{"id":"arxiv:2108.06325","title":"Continual Backprop: Stochastic Gradient Descent with Persistent Randomness","date":"2021-08-13","section":"Data, training, optimization","quality":-0.471382,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/","source_url":"https://arxiv.org/abs/2108.06325","text_available":true,"authors":["Shibhansh Dohare","Richard S. Sutton","A. Rupam Mahmood"],"outgoing":[{"source":"arxiv:2108.06325","target":"arxiv:1710.05941","evidence":"id","id":"arxiv:1710.05941","title":"Searching for Activation Functions","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8815985e018dbc2066450e5afbb9260b48d5b03a4f41bc23285755ccf3185454/"},{"source":"arxiv:2108.06325","target":"doi:10.1073/pnas.1611835114","evidence":"title","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:2205.07802","target":"arxiv:2108.06325","evidence":"id","id":"arxiv:2205.07802","title":"The Primacy Bias in Deep Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4e2aca370b941bb16bb6fefc359c420f6d4bde5a956bedaf0a84e6e71febc5bc/"},{"source":"arxiv:2607.05609","target":"arxiv:2108.06325","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":"doi:10.1038/s41586-024-07711-7","target":"arxiv:2108.06325","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/"}],"text_parts":[{"start":0,"end":15709,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html"},{"start":15709,"end":31652,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-2.html"},{"start":31652,"end":44252,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html"}],"contents":[{"title":"Continual Backprop: Stochastic Gradient Descent with Persistent Randomness","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"Abstract","offset":78,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"1 Introduction","offset":1229,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"2 Non-stationary problems","offset":5855,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"2.1 Bit-Flipping problem","offset":7465,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"2.2 Permuted MNIST","offset":10284,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"2.3 Slippery Ant","offset":11089,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"3 Backprop loses the ability to adapt under extended tracking","offset":11717,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-1.html#text"},{"title":"4 Continual Backprop","offset":16440,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-2.html#text"},{"title":"5 Related works","offset":29693,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-2.html#text"},{"title":"6 Conclusion and limitations","offset":32536,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix","offset":34251,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix A More experiments with Backprop on the Bit-Flipping Problem","offset":34264,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix B More experiments with CBP on the Bit-Flipping Problem","offset":35540,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix C Ablation Study for the utility measure","offset":36604,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix D Continual PPO","offset":38612,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix E Compute Used","offset":39585,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix F Affect of the speed of the distribution change on decaying plasticity","offset":40256,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix G A deeper look in the networks: Feature Saturation and Vanishing Gradients","offset":40833,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Appendix H Affect of network size on decaying plasticity","offset":42610,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Acknowledgements.","offset":43262,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"},{"title":"Instructions for reporting errors","offset":43288,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/part-3.html#text"}],"text_code_points":44252,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/73c90c52bffd8b5b2485d93607350545d822f361c29d3c0063b8d3025576e793/source.md"}