{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,5,1]],"date-time":"2026-05-01T01:23:58Z","timestamp":1777598638410,"version":"3.51.4"},"reference-count":73,"publisher":"MIT Press - Journals","issue":"8","content-domain":{"domain":["direct.mit.edu"],"crossmark-restriction":true},"short-container-title":[],"published-print":{"date-parts":[[2021,7,26]]},"abstract":"<jats:p>Deep learning is often criticized by two serious issues that rarely exist in natural nervous systems: overfitting and catastrophic forgetting. It can even memorize randomly labeled data, which has little knowledge behind the instance-label pairs. When a deep network continually learns over time by accommodating new tasks, it usually quickly overwrites the knowledge learned from previous tasks. Referred to as the neural variability, it is well known in neuroscience that human brain reactions exhibit substantial variability even in response to the same stimulus. This mechanism balances accuracy and plasticity\/flexibility in the motor learning of natural nervous systems. Thus, it motivates us to design a similar mechanism, named artificial neural variability (ANV), that helps artificial neural networks learn some advantages from \u201cnatural\u201d neural networks. We rigorously prove that ANV plays as an implicit regularizer of the mutual information between the training data and the learned model. This result theoretically guarantees ANV a strictly improved generalizability, robustness to label noise, and robustness to catastrophic forgetting. We then devise a neural variable risk minimization (NVRM) framework and neural variable optimizers to achieve ANV for conventional network architectures in practice. The empirical studies demonstrate that NVRM can effectively relieve overfitting, label noise memorization, and catastrophic forgetting at negligible costs.<\/jats:p>","DOI":"10.1162\/neco_a_01403","type":"journal-article","created":{"date-parts":[[2021,6,7]],"date-time":"2021-06-07T21:40:02Z","timestamp":1623102002000},"page":"2163-2192","update-policy":"https:\/\/doi.org\/10.1162\/mitpressjournals.corrections.policy","source":"Crossref","is-referenced-by-count":40,"title":["Artificial Neural Variability for Deep Learning: On Overfitting, Noise Memorization, and Catastrophic Forgetting"],"prefix":"10.1162","volume":"33","author":[{"given":"Zeke","family":"Xie","sequence":"first","affiliation":[{"name":"University of Tokyo, Bunkyo-ku, Tokyo 113-0333, Japan and RIKEN Center for AIP, Chuo-ku, Tokyo 103-0027, Japan xie@ms.k.u-tokyo.ac.jp"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Fengxiang","family":"He","sequence":"additional","affiliation":[{"name":"University of Sydney, Level 1, Chippendale NSW 2008, Australia fengxiang.he@sydney.edu.au"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Shaopeng","family":"Fu","sequence":"additional","affiliation":[{"name":"University of Sydney, Level 1, Chippendale NSW 2008, Australia shfu7008@sydney.edu.au"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Issei","family":"Sato","sequence":"additional","affiliation":[{"name":"University of Tokyo, Bunkyo-ku, Tokyo 113-0333, Japan, and RIKEN Center for AIP, Chuo-ku, Tokyo 103-0027, Japan sato@k.u-tokyo.ac.jp"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Dacheng","family":"Tao","sequence":"additional","affiliation":[{"name":"University of Sydney, Level 1, Chippendale NSW 2008, Australi Dacheng.Tao@uts.edu.au"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Masashi","family":"Sugiyama","sequence":"additional","affiliation":[{"name":"RIKEN Center for AIP, Chuo-ku, Tokyo 103-0027, Japan, and University of Tokyo, Bunkyo-ku, Tokyo 113-0333, Japan sugi@k.u-tokyo.ac.jp"}],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"281","published-online":{"date-parts":[[2021,7,26]]},"reference":[{"key":"2021072618251065400_B1","author":"Achille","year":"2019","journal-title":"Where is the information in a deep neuralnetwork?"