{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"institution":[{"name":"bioRxiv"}],"indexed":{"date-parts":[[2026,10,24]],"date-time":"2026-10-24T14:37:04Z","timestamp":1792852624066,"version":"4.5.0"},"posted":{"date-parts":[[2026,9,14]]},"group-title":"Neuroscience","reference-count":24,"publisher":"openRxiv","license":[{"start":{"date-parts":[[2026,9,14]],"date-time":"2026-09-14T00:00:00Z","timestamp":1789344000000},"content-version":"vor","delay-in-days":0,"URL":"http:\/\/creativecommons.org\/licenses\/by-nc-nd\/4.0\/"}],"funder":[{"name":"Shanahan Family Foundation"},{"award":["2223725"],"award-info":[{"award-number":["2223725"]}],"id":[{"id":"https:\/\/ror.org\/021nxhr62","id-type":"ROR","asserted-by":"publisher"}]},{"award":["2424124"],"award-info":[{"award-number":["2424124"]}],"id":[{"id":"https:\/\/ror.org\/021nxhr62","id-type":"ROR","asserted-by":"publisher"}]},{"award":["R01EB029813"],"award-info":[{"award-number":["R01EB029813"]}],"id":[{"id":"https:\/\/ror.org\/01cwqze88","id-type":"ROR","asserted-by":"publisher"}]},{"award":["RF1DA055669"],"award-info":[{"award-number":["RF1DA055669"]}],"id":[{"id":"https:\/\/ror.org\/01cwqze88","id-type":"ROR","asserted-by":"publisher"}]}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"accepted":{"date-parts":[[2026,10,21]]},"abstract":"<jats:title>Abstract<\/jats:title>\n                <jats:p>Biological neural systems learn new behaviors while retaining earlier ones, whereas sequentially trained artificial networks often overwrite parameters and forget previously learned behaviors. We introduce the Context-modulated Synaptic-states Recurrent Neural Network (CoSyn-RNN), a new continual-learning model inspired by heterogeneous synaptic states and neuromodulatory control. CoSyn-RNN combines sparse recurrent allocation with a discrete protection rule: after each task, recurrent neurons whose synaptic weights cross a fixed threshold form a task-specific neuron group, and their associated parameters enter a protected state during later learning. A neuron-specific gain and a task-cued mask over a fixed base readout control how recurrent activity contributes to behavior. Across various task sequences consisting of many cognitive tasks, training and validation losses and accuracies were maintained throughout learning. The protection rule produced modular, task-ordered recurrent structure compatible with forward reuse of earlier computations. Pairwise recruitment measurements further defined minimum- and maximum-cost curricula that produced markedly different final recurrent structures and different remaining capacity for future learning. Thus, CoSyn-RNN provides a biologically inspired model in which protected synaptic states, sparse allocation, and context-dependent modulation support continual learning.<\/jats:p>","DOI":"10.64898\/2026.09.11.750704","type":"posted-content","created":{"date-parts":[[2026,9,14]],"date-time":"2026-09-14T20:35:13Z","timestamp":1789418113000},"source":"Crossref","is-referenced-by-count":0,"title":["Discrete synaptic states and context-modulated readouts support continual learning"],"prefix":"10.64898","author":[{"ORCID":"https:\/\/orcid.org\/0009-0005-7850-7658","authenticated-orcid":false,"given":"Yuan","family":"Gao","sequence":"first","affiliation":[{"name":"Allen Institute, Seattle, WA, USA"},{"name":"University of Washington, Seattle, WA, USA"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-2629-7100","authenticated-orcid":false,"given":"Stefan","family":"Mihalas","sequence":"additional","affiliation":[{"name":"Allen Institute, Seattle, WA, USA"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-2955-2568","authenticated-orcid":false,"given":"Denis","family":"Turcu","sequence":"additional","affiliation":[{"name":"Allen Institute, Seattle, WA, USA"},{"name":"University of Washington, Seattle, WA, USA"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"54368","reference":[{"key":"2026102406550634000_2026.09.11.750704v2.1","doi-asserted-by":"publisher","DOI":"10.1038\/385533a0"},{"key":"2026102406550634000_2026.09.11.750704v2.2","doi-asserted-by":"publisher","DOI":"10.1126\/science.8036517"},{"key":"2026102406550634000_2026.09.11.750704v2.3","doi-asserted-by":"publisher","DOI":"10.1073\/pnas.1803839115"},{"issue":"12","key":"2026102406550634000_2026.09.11.750704v2.4","doi-asserted-by":"crossref","first-page":"1185","DOI":"10.1038\/s42256-022-00568-3","article-title":"Three types of incremental learning","volume":"4","year":"2022","journal-title":"Nature Machine