{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2024,8,7]],"date-time":"2024-08-07T12:38:37Z","timestamp":1723034317858},"reference-count":40,"publisher":"MIT Press - Journals","issue":"4","content-domain":{"domain":["direct.mit.edu"],"crossmark-restriction":true},"short-container-title":[],"published-print":{"date-parts":[[2021,3,26]]},"abstract":"<jats:p>We study the learning dynamics and the representations emerging in recurrent neural networks (RNNs) trained to integrate one or multiple temporal signals. Combining analytical and numerical investigations, we characterize the conditions under which an RNN with n neurons learns to integrate D(\u226an) scalar signals of arbitrary duration. We show, for linear, ReLU, and sigmoidal neurons, that the internal state lives close to a D-dimensional manifold, whose shape is related to the activation function. Each neuron therefore carries, to various degrees, information about the value of all integrals. We discuss the deep analogy between our results and the concept of mixed selectivity forged by computational neuroscientists to interpret cortical recordings.<\/jats:p>","DOI":"10.1162\/neco_a_01366","type":"journal-article","created":{"date-parts":[[2021,1,29]],"date-time":"2021-01-29T23:13:43Z","timestamp":1611962023000},"page":"1063-1112","update-policy":"http:\/\/dx.doi.org\/10.1162\/mitpressjournals.corrections.policy","source":"Crossref","is-referenced-by-count":1,"title":["Low-Dimensional Manifolds Support Multiplexed Integrations in Recurrent Neural Networks"],"prefix":"10.1162","volume":"33","author":[{"given":"Arnaud","family":"Fanthomme","sequence":"first","affiliation":[{"name":"Laboratoire de Physique de l'Ecole Normale Sup\u00e9rieure PSLand CNRS UMR 8023, Sorbonne Universit\u00e9, 75005 Paris, France, arnaud.fanthomme@phys.ens.fr"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"R\u00e9mi","family":"Monasson","sequence":"additional","affiliation":[{"name":"Laboratoire de Physique de l'Ecole Normale Sup\u00e9rieure PSLand CNRS UMR 8023, Sorbonne Universit\u00e9, 75005 Paris, France, remi.monasson@phys.ens.fr"}],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"281","published-online":{"date-parts":[[2021,3,26]]},"reference":[{"issue":"4","key":"2021041320044075500_B1","doi-asserted-by":"crossref","first-page":"494","DOI":"10.1038\/nn1877","article-title":"Functional dissection of circuitry in a neural integrator","volume":"10","author":"Aksay","year":"2007","journal-title":"Nature Neuroscience"},{"issue":"21","key":"2021041320044075500_B2","doi-asserted-by":"crossref","first-page":"3387","DOI":"10.1093\/bioinformatics\/btx431","article-title":"DeepLoc: Prediction of protein subcellular localization using deep learning","volume":"33","author":"Almagro Armenteros","year":"2017","journal-title":"Bioinformatics"},{"key":"2021041320044075500_B3","author":"Amodei","year":"2015","journal-title":"Deep Speech 2: End-to-end speech recognition in English and Mandarin"},{"key":"2021041320044075500_B4","author":"Arora","year":"2019","journal-title":"Implicit regularization in deep matrix factorization."},{"issue":"7705","key":"2021041320044075500_B5","doi-asserted-by":"crossref","DOI":"10.1038\/s41586-018-0102-6","article-title":"Vector-based navigation using grid-like representations in artificial agents","volume":"557","author":"Banino","year":"2018","journal-title":"Nature"},{"key":"2021041320044075500_B6","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1016\/j.conb.2017.06.003","article-title":"Recurrent neural networks as versatile tools of neuroscience research","volume":"46","author":"Barak","year":"2017","journal-title":"Current Opinion in Neurobiology"},{"key":"2021041320044075500_B7","author":"Chung","year":"2014","journal-title":"Empirical evaluation of gated recurrent neural networks on sequence modeling."},{"key":"2021041320044075500_B8","author":"Collins","year":"2017","journal-title":"Capacity and trainability in recurrent neural networks."