},{"issue":"1","key":"2021072618251065400_B2","first-page":"1947","article-title":"Emergence of invariance and disentanglement in deep representations","volume":"19","author":"Achille","year":"2018","journal-title":"Journal of Machine Learning Research"},{"key":"2021072618251065400_B3","first-page":"139","article-title":"Memory aware synapses: Learning what (not) to forget.","author":"Aljundi","year":"2018","journal-title":"Proceedings of the European Conference on Computer Vision"},{"key":"2021072618251065400_B4","first-page":"6155","volume-title":"Advances in neural information processing systems","author":"Allen-Zhu","year":"2019"},{"key":"2021072618251065400_B5","first-page":"242","article-title":"A convergence theory for deep learning via over-parameterization.","author":"Allen-Zhu","year":"2019","journal-title":"Proceedings of the International Conference on Machine Learning"},{"issue":"3","key":"2021072618251065400_B6","doi-asserted-by":"crossref","first-page":"643","DOI":"10.1162\/neco.1996.8.3.643","article-title":"The effects of adding noise during backpropagation training on a generalization performance","volume":"8","author":"An","year":"1996","journal-title":"Neural Computation"},{"key":"2021072618251065400_B7","article-title":"On the optimization of deep networks: Implicit acceleration by overparameterization.","author":"Arora","year":"2018","journal-title":"Proceedings of the 35th International Conference on Machine Learning"},{"key":"2021072618251065400_B8","first-page":"233","article-title":"A closer look at memorization in deep networks.","author":"Arpit","year":"2017","journal-title":"Proceedings of the International Conference on Machine Learning"},{"key":"2021072618251065400_B9","first-page":"1613","article-title":"Weight uncertainty in neural networks.","volume":"37","author":"Blundell","year":"2015","journal-title":"Proceedings of the 32nd International Conference on Machine Learning"},{"issue":"9","key":"2021072618251065400_B10","article-title":"Online learning and stochastic approximations","volume":"17","author":"Bottou","year":"1998","journal-title":"On-Line Learning in Neural Networks"},{"key":"2021072618251065400_B11","author":"Carbone","year":"2020","journal-title":"Robustness of Bayesian neural networks to gradient-based attacks"},{"issue":"6","key":"2021072618251065400_B12","doi-asserted-by":"crossref","first-page":"1215","DOI":"10.1162\/neco_a_01192","article-title":"Improving the antinoise ability of DNNs via a bio-inspired noise adaptive activation function rand softplus","volume":"31","author":"Chen","year":"2019","journal-title":"Neural Computation"},{"issue":"3","key":"2021072618251065400_B13","doi-asserted-by":"crossref","DOI":"10.1038\/nn.2501","article-title":"Stimulus onset quenches neural variability: A widespread cortical phenomenon","volume":"13","author":"Churchland","year":"2010","journal-title":"Nature Neuroscience"},{"issue":"14","key":"2021072618251065400_B14","doi-asserted-by":"crossref","first-page":"3697","DOI":"10.1523\/JNEUROSCI.3762-05.2006","article-title":"Neural variability in premotor cortex provides a signature of motor preparation","volume":"26","author":"Churchland","year":"2006","journal-title":"Journal of Neuroscience"},{"key":"2021072618251065400_B15","author":"Cover","year":"2012","journal-title":"Elements of information theory"},{"issue":"4","key":"2021072618251065400_B16","doi-asserted-by":"crossref","first-page":"303","DOI":"10.1007\/BF02551274","article-title":"Approximation by superpositions of a sigmoidal function","volume":"2","author":"Cybenko","year":"1989","journal-title":"Mathematics of Control, Signals and Systems"},{"key":"2021072618251065400_B17","first-page":"1019","article-title":"Sharp minima can generalize for deep nets.","author":"Dinh","year":"2017","journal-title":"Proceedings of the International Conference on Machine Learning"},{"issue":"6","key":"2021072618251065400_B18","doi-asserted-by":"crossref","first-page":"322","DOI":"10.1016\/j.tics.2015.04.005","article-title":"Neural variability: Friend or foe?","volume":"19","author":"Dinstein","year":"2015","journal-title":"Trends in Cognitive Sciences"},{"key":"2021072618251065400_B19","author":"Doan","year":"2020","journal-title":"A theoretical analysis of catastrophic forgetting through the NTK overlap matrix"},{"issue":"2","key":"2021072618251065400_B20","doi-asserted-by":"crossref","first-page":"674","DOI":"10.1137\/110831659","article-title":"Randomized smoothing for stochastic optimization","volume":"22","author":"Duchi","year":"2012","journal-title":"SIAM Journal on Optimization"},{"issue":"12","key":"2021072618251065400_B21","doi-asserted-by":"crossref","first-page":"1860","DOI":"10.2522\/ptj.2010090","article-title":"Perspective on variability in the development of human