{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,6,4]],"date-time":"2026-06-04T20:05:29Z","timestamp":1780603529588,"version":"3.54.1"},"reference-count":77,"publisher":"Institute of Electrical and Electronics Engineers (IEEE)","issue":"6","license":[{"start":{"date-parts":[[2026,6,1]],"date-time":"2026-06-01T00:00:00Z","timestamp":1780272000000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/ieeexplore.ieee.org\/Xplorehelp\/downloads\/license-information\/IEEE.html"},{"start":{"date-parts":[[2026,6,1]],"date-time":"2026-06-01T00:00:00Z","timestamp":1780272000000},"content-version":"stm-asf","delay-in-days":0,"URL":"https:\/\/doi.org\/10.15223\/policy-029"},{"start":{"date-parts":[[2026,6,1]],"date-time":"2026-06-01T00:00:00Z","timestamp":1780272000000},"content-version":"stm-asf","delay-in-days":0,"URL":"https:\/\/doi.org\/10.15223\/policy-037"}],"funder":[{"DOI":"10.13039\/501100001809","name":"National Natural Science Foundation of China","doi-asserted-by":"publisher","award":["62402023"],"award-info":[{"award-number":["62402023"]}],"id":[{"id":"10.13039\/501100001809","id-type":"DOI","asserted-by":"publisher"}]},{"DOI":"10.13039\/501100001809","name":"National Natural Science Foundation of China","doi-asserted-by":"publisher","award":["62476181"],"award-info":[{"award-number":["62476181"]}],"id":[{"id":"10.13039\/501100001809","id-type":"DOI","asserted-by":"publisher"}]}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["IEEE Trans. Neural Netw. Learning Syst."],"published-print":{"date-parts":[[2026,6]]},"DOI":"10.1109\/tnnls.2025.3640733","type":"journal-article","created":{"date-parts":[[2025,12,17]],"date-time":"2025-12-17T18:47:14Z","timestamp":1765997234000},"page":"2839-2853","source":"Crossref","is-referenced-by-count":0,"title":["Spurious Local Minima Provably Exist for Deep CNNs: Theory and Application"],"prefix":"10.1109","volume":"37","author":[{"ORCID":"https:\/\/orcid.org\/0000-0002-3482-6930","authenticated-orcid":false,"given":"Bo","family":"Liu","sequence":"first","affiliation":[{"name":"College of Computer Science, Beijing University of Technology, Beijing, China"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Keyi","family":"Fu","sequence":"additional","affiliation":[{"name":"College of Computer Science, Beijing University of Technology, Beijing, China"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-8224-9891","authenticated-orcid":false,"given":"Tongtong","family":"Yuan","sequence":"additional","affiliation":[{"name":"College of Computer Science, Beijing University of Technology, Beijing, China"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Shen","family":"Geng","sequence":"additional","affiliation":[{"name":"College of Computer Science, Beijing University of Technology, Beijing, China"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"263","reference":[{"key":"ref1","doi-asserted-by":"publisher","DOI":"10.1109\/5.726791"},{"key":"ref2","doi-asserted-by":"publisher","DOI":"10.1145\/3065386"},{"key":"ref3","first-page":"1","article-title":"Very deep convolutional networks for large-scale image recognition","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Simonyan"},{"key":"ref4","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR.2015.7298594"},{"key":"ref5","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR.2016.90"},{"key":"ref6","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR.2017.243"},{"key":"ref7","doi-asserted-by":"publisher","DOI":"10.5555\/2969033.2969154"},{"key":"ref8","first-page":"1","article-title":"Qualitatively characterizing neural network optimization problems","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Goodfellow"},{"key":"ref9","article-title":"Theory II: Landscape of the empirical risk in deep learning","author":"Liao","year":"2017","journal-title":"arXiv:1703.09833"},{"key":"ref10","first-page":"1","article-title":"Topology and geometry of half-rectified network optimization","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Freeman"},{"key":"ref11","first-page":"1309","article-title":"Essentially no barriers in neural network energy landscape","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Draxler"},{"key":"ref12","first-page":"1","article-title":"Loss surfaces, mode connectivity, and fast ensembling of DNNs","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"Garipov"},{"key":"ref13","first-page":"1","article-title":"Visualizing the loss landscape of neural nets","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"Li"},{"key":"ref14","first-page":"1","article-title":"Ringing ReLUs: Harmonic distortion analysis of nonlinear feedforward networks","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Mehmeti-G\u00f6pel"},{"key":"ref15","doi-asserted-by":"publisher","DOI":"10.1016\/0893-6080(89)90014-2"},{"key":"ref16","first-page":"1","article-title":"Deep learning without poor local minima","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"Kawaguchi"},{"key":"ref17","article-title":"Depth creates no bad local minima","author":"Lu","year":"2017","journal-title":"arXiv:1702.08580"},{"key":"ref18","first-page":"2902","article-title":"Deep linear networks with arbitrary loss: All local minima are