2000 character limit reached
Diagnosing Catastrophe: Large parts of accuracy loss in continual learning can be accounted for by readout misalignment (2310.05644v1)
Published 9 Oct 2023 in cs.LG and cs.CV
Abstract: Unlike primates, training artificial neural networks on changing data distributions leads to a rapid decrease in performance on old tasks. This phenomenon is commonly referred to as catastrophic forgetting. In this paper, we investigate the representational changes that underlie this performance decrease and identify three distinct processes that together account for the phenomenon. The largest component is a misalignment between hidden representations and readout layers. Misalignment occurs due to learning on additional tasks and causes internal representations to shift. Representational geometry is partially conserved under this misalignment and only a small part of the information is irrecoverably lost. All types of representational changes scale with the dimensionality of hidden representations. These insights have implications for deep learning applications that need to be continuously updated, but may also aid aligning ANN models to the rather robust biological vision.
- \APACrefYearMonthDay2022. \BBOQ\APACrefatitleProbing representation forgetting in supervised and unsupervised continual learning Probing representation forgetting in supervised and unsupervised continual learning.\BBCQ \BIn \APACrefbtitleProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Proceedings of the ieee/cvf conference on computer vision and pattern recognition (\BPGS 16712–16721). \PrintBackRefs\CurrentBib
- \APACinsertmetastarfrench1999catastrophic{APACrefauthors}French, R\BPBIM. \APACrefYearMonthDay1999. \BBOQ\APACrefatitleCatastrophic forgetting in connectionist networks Catastrophic forgetting in connectionist networks.\BBCQ \APACjournalVolNumPagesTrends in cognitive sciences34128–135. \PrintBackRefs\CurrentBib
- \APACinsertmetastargower1975generalized{APACrefauthors}Gower, J\BPBIC. \APACrefYearMonthDay1975. \BBOQ\APACrefatitleGeneralized procrustes analysis Generalized procrustes analysis.\BBCQ \APACjournalVolNumPagesPsychometrika4033–51. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2020. \BBOQ\APACrefatitleEmbracing change: Continual learning in deep neural networks Embracing change: Continual learning in deep neural networks.\BBCQ \APACjournalVolNumPagesTrends in cognitive sciences24121028–1040. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2017. \BBOQ\APACrefatitleOvercoming catastrophic forgetting in neural networks Overcoming catastrophic forgetting in neural networks.\BBCQ \APACjournalVolNumPagesProceedings of the national academy of sciences114133521–3526. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2009. \BBOQ\APACrefatitleLearning multiple layers of features from tiny images Learning multiple layers of features from tiny images.\BBCQ \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2021. \BBOQ\APACrefatitleContinual learning in deep networks: an analysis of the last layer Continual learning in deep networks: an analysis of the last layer.\BBCQ \APACjournalVolNumPagesarXiv preprint arXiv:2106.01834. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2022. \BBOQ\APACrefatitleEnergy-based models for continual learning Energy-based models for continual learning.\BBCQ \BIn \APACrefbtitleConference on Lifelong Learning Agents Conference on lifelong learning agents (\BPGS 1–22). \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2017. \BBOQ\APACrefatitleLearning without forgetting Learning without forgetting.\BBCQ \APACjournalVolNumPagesIEEE transactions on pattern analysis and machine intelligence40122935–2947. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay1989. \BBOQ\APACrefatitleCatastrophic interference in connectionist networks: The sequential learning problem Catastrophic interference in connectionist networks: The sequential learning problem.\BBCQ \BIn \APACrefbtitlePsychology of learning and motivation Psychology of learning and motivation (\BVOL 24, \BPGS 109–165). \APACaddressPublisherElsevier. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2022. \BBOQ\APACrefatitleWide neural networks forget less catastrophically Wide neural networks forget less catastrophically.\BBCQ \BIn \APACrefbtitleInternational Conference on Machine Learning International conference on machine learning (\BPGS 15699–15717). \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2019. \BBOQ\APACrefatitleContinual lifelong learning with neural networks: A review Continual lifelong learning with neural networks: A review.\BBCQ \APACjournalVolNumPagesNeural networks11354–71. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2020. \BBOQ\APACrefatitleAnatomy of catastrophic forgetting: Hidden representations and task semantics Anatomy of catastrophic forgetting: Hidden representations and task semantics.\BBCQ \APACjournalVolNumPagesarXiv preprint arXiv:2007.07400. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2022. \BBOQ\APACrefatitleEffect of scale on catastrophic forgetting in neural networks Effect of scale on catastrophic forgetting in neural networks.\BBCQ \BIn \APACrefbtitleInternational Conference on Learning Representations. International conference on learning representations. \PrintBackRefs\CurrentBib
- \APACinsertmetastartorgerson1952multidimensional{APACrefauthors}Torgerson, W\BPBIS. \APACrefYearMonthDay1952. \BBOQ\APACrefatitleMultidimensional scaling: I. Theory and method Multidimensional scaling: I. theory and method.\BBCQ \APACjournalVolNumPagesPsychometrika174401–419. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2019. \BBOQ\APACrefatitleThree scenarios for continual learning Three scenarios for continual learning.\BBCQ \APACjournalVolNumPagesarXiv preprint arXiv:1904.07734. \PrintBackRefs\CurrentBib
- \APACrefYearMonthDay2017. \BBOQ\APACrefatitleContinual learning through synaptic intelligence Continual learning through synaptic intelligence.\BBCQ \BIn \APACrefbtitleInternational conference on machine learning International conference on machine learning (\BPGS 3987–3995). \PrintBackRefs\CurrentBib
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Collections
Sign up for free to add this paper to one or more collections.