action","volume":"90","author":"Fetters","year":"2010","journal-title":"Physical Therapy"},{"issue":"3","key":"2021072618251065400_B22","doi-asserted-by":"crossref","first-page":"183","DOI":"10.1016\/0893-6080(89)90003-8","article-title":"On the approximate realization of continuous mappings by neural networks","volume":"2","author":"Funahashi","year":"1989","journal-title":"Neural Networks"},{"key":"2021072618251065400_B23","author":"Goodfellow","year":"2013","journal-title":"An empirical investigation of catastrophic forgetting in gradient-based neural networks"},{"key":"2021072618251065400_B24","first-page":"2348","volume-title":"Advances in neural information processing systems","author":"Graves","year":"2011"},{"key":"2021072618251065400_B25","first-page":"8527","volume-title":"Advances in neural information processing systems","author":"Han","year":"2018"},{"key":"2021072618251065400_B26","author":"Harutyunyan","year":"2020","journal-title":"Improving generalization by controlling label-noise information in neural network weights."},{"key":"2021072618251065400_B27","first-page":"1141","volume-title":"Advances in neural information processing systems","author":"He","year":"2019"},{"key":"2021072618251065400_B28","article-title":"Deep residual learning for image recognition.","author":"He","year":"2016","journal-title":"Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition"},{"issue":"2","key":"2021072618251065400_B29","doi-asserted-by":"crossref","first-page":"87","DOI":"10.1038\/nrn1323","article-title":"Insights into the ageing mind: A view from cognitive neuroscience","volume":"5","author":"Hedden","year":"2004","journal-title":"Nature Reviews Neuroscience"},{"key":"2021072618251065400_B30","doi-asserted-by":"crossref","first-page":"5","DOI":"10.1145\/168304.168306","article-title":"Keeping the neural networks simple by minimizing the description length of the weights.","author":"Hinton","year":"1993","journal-title":"Proceedings of the Sixth Annual Conference on Computational Learning Theory"},{"issue":"1","key":"2021072618251065400_B31","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1162\/neco.1997.9.1.1","article-title":"Flat minima","volume":"9","author":"Hochreiter","year":"1997","journal-title":"Neural Computation"},{"issue":"8","key":"2021072618251065400_B32","doi-asserted-by":"crossref","first-page":"1069","DOI":"10.1016\/S0893-6080(09)80018-X","article-title":"Some new results on neural network approximation","volume":"6","author":"Hornik","year":"1993","journal-title":"Neural Networks"},{"issue":"5","key":"2021072618251065400_B33","doi-asserted-by":"crossref","first-page":"359","DOI":"10.1016\/0893-6080(89)90020-8","article-title":"Multilayer feedforward networks are universal approximators","volume":"2","author":"Hornik","year":"1989","journal-title":"Neural Networks"},{"issue":"2","key":"2021072618251065400_B34","doi-asserted-by":"crossref","first-page":"330","DOI":"10.1162\/neco_a_01155","article-title":"Calculating the mutual information between two spike trains","volume":"31","author":"Houghton","year":"2019","journal-title":"Neural Computation"},{"key":"2021072618251065400_B35","author":"Jastrz\u0229bski","year":"2017","journal-title":"Three factors influencing minima in SGD."},{"issue":"7","key":"2021072618251065400_B36","doi-asserted-by":"crossref","first-page":"1462","DOI":"10.1162\/neco_a_01195","article-title":"Effect of depth and width on local minima in deep learning","volume":"31","author":"Kawaguchi","year":"2019","journal-title":"Neural Computation"},{"key":"2021072618251065400_B37","article-title":"On large-batch training for deep learning: Generalization gap and sharp minima.","author":"Keskar","year":"2017","journal-title":"Proceedings of the International Conference on Learning Representations"},{"key":"2021072618251065400_B38","first-page":"2611","article-title":"Fast and scalable Bayesian deep learning by weight-perturbation in Adam.","author":"Khan","year":"2018","journal-title":"Proceedings of the International Conference on Machine Learning"},{"key":"2021072618251065400_B39","author":"Kingma","year":"2014","journal-title":"Adam: A method for stochastic optimization."