{"id":"doi:10.1038/s41586-024-07711-7","title":"Maintaining Plasticity in Deep Continual Learning","date":"2023-06-23","section":"Data, training, optimization","quality":0.362714,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/","source_url":"https://doi.org/10.1038/s41586-024-07711-7","text_available":true,"authors":["Shibhansh Dohare","J. Fernando Hernandez-Garcia","Parash Rahman","A. Rupam Mahmood"],"outgoing":[{"source":"doi:10.1038/s41586-024-07711-7","target":"arxiv:1710.05941","evidence":"title","id":"arxiv:1710.05941","title":"Searching for Activation Functions","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8815985e018dbc2066450e5afbb9260b48d5b03a4f41bc23285755ccf3185454/"},{"source":"doi:10.1038/s41586-024-07711-7","target":"arxiv:1803.03635","evidence":"title","id":"arxiv:1803.03635","title":"The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a9d7f0235205afcc709e3e74f28cc7707831bcb1d46239c418cc7e4b624d23bd/"},{"source":"doi:10.1038/s41586-024-07711-7","target":"arxiv:2011.14826","evidence":"title","id":"arxiv:2011.14826","title":"Revisiting Rainbow: Promoting more Insightful and Inclusive Deep Reinforcement Learning Research","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2bf20803a5d9e686543fe5f3967e69b5aa481ba29ff91c9ca3fbeb147ab8c0aa/"},{"source":"doi:10.1038/s41586-024-07711-7","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":"doi:10.1038/s41586-024-07711-7","target":"arxiv:2205.07802","evidence":"title","id":"arxiv:2205.07802","title":"The Primacy Bias in Deep Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4e2aca370b941bb16bb6fefc359c420f6d4bde5a956bedaf0a84e6e71febc5bc/"},{"source":"doi:10.1038/s41586-024-07711-7","target":"arxiv:2305.19452","evidence":"title","id":"arxiv:2305.19452","title":"Bigger, Better, Faster: Human-level Atari with human-level efficiency","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/727a7ee43fd005401d34d4591a2f49d85175f9366ebf2c9e7a862b340787614e/"},{"source":"doi:10.1038/s41586-024-07711-7","target":"doi:10.1073/pnas.1611835114","evidence":"id","id":"doi:10.1073/pnas.1611835114","title":"Overcoming catastrophic forgetting in neural networks","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/849082368a44d06b93ef3ba95bcfddc6672b7a5e8064ad2b2d19601639266e8b/"}],"incoming":[{"source":"arxiv:2403.05175","target":"doi:10.1038/s41586-024-07711-7","evidence":"id","id":"arxiv:2403.05175","title":"Continual Learning and Catastrophic Forgetting","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3ed9ae586e8898d0235553d62c5f002609d3d112cbec54cc46655d19ff736400/"},{"source":"arxiv:2607.05609","target":"doi:10.1038/s41586-024-07711-7","evidence":"id","id":"arxiv:2607.05609","title":"To Retain or to Adapt? Generalizing Continual Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de2d99f43688431acd6fa456a50322ed0f00aa30d884a6767719f8a73b3b7c45/"},{"source":"arxiv:2607.07847","target":"doi:10.1038/s41586-024-07711-7","evidence":"title","id":"arxiv:2607.07847","title":"When Does Continual Learning Require Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/86e88c6cf21a66879eab5c156913fc85d14eb69c394c78c33c0e81d144e528eb/"}],"text_parts":[{"start":0,"end":15435,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html"},{"start":15435,"end":30859,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-2.html"},{"start":30859,"end":46576,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-3.html"},{"start":46576,"end":62011,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-4.html"},{"start":62011,"end":76861,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-5.html"},{"start":76861,"end":92351,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-6.html"},{"start":92351,"end":97615,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html"}],"contents":[{"title":"Loss of plasticity in deep continual learning","offset":549,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Abstract","offset":1017,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Similar content being viewed by others","offset":2390,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Engineering flexible machine learning systems by traversing functionally invariant paths","offset":2434,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Variational tensor neural networks for deep learning","offset":2528,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Physical deep learning with biologically inspired training method: gradient-free approach for physical hardware","offset":2586,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Explore related subjects","offset":2703,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Main","offset":2768,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Plasticity loss in supervised learning","offset":6969,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-1.html#text"},{"title":"Plasticity loss in reinforcement learning","offset":16782,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-2.html#text"},{"title":"Maintaining plasticity","offset":23278,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-2.html#text"},{"title":"Discussion","offset":27487,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-2.html#text"},{"title":"Methods","offset":29561,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-2.html#text"},{"title":"Specifics of continual backpropagation","offset":29573,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-2.html#text"},{"title":"Algorithm 1","offset":32788,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-3.html#text"},{"title":"Details of Continual ImageNet","offset":33963,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-3.html#text"},{"title":"Class-incremental CIFAR-100","offset":37771,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-3.html#text"},{"title":"Robust loss of plasticity in permuted MNIST","offset":43452,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-3.html#text"},{"title":"Loss of plasticity with different activations in the Slowly-Changing Regression problem","offset":48554,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-4.html#text"},{"title":"Understanding loss of plasticity","offset":54006,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-4.html#text"},{"title":"Existing deep-learning methods for mitigating loss of plasticity","offset":61941,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-4.html#text"},{"title":"Details and further analysis in reinforcement learning","offset":73572,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-5.html#text"},{"title":"Extended discussion","offset":81268,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-6.html#text"},{"title":"Data availability","offset":93029,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Code availability","offset":93252,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Supplementary information","offset":93350,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Peer Review File (download PDF )","offset":93380,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Rights and permissions","offset":93418,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"About this article","offset":94324,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Cite this article","offset":94347,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Share this article","offset":94755,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Comments","offset":94995,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Associated content","offset":96075,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Switching between tasks can cause AI to lose the ability to learn","offset":96098,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"AI can’t learn new things forever — an algorithm can fix that","offset":96197,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Search","offset":96316,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Quick links","offset":96327,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"nature.com footer links","offset":96491,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"About Nature Portfolio","offset":96519,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Discover content","offset":96599,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Publishing policies","offset":96684,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Author & Researcher services","offset":96749,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Libraries & institutions","offset":96902,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Advertising & partnerships","offset":97014,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Professional development","offset":97113,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"},{"title":"Regional websites","offset":97194,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/part-7.html#text"}],"text_code_points":97615,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5d46327415fe38f2b9160b874735530939961a71536f7f1ac4c0dc7a3af3224/source.md"}