drift Model","offset":44330,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"4.5 The Optimization Bias Matters","offset":46704,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-3.html#text"},{"title":"5 Generalizing CL beyond the JTL/ITL dichotomy","offset":48208,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"5.1 From implicit to explicit future models","offset":48260,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Definition 9 (Predictive CL Algorithm) .","offset":50296,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"5.2 Proof of Concept: the Window Predictive CL Algorithm","offset":51686,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Definition 10 (Mean Squared Displacement (MSD)) .","offset":52905,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Lemma 11 (Asymptotic Optimality of Window Size) .","offset":53544,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Beyond the Window: richer future models.","offset":55865,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"6 Interdisciplinary Literature Review","offset":56616,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Online Learning and Stochastic Optimization.","offset":57042,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Multi-Task Learning (MTL).","offset":59229,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Learning To Learn and Online Meta-Learning.","offset":60456,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"Continual Learning.","offset":62066,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-4.html#text"},{"title":"7 Discussion and Conclusion","offset":64981,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Limitations.","offset":69230,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Acknowledgements","offset":70179,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Appendix","offset":71222,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Appendix A Theoretical proofs","offset":71235,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"A.1 Introduction to the Framework of Kumar et al. (2025)","offset":71280,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"The Environment.","offset":71343,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Agent and Learning Algorithm.","offset":72141,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Continual Learning Objective.","offset":72822,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Supervised Specialization.","offset":74116,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"A.2 Analysis Setup and Formalism","offset":74985,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Task-Based Formulation.","offset":75288,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Definition (Task-Based Data Generating Process) .","offset":75412,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Batched Observations.","offset":75824,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"The Data-Generating Process.","offset":76285,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Measures of Error and Performance.","offset":76978,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"A.2.1 The Algorithm as a Markov Process","offset":77873,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-5.html#text"},{"title":"Convergence.","offset":79610,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Stochastic Gradient Descent Instantiations.","offset":80087,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"ITL SGD ( 𝒜 ITL \\mathcal{A}_{\\text{ITL}} ).","offset":80465,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"JTL SGD ( 𝒜 JTL \\mathcal{A}_{\\text{JTL}} ).","offset":80871,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"A.2.2 Why use reset ITL as the reference?","offset":81608,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"A.3 Decomposition: Instability and Transient Error","offset":83269,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Definition 12 (Task-wise Convergence) .","offset":83510,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Definition 13 (Instability ℐ k \\mathcal{I}^{k} ) .","offset":84382,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Definition 14 (Transient Error δ k \\delta^{k} ) .","offset":84869,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Lemma 15 (Exact TE Decomposition) .","offset":85553,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Proof.","offset":86011,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Lemma 16 (The Critical Task Duration) .","offset":86850,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Proof.","offset":87732,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"On the Sub-Linear Transient Error Assumption.","offset":88956,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"A.4 Exact Analysis in the Quadratic Regime","offset":90471,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Assumption 17 (Quadratic Tasks) .","offset":90793,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"A.4.1 SGD Dynamics on Quadratic Landscapes","offset":91967,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Lemma 18 (Convergence of the Mean) .","offset":92167,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Proof.","offset":92766,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Lemma 19 (Parameter Covariance) .","offset":93131,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Proof.","offset":93750,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-6.html#text"},{"title":"Lemma 20 (ITL Gradient Noise and Parameter Covariance) .","offset":95992,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Proof.","offset":96660,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Lemma 21 (JTL Gradient Noise Covariance) .","offset":99157,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Proof.","offset":99785,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"A.4.2 JTL Geometry","offset":101824,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Assumption 22 (Task Alignment) .","offset":102273,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Assumption 23 (Bounded Curvature) .","offset":103321,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"A.4.3 Exact Instability","offset":103804,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Theorem 24 (Exact Instability Decomposition) .","offset":104037,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Proof.","offset":104939,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"A.4.4 Exact Transient Error","offset":109085,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Theorem 25 (The Transient Potential Law) .","offset":109318,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-7.html#text"},{"title":"Geometric Interpretation.","offset":110539,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"Proof.","offset":111514,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"A.5 Asymptotic Dynamics of the First and Second Order Moments","offset":116741,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"Definition 26 (Growth Rate) .","offset":117620,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"A.5.1 Geometric Drift","offset":118158,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"Lemma 27 (Geometric Mean Drift) .","offset":118830,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"Proof.","offset":119295,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"Lemma 