},{"issue":"13","key":"2021072618251065400_B40","doi-asserted-by":"crossref","first-page":"3521","DOI":"10.1073\/pnas.1611835114","article-title":"Overcoming catastrophic forgetting in neural networks","volume":"114","author":"Kirkpatrick","year":"2017","journal-title":"Proceedings of the National Academy of Sciences"},{"key":"2021072618251065400_B41","first-page":"5436","article-title":"Being Bayesian, even just a bit, fixes overconfidence in RElU networks.","author":"Kristiadi","year":"2020","journal-title":"Proceedings of the International Conference on Machine Learning"},{"key":"2021072618251065400_B42","author":"Krizhevsky","year":"2009","journal-title":"Learning multiple layers of features from tiny images"},{"key":"2021072618251065400_B43","author":"LeCun","year":"1998","journal-title":"The MNIST database of handwritten digits"},{"issue":"7553","key":"2021072618251065400_B44","doi-asserted-by":"crossref","DOI":"10.1038\/nature14539","article-title":"Deep learning","volume":"521","author":"LeCun","year":"2015","journal-title":"Nature"},{"key":"2021072618251065400_B45","first-page":"8157","volume-title":"Advances in neural information processing system","author":"Li","year":"2018"},{"key":"2021072618251065400_B46","doi-asserted-by":"crossref","first-page":"60","DOI":"10.1016\/j.media.2017.07.005","article-title":"A survey on deep learning in medical image analysis","volume":"42","author":"Litjens","year":"2017","journal-title":"Medical Image Analysis"},{"key":"2021072618251065400_B47","article-title":"Decoupled weight decay regularization.","author":"Loshchilov","year":"2018","journal-title":"Proceedings of the International Conference on Learning Representations"},{"key":"2021072618251065400_B48","doi-asserted-by":"crossref","first-page":"164","DOI":"10.1145\/307400.307435","article-title":"Pac-Bayesian model averaging.","author":"McAllester","year":"1999","journal-title":"Proceedings of the 12th Annual Conference on Computational Learning Theory"},{"issue":"3","key":"2021072618251065400_B49","doi-asserted-by":"crossref","first-page":"355","DOI":"10.1023\/A:1007618624809","article-title":"Some Pac-Bayesian theorems","volume":"37","author":"McAllester","year":"1999","journal-title":"Machine Learning"},{"key":"2021072618251065400_B50","doi-asserted-by":"crossref","first-page":"109","DOI":"10.1016\/S0079-7421(08)60536-8","article-title":"Catastrophic interference in connectionist networks: The sequential learning problem","volume":"24","author":"McCloskey","year":"1989","journal-title":"Psychology of Learning and Motivation"},{"issue":"2","key":"2021072618251065400_B51","doi-asserted-by":"crossref","first-page":"271","DOI":"10.1016\/j.bandc.2012.12.001","article-title":"Impact of Parkinson's disease and dopaminergic medication on adaptation to explicit and implicit visuomotor perturbations","volume":"81","author":"Mongeon","year":"2013","journal-title":"Brain and Cognition"},{"key":"2021072618251065400_B52","author":"Neelakantan","year":"2015","journal-title":"Adding gradient noise improves learning for very deep networks."},{"issue":"2","key":"2021072618251065400_B53","doi-asserted-by":"crossref","first-page":"527","DOI":"10.1007\/s10208-015-9296-2","article-title":"Random gradient-free minimization of convex functions","volume":"17","author":"Nesterov","year":"2017","journal-title":"Foundations of Computational Mathematics"},{"key":"2021072618251065400_B54","first-page":"5947","article-title":"Exploring generalization in deep learning.","volume":"30","author":"Neyshabur","year":"2017","journal-title":"Advances in neural information processing systems"},{"issue":"1","key":"2021072618251065400_B55","doi-asserted-by":"crossref","first-page":"386","DOI":"10.1152\/jn.00018.2011","article-title":"Changes in the neural control of a complex motor sequence during learning","volume":"106","author":"\u00d6lveczky","year":"2011","journal-title":"Journal of Neurophysiology"},{"key":"2021072618251065400_B56","doi-asserted-by":"crossref","first-page":"54","DOI":"10.1016\/j.neunet.2019.01.012","article-title":"Continual lifelong learning with neural networks: A review","volume":"113","author":"Parisi","year":"2019","journal-title":"Neural Networks"},{"key":"2021072618251065400_B57","first-page":"4510","article-title":"Mobilenetv2: Inverted residuals and linear bottlenecks.","author":"Sandler","year":"2018","journal-title":"Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition"},{"issue":"7587","key":"2021072618251065400_B58","doi-asserted-by":"crossref","DOI":"10.1038\/nature16961","article-title":"Mastering the game of go with deep neural networks