global","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Laurent"},{"key":"ref19","first-page":"1","article-title":"Global optimality conditions for deep neural networks","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Yun"},{"key":"ref20","article-title":"Learning deep models: Critical points and local openness","author":"Nouiehed","year":"2018","journal-title":"arXiv:1803.02968"},{"key":"ref21","article-title":"Depth creates no more spurious local minima","author":"Zhang","year":"2019","journal-title":"arXiv:1901.09827"},{"key":"ref22","first-page":"1","article-title":"Exact solutions of a deep linear network","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"Ziyin"},{"key":"ref23","first-page":"8279","article-title":"How SGD selects the global minima in over-parameterized learning: A dynamical stability perspective","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","volume":"31","author":"Wu"},{"key":"ref24","doi-asserted-by":"publisher","DOI":"10.52202\/068431-0338"},{"key":"ref25","first-page":"1","article-title":"The implicit bias of minima stability in multivariate shallow ReLU networks","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Nacson"},{"key":"ref26","first-page":"1","article-title":"A loss curvature perspective on training instability in deep learning","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Gilmer"},{"key":"ref27","first-page":"1","article-title":"A precise characterization of SGD stability using loss surface geometry","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Dexter"},{"key":"ref28","first-page":"1","article-title":"Matrix completion has no spurious local minimum","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"Ge"},{"key":"ref29","doi-asserted-by":"publisher","DOI":"10.1109\/TIT.2018.2854560"},{"key":"ref30","first-page":"1329","article-title":"On the power of over-parametrization in neural networks with quadratic activation","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Du"},{"key":"ref31","first-page":"1","article-title":"Identity matters in deep learning","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Hardt"},{"key":"ref32","article-title":"Avoiding spurious local minima in deep quadratic networks","author":"Kazemipour","year":"2020","journal-title":"arXiv:2001.00098"},{"key":"ref33","first-page":"1","article-title":"Spurious local minima are common in two-layer ReLU neural networks","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Safran"},{"key":"ref34","article-title":"Local minima in training of neural networks","author":"Swirszcz","year":"2016","journal-title":"arXiv:1611.06310"},{"key":"ref35","first-page":"1","article-title":"Critical points of neural networks: Analytical forms and landscape properties","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Zhou"},{"key":"ref36","first-page":"1","article-title":"Small nonlinearities in activation functions create bad local minima in neural networks","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Yun"},{"key":"ref37","doi-asserted-by":"publisher","DOI":"10.1287\/moor.2021.1228"},{"key":"ref38","first-page":"1","article-title":"Bounds on over-parameterization for guaranteed existence of descent paths in shallow ReLU networks","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Sharifnassab"},{"key":"ref39","first-page":"1","article-title":"Piecewise linear activations substantially shape the loss surfaces of neural networks","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"He"},{"key":"ref40","first-page":"1","article-title":"Truth or backpropaganda? An empirical investigation of deep learning theory","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Goldblum"},{"key":"ref41","article-title":"Exponentially vanishing sub-optimal local minima in multilayer neural networks","author":"Soudry","year":"2017","journal-title":"arXiv:1702.05777"},{"key":"ref42","first-page":"1","article-title":"The hidden convex optimization landscape of regularized two-layer relu networks: An exact characterization of optimal solutions","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Wang"},{"key":"ref43","first-page":"1","article-title":"Analytic study of families of spurious minima in two-layer ReLU neural networks: A tale of symmetry II","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","volume":"34","author":"Arjevani"},{"key":"ref44","doi-asserted-by":"publisher","DOI":"10.52202\/068431-2719"},{"key":"ref45","first-page":"1","article-title":"Does a sparse ReLU network training problem always admit an optimum?","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"Le"},{"key":"ref46","doi-asserted-by":"publisher","DOI":"10.1016\/j.neunet.2021.08.005"},{"key":"ref47","first-page":"1339","article-title":"Gradient descent learns one-hidden-layer CNN: Don\u2019t be afraid of spurious local minima","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Du"},{"key":"ref48","first-page":"7594","article-title":"Toward understanding the importance of noise in training neural networks","volume-title":"Proc. 36th Int. Conf. Mach. Learn.","author":"Zhou"},{"key":"ref49","first-page":"605","article-title":"Globally optimal gradient descent for a ConvNet with Gaussian