},{"key":"2021041320044075500_B9","first-page":"4414","volume-title":"Advances in neural information processing systems","author":"Denton","year":"2017"},{"issue":"2","key":"2021041320044075500_B10","doi-asserted-by":"crossref","first-page":"179","DOI":"10.1207\/s15516709cog1402_1","article-title":"Finding structure in time","volume":"14","author":"Elman","year":"1990","journal-title":"Cognitive Science"},{"key":"2021041320044075500_B11","first-page":"2663","article-title":"Understanding and controlling memory in recurrent neural networks","author":"Haviv","year":"2019","journal-title":"Proceedings of the International Conference on Machine Learning"},{"key":"2021041320044075500_B12","author":"Kingma","year":"2017","journal-title":"Adam: A method for stochastic optimization"},{"key":"2021041320044075500_B13","author":"Leavitt","year":"2020","journal-title":"Selectivity considered harmful: Evaluating the causal impact of class selectivity in DNNs"},{"key":"2021041320044075500_B14","author":"Leavitt","year":"2020","journal-title":"On the relationship between class selectivity, dimensionality, and robustness."},{"key":"2021041320044075500_B15","doi-asserted-by":"crossref","first-page":"371","DOI":"10.1007\/978-1-4757-9800-5_60","volume-title":"Computational neuroscience","author":"Lee","year":"1997"},{"key":"2021041320044075500_B16","first-page":"35","article-title":"Diverse image-to-image translation via disentangled representations","author":"Lee","year":"2018","journal-title":"Proceedings of the European Conference on Computer Vision"},{"key":"2021041320044075500_B17","author":"Li","year":"2017","journal-title":"Understanding neural networks through representation erasure"},{"key":"2021041320044075500_B18","author":"Lipton","year":"2015","journal-title":"A critical review of recurrent neural networks for sequence learning."},{"key":"2021041320044075500_B19","author":"Luong","year":"2016","journal-title":"Multitask sequence to sequence learning."},{"issue":"7474","key":"2021041320044075500_B20","doi-asserted-by":"crossref","first-page":"78","DOI":"10.1038\/nature12742","article-title":"Context-dependent computation by recurrent dynamics in prefrontal cortex","volume":"503","author":"Mante","year":"2013","journal-title":"Nature"},{"issue":"3","key":"2021041320044075500_B21","doi-asserted-by":"crossref","first-page":"609","DOI":"10.1016\/j.neuron.2018.07.003","article-title":"Linking connectivity, dynamics and computations in low-rank recurrent neural networks","volume":"99","author":"Mastrogiuseppe","year":"2018","journal-title":"Neuron"},{"key":"2021041320044075500_B22","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1016\/j.dsp.2017.10.011","article-title":"Methods for interpreting and understanding deep neural networks","volume":"73","author":"Montavon","year":"2018","journal-title":"Digital Signal Processing"},{"issue":"3","key":"2021041320044075500_B23","doi-asserted-by":"crossref","DOI":"10.23915\/distill.00010","article-title":"The building blocks of interpretability","volume":"3","author":"Olah","year":"2018","journal-title":"Distill"},{"issue":"10","key":"2021041320044075500_B24","doi-asserted-by":"crossref","first-page":"1345","DOI":"10.1109\/TKDE.2009.191","article-title":"A survey on transfer learning","volume":"22","author":"Pan","year":"2010","journal-title":"IEEE Transactions on Knowledge and Data Engineering"},{"key":"2021041320044075500_B25","first-page":"8024","volume-title":"Advances in neural information processing systems","author":"Paszke","year":"2019"},{"key":"2021041320044075500_B26","author":"Radford","year":"2017","journal-title":"Learning to generate reviews and discovering sentiment."},{"issue":"11","key":"2021041320044075500_B27","doi-asserted-by":"crossref","first-page":"1761","DOI":"10.1038\/s41593-019-0520-2","article-title":"A deep learning framework for neuroscience","volume":"22","author":"Richards","year":"2019","journal-title":"Nature Neuroscience"},{"issue":"7451","key":"2021041320044075500_B28","doi-asserted-by":"crossref","first-page":"585","DOI":"10.1038\/nature12160","article-title":"The importance of mixed selectivity in complex cognitive asks","volume":"497","author":"Rigotti","year":"2013","journal-title":"Nature"},{"issue":"1","key":"2021041320044075500_B29","doi-asserted-by":"crossref","first-page":"33","DOI":"10.1146\/annurev.ne.12.030189.000341","article-title":"Integrating