and tree search","volume":"529","author":"Silver","year":"2016","journal-title":"Nature"},{"key":"2021072618251065400_B59","article-title":"Very deep convolutional networks for large-scale image recognition.","author":"Simonyan","year":"2015","journal-title":"Proceedings of the 3rd International Conference on Learning Representations"},{"issue":"5","key":"2021072618251065400_B60","doi-asserted-by":"crossref","first-page":"389","DOI":"10.1038\/nrn1668","article-title":"Neuronal variability: Noise or part of the signal?","volume":"6","author":"Stein","year":"2005","journal-title":"Nature Reviews Neuroscience"},{"key":"2021072618251065400_B61","article-title":"On the importance of initialization and momentum in deep learning.","author":"Sutskever","year":"2013","journal-title":"Proceedings of the International Conference on Machine Learning"},{"issue":"7173","key":"2021072618251065400_B62","doi-asserted-by":"crossref","first-page":"1240","DOI":"10.1038\/nature06390","article-title":"Performance variability enables adaptive plasticity of \u201ccrystallized\u201d adult birdsong","volume":"450","author":"Tumer","year":"2007","journal-title":"Nature"},{"key":"2021072618251065400_B63","first-page":"681","article-title":"Bayesian learning via stochastic gradient Langevin dynamics.","author":"Welling","year":"2011","journal-title":"Proceedings of the 28th International Conference on Machine Learning"},{"key":"2021072618251065400_B64","author":"Wen","year":"2018","journal-title":"Smoothout: Smoothing out sharp minima to improve generalization in deep learning."},{"key":"2021072618251065400_B65","author":"Witten","year":"2016","journal-title":"Data mining: Practical machine learning tools and techniques"},{"key":"2021072618251065400_B66","author":"Xie","year":"2020","journal-title":"Stable weight decay regularization"},{"key":"2021072618251065400_B67","article-title":"A diffusion theory for deep learning dynamics: Stochastic gradient descent exponentially favors flat minima.","author":"Xie","year":"2021","journal-title":"Proceedings of the International Conference on Learning Representations"},{"key":"2021072618251065400_B68","author":"Xie","year":"2020","journal-title":"Adai: Separating the effects of adaptive learning rate and momentum inertia."},{"key":"2021072618251065400_B69","first-page":"2524","volume-title":"Advances in neural information processing systems","author":"Xu","year":"2017"},{"key":"2021072618251065400_B70","first-page":"3987","article-title":"Continual learning through synaptic intelligence.","volume":"70","author":"Zenke","year":"2017","journal-title":"Proceedings of the 34th International Conference on Machine Learning"},{"key":"2021072618251065400_B71","author":"Zhang","year":"2017","journal-title":"Understanding deep learning requires rethinking generalization."},{"key":"2021072618251065400_B72","article-title":"Toward understanding the importance of noise in training neural networks.","author":"Zhou","year":"2019","journal-title":"Proceedings of the International Conference on Machine Learning"},{"key":"2021072618251065400_B73","first-page":"7654","article-title":"The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects.","author":"Zhu","year":"2019","journal-title":"Proceedings of the International Conference on Machine Learning"}],"container-title":["Neural Computation"],"original-title":[],"language":"en","link":[{"URL":"http:\/\/direct.mit.edu\/neco\/article-pdf\/33\/8\/2163\/1930889\/neco_a_01403.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"http:\/\/direct.mit.edu\/neco\/article-pdf\/33\/8\/2163\/1930889\/neco_a_01403.pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2021,7,26]],"date-time":"2021-07-26T18:27:06Z","timestamp":1627324026000},"score":1,"resource":{"primary":{"URL":"https:\/\/direct.mit.edu\/neco\/article\/33\/8\/2163\/101866\/Artificial-Neural-Variability-for-Deep-Learning-On"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2021,7,26]]},"references-count":73,"journal-issue":{"issue":"8","published-online":{"date-parts":[[2021,7,26]]},"published-print":{"date-parts":[[2021,7,26]]}},"URL":"https:\/\/doi.org\/10.1162\/neco_a_01403","relation":{},"ISSN":["0899-7667","1530-888X"],"issn-type":[{"value":"0899-7667","type":"print"},{"value":"1530-888X","type":"electronic"}],"subject":[],"published-other":{"date-parts":[[2021,8]]},"published":{"date-parts":[[2021,7,26]]}}}