inputs","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Brutzkus"},{"key":"ref50","first-page":"3730","article-title":"Optimization landscape and expressivity of deep CNNs","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Nguyen"},{"key":"ref51","doi-asserted-by":"publisher","DOI":"10.1109\/TNNLS.2022.3204319"},{"key":"ref52","article-title":"Spurious valleys in two-layer neural network optimization landscapes","author":"Venturi","year":"2018","journal-title":"arXiv:1802.06384"},{"key":"ref53","first-page":"4790","article-title":"On connected sublevel sets in deep learning","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Nguyen"},{"key":"ref54","first-page":"1","article-title":"On the loss landscape of a class of deep neural networks with no bad local valleys","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Nguyen"},{"key":"ref55","article-title":"On the benefit of width for neural networks: Disappearance of bad basins","author":"Li","year":"2018","journal-title":"arXiv:1812.11039"},{"key":"ref56","first-page":"1","article-title":"Asymmetric valleys: Beyond sharp and flat local minima","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"He"},{"key":"ref57","first-page":"1","article-title":"Taxonomizing local versus global structure in neural network loss landscapes","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","author":"Yang"},{"key":"ref58","first-page":"9722","article-title":"Geometry of the loss landscape in overparameterized neural networks: Symmetries and invariances","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Simsek"},{"key":"ref59","doi-asserted-by":"publisher","DOI":"10.1088\/1742-5468\/ac9832"},{"key":"ref60","first-page":"1","article-title":"Symmetries, flat minima, and the conserved quantities of gradient flow","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Zhao"},{"key":"ref61","first-page":"2798","article-title":"Geometry of neural network loss surfaces via random matrix theory","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Pennington"},{"key":"ref62","first-page":"1","article-title":"The Hessian perspective into the nature of convolutional neural networks","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Singh"},{"key":"ref63","first-page":"1","article-title":"A diffusion theory for deep learning dynamics: Stochastic gradient descent exponentially favors flat minima","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Xie"},{"key":"ref64","first-page":"1","article-title":"Sharpness-aware minimization for efficiently improving generalization","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Foret"},{"key":"ref65","doi-asserted-by":"publisher","DOI":"10.52202\/075280-3165"},{"key":"ref66","first-page":"118701","article-title":"Improving generalization and convergence by enhancing implicit regularization","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","volume":"37","author":"Weinan"},{"key":"ref67","first-page":"1","article-title":"Sharpness-aware minimization and the edge of stability","volume":"25","author":"Long","year":"2024","journal-title":"J. Mach. Learn. Res."},{"key":"ref68","first-page":"1","article-title":"Sharpness-aware minimization efficiently selects flatter minima late in training","volume-title":"Proc. Int. Conf. Learn. Represent.","author":"Zhou"},{"key":"ref69","first-page":"1675","article-title":"Gradient descent finds global minima of deep neural networks","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Du"},{"key":"ref70","first-page":"1","article-title":"A convergence theory for deep learning via over-parameterization","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Allen-Zhu"},{"key":"ref71","doi-asserted-by":"publisher","DOI":"10.1007\/s10994-019-05839-6"},{"key":"ref72","first-page":"8056","article-title":"On the proof of global convergence of gradient descent for deep ReLU networks with linear widths","volume-title":"Proc. Int. Conf. Mach. Learn.","author":"Nguyen"},{"key":"ref73","doi-asserted-by":"publisher","DOI":"10.52202\/075280-1070"},{"key":"ref74","first-page":"13868","article-title":"Semi-flat minima and saddle points by embedding neural networks to overparameterization","volume-title":"Proc. Adv. Neural Inf. Process. Syst.","volume":"32","author":"Fukumizu"},{"key":"ref75","doi-asserted-by":"publisher","DOI":"10.1016\/S0893-6080(00)00009-5"},{"key":"ref76","doi-asserted-by":"publisher","DOI":"10.4208\/jml.220108"},{"key":"ref77","doi-asserted-by":"publisher","DOI":"10.1017\/CBO9780511804441"}],"container-title":["IEEE Transactions on Neural Networks and Learning Systems"],"original-title":[],"link":[{"URL":"http:\/\/xplorestaging.ieee.org\/ielx8\/5962385\/11549955\/11301825.pdf?arnumber=11301825","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,6,4]],"date-time":"2026-06-04T19:56:48Z","timestamp":1780603008000},"score":1,"resource":{"primary":{"URL":"https:\/\/ieeexplore.ieee.org\/document\/11301825\/"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,6]]},"references-count":77,"journal-issue":{"issue":"6"},"URL":"https:\/\/doi.org\/10.1109\/tnnls.2025.3640733","relation":{},"ISSN":["2162-237X","2162-2388"],"issn-type":[{"value":"2162-237X","type":"print"},{"value":"2162-2388","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,6]]}}}