with neurons","volume":"12","author":"Robinson","year":"1989","journal-title":"Annual Review of Neuroscience"},{"key":"2021041320044075500_B30","author":"Saxe","year":"2014","journal-title":"Exact solutions to the nonlinear dynamics of learning in deep linear neural networks"},{"issue":"1","key":"2021041320044075500_B31","doi-asserted-by":"crossref","DOI":"10.1103\/PhysRevResearch.2.013111","article-title":"Dynamics of random recurrent networks with correlated low-rank structure","volume":"2","author":"Schuessler","year":"2020","journal-title":"Physical Review Research"},{"key":"2021041320044075500_B32","author":"Schuessler","year":"2020","journal-title":"The interplay between randomness and structure during learning in RNNs"},{"issue":"23","key":"2021041320044075500_B33","doi-asserted-by":"crossref","first-page":"13339","DOI":"10.1073\/pnas.93.23.13339","article-title":"How the brain keeps the eyes still","volume":"93","author":"Seung","year":"1996","journal-title":"Proceedings of the National Academy of Sciences"},{"issue":"2","key":"2021041320044075500_B34","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1371\/journal.pcbi.1004792","article-title":"Training excitatory-inhibitory recurrent neural networks for cognitive tasks: A simple and flexible framework","volume":"12","author":"Song","year":"2016","journal-title":"PLOS Computational Biology"},{"key":"2021041320044075500_B35","volume-title":"Fundamental neuroscience","author":"Squire","year":"2012","edition":"4th"},{"issue":"3","key":"2021041320044075500_B36","doi-asserted-by":"crossref","first-page":"626","DOI":"10.1162\/NECO_a_00409","article-title":"Opening the black box: Low-dimensional dynamics in high-dimensional recurrent neural networks","volume":"25","author":"Sussillo","year":"2012","journal-title":"Neural Computation"},{"key":"2021041320044075500_B37","doi-asserted-by":"crossref","first-page":"100","DOI":"10.1016\/j.neunet.2019.03.005","article-title":"Recent advances in physical reservoir computing: A review","volume":"115","author":"Tanaka","year":"2019","journal-title":"Neural Networks"},{"issue":"3","key":"2021041320044075500_B38","doi-asserted-by":"crossref","first-page":"261","DOI":"10.1038\/s41592-019-0686-2","article-title":"SciPy 1.0: Fundamental algorithms for computing in Python","volume":"17","author":"Virtanen","year":"2020","journal-title":"Nature Methods"},{"issue":"4","key":"2021041320044075500_B39","doi-asserted-by":"crossref","first-page":"1314","DOI":"10.1523\/JNEUROSCI.3733-05.2006","article-title":"A recurrent network mechanism of time integration in perceptual decisions","volume":"26","author":"Wong","year":"2006","journal-title":"Journal of Neuroscience"},{"issue":"1","key":"2021041320044075500_B40","doi-asserted-by":"crossref","first-page":"27","DOI":"10.1631\/FITEE.1700808","article-title":"Visual interpretability for deep learning: A survey","volume":"19","author":"Zhang","year":"2018","journal-title":"Frontiers of Information Technology and Electronic Engineering"}],"container-title":["Neural Computation"],"original-title":[],"language":"en","link":[{"URL":"http:\/\/direct.mit.edu\/neco\/article-pdf\/33\/4\/1063\/1902263\/neco_a_01366.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"http:\/\/direct.mit.edu\/neco\/article-pdf\/33\/4\/1063\/1902263\/neco_a_01366.pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2021,4,15]],"date-time":"2021-04-15T00:15:14Z","timestamp":1618445714000},"score":1,"resource":{"primary":{"URL":"https:\/\/direct.mit.edu\/neco\/article\/33\/4\/1063\/97474\/Low-Dimensional-Manifolds-Support-Multiplexed"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2021]]},"references-count":40,"journal-issue":{"issue":"4","published-online":{"date-parts":[[2021,3,26]]},"published-print":{"date-parts":[[2021,3,26]]}},"URL":"https:\/\/doi.org\/10.1162\/neco_a_01366","relation":{},"ISSN":["0899-7667","1530-888X"],"issn-type":[{"value":"0899-7667","type":"print"},{"value":"1530-888X","type":"electronic"}],"subject":[],"published-other":{"date-parts":[[2021]]},"published":{"date-parts":[[2021]]}}}