28 (Geometric Covariance Drift) .","offset":123210,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"Proof.","offset":123690,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-8.html#text"},{"title":"Lemma 29 (Expected Asymptotic Instability) .","offset":125773,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"Proof.","offset":126368,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"A.5.2 Residual Errors","offset":128738,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"Lemma 30 (Contraction of Residuals) .","offset":130879,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"Proof.","offset":131317,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"Lemma 31 (Asymptotic Dominance of Stable Recursions) .","offset":133553,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"Proof.","offset":134220,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"A.5.3 Asymptotic Transfer","offset":135601,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"Theorem 32 (Asymptotic Condition for Positive Transfer Efficiency) .","offset":135634,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"Proof.","offset":136265,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-9.html#text"},{"title":"A.6 (Non-asymptotic) Continuous Drift Analysis in the Quadratic Model","offset":142916,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-10.html#text"},{"title":"A.6.1 Setup and Definitions","offset":143237,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-10.html#text"},{"title":"A.6.2 Derivation of Expected Instability","offset":145138,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-10.html#text"},{"title":"A.6.3 Derivation of Expected Transient Error","offset":149900,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-10.html#text"},{"title":"ITL Agent.","offset":150378,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-10.html#text"},{"title":"JTL Agent.","offset":152173,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-10.html#text"},{"title":"A.6.4 Non-asymptotic Condition for Positive Transfer","offset":157805,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"A.7 Window Algorithm as an Instance of Predictive CL","offset":159725,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"A.7.1 Asymptotic Analysis in the Quadratic Loss Setting","offset":160749,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Definition 33 (Mean Squared Displacement) .","offset":161130,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Remark (Relationship between MSD and Growth Rate) .","offset":161792,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"A.7.2 Window Geometric Drift","offset":162297,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Lemma 34 (Window Geometric Drift) .","offset":163737,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Proof.","offset":164057,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Lemma 35 (Window Covariance Drift) .","offset":167103,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Proof.","offset":167629,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Lemma 36 (Asymptotic Instability of the Window Agent) .","offset":169824,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"Proof.","offset":170177,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-11.html#text"},{"title":"A.7.3 Asymptotically Optimal Window","offset":172334,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Lemma 37 (Asymptotic Optimality of Window Size) .","offset":172625,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Proof.","offset":173704,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Case 1: Bounded Regime.","offset":176371,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Case 2: Linear Drift Regime.","offset":176863,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Case 3: Diffusive Regime.","offset":177516,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"A.7.4 Window Agent for Continuously Drifting Tasks","offset":178455,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"A.7.5 Instability of the Window Agent","offset":179502,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Centroid Lag.","offset":179881,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Diversity Noise.","offset":181333,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"A.7.6 Transient Error of the Window Agent","offset":183692,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"A.7.7 The Optimal Window W ⋆ W^{\\star}","offset":185630,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-12.html#text"},{"title":"Appendix B Empirical Setup","offset":187842,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.1 Simulations","offset":187874,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.1.1 Shallow Linear Experiments","offset":188058,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.1.2 Deep Linear Experiments","offset":189357,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.1.3 Window Predictive CL Algorithm Experiments","offset":193209,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"Real-Time Trajectory Evaluation with W ⋆ W^{\\star} :","offset":193721,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"Transfer Efficiency Landscapes:","offset":194518,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.2 Real-World Benchmarks","offset":195551,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.2.1 CLEAR","offset":195583,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.2.2 MULTIDATASET (MD5)","offset":196196,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.2.3 Meta-World","offset":197326,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.3 Synthetic Benchmarks","offset":198597,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.3.1 Permuted CIFAR-10","offset":198628,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.3.2 Shuffled CIFAR-10","offset":199393,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"B.4 Training Procedures & Networks","offset":200201,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"Supervised Learning","offset":200242,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-13.html#text"},{"title":"Reinforcement Learning","offset":201855,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"B.4.1 Hyperparameters","offset":204979,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"Supervised Learning","offset":205008,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"Reinforcement Learning","offset":205693,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"B.5 Additional Figures and Empirical Substantiation","offset":209699,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"B.5.1 Pairwise Transfer is Asymmetric, and Why Ordering Matters","offset":210169,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"B.5.2 Proxy measures of instability","offset":213453,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"B.5.3 Scaling Deep Linear Networks Experiments","offset":214796,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"B.5.4 Real-time training trajectories","offset":216669,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-14.html#text"},{"title":"Instructions for reporting errors","offset":218793,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/part-15.html#text"}],"text_code_points":219757,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/source.md"}