Intelligence"},{"key":"2026102406550634000_2026.09.11.750704v2.5","doi-asserted-by":"publisher","DOI":"10.1016\/j.neuron.2005.02.001"},{"key":"2026102406550634000_2026.09.11.750704v2.6","doi-asserted-by":"publisher","DOI":"10.1073\/pnas.0905110106"},{"issue":"1","key":"2026102406550634000_2026.09.11.750704v2.7","first-page":"2020","article-title":"Endoplasmic reticulum visits highly active spines and prevents runaway potentiation of synapses","volume":"11","year":"5083","journal-title":"Nature Communications"},{"key":"2026102406550634000_2026.09.11.750704v2.8","doi-asserted-by":"publisher","DOI":"10.1038\/s41593-021-00824-6"},{"key":"2026102406550634000_2026.09.11.750704v2.9","doi-asserted-by":"publisher","DOI":"10.1073\/pnas.1611835114"},{"key":"2026102406550634000_2026.09.11.750704v2.10","unstructured":"Friedemann Zenke , Ben Poole , and Surya Ganguli . Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987\u20133995. Pmlr, 2017."},{"issue":"12","key":"2026102406550634000_2026.09.11.750704v2.11","first-page":"2935","article-title":"Learning without forgetting","volume":"40","year":"2017","journal-title":"IEEE transactions on pattern analysis and machine intelligence"},{"key":"2026102406550634000_2026.09.11.750704v2.12","doi-asserted-by":"crossref","unstructured":"Romik Ghosh , Dana Mastrovito , and Stefan Mihalas . A biologically plausible model for continual learning using synaptic weight attractors. bioRxiv, pages 2021\u201310, 2021.","DOI":"10.1101\/2021.10.20.465035"},{"key":"2026102406550634000_2026.09.11.750704v2.13","unstructured":"Andrei A Rusu , Neil C Rabinowitz , Guillaume Desjardins , Hubert Soyer , James Kirkpatrick , Koray Kavukcuoglu , Razvan Pascanu , and Raia Hadsell . Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016."},{"key":"2026102406550634000_2026.09.11.750704v2.14","doi-asserted-by":"crossref","unstructured":"Arun Mallya and Svetlana Lazebnik . Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765\u20137773, 2018.","DOI":"10.1109\/CVPR.2018.00810"},{"key":"2026102406550634000_2026.09.11.750704v2.15","unstructured":"Jaehong Yoon , Eunho Yang , Jeongtae Lee , and Sung Ju Hwang . Lifelong learning with dynamically expandable networks. In International Conference on Learning Representations, 2018."},{"key":"2026102406550634000_2026.09.11.750704v2.16","first-page":"14387","article-title":"Organizing recurrent network dynamics by task-computation to enable continual learning","volume":"33","year":"2020","journal-title":"Advances in neural information processing systems"},{"key":"2026102406550634000_2026.09.11.750704v2.17","unstructured":"Matthew J. Tilley , Michelle Miller , and David J. Freedman . Artificial neuronal ensembles with learned context dependent gating. In International Conference on Learning Representations, 2023."},{"issue":"1","key":"2026102406550634000_2026.09.11.750704v2.18","doi-asserted-by":"crossref","first-page":"8275","DOI":"10.1038\/s41467-024-52289-3","article-title":"Rapid context inference in a thalamocortical model using recurrent neural networks","volume":"15","year":"2024","journal-title":"Nature Communications"},{"key":"2026102406550634000_2026.09.11.750704v2.19","doi-asserted-by":"publisher","DOI":"10.1038\/s41593-024-01668-6"},{"key":"2026102406550634000_2026.09.11.750704v2.20","first-page":"90826","article-title":"Separating the \u2018what\u2019 and \u2018how\u2019 of compositional computation to enable reuse and continual learning","volume":"38","year":"2025","journal-title":"Advances in Neural Information Processing Systems"},{"key":"2026102406550634000_2026.09.11.750704v2.21","doi-asserted-by":"publisher","DOI":"10.1038\/s41593-018-0310-2"},{"key":"2026102406550634000_2026.09.11.750704v2.22","first-page":"1954","article-title":"Structured flexibility in recurrent neural networks via neuromodulation","volume":"37","year":"2024","journal-title":"Advances in Neural Information Processing Systems"},{"key":"2026102406550634000_2026.09.11.750704v2.23","doi-asserted-by":"publisher","DOI":"10.1137\/0110015"},{"key":"2026102406550634000_2026.09.11.750704v2.24","first-page":"34578","volume-title":"In Proceedings of the 42nd International Conference on Machine Learning","volume":"267","year":"2025"}],"container-title":[],"original-title":[],"link":[{"URL":"https:\/\/syndication.highwire.org\/content\/doi\/10.64898\/2026.09.11.750704","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,10,24]],"date-time":"2026-10-24T13:55:13Z","timestamp":1792850113000},"score":1,"resource":{"primary":{"URL":"http:\/\/biorxiv.org\/lookup\/doi\/10.64898\/2026.09.11.750704"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,9,14]]},"references-count":24,"URL":"https:\/\/doi.org\/10.64898\/2026.09.11.750704","relation":{},"subject":[],"published":{"date-parts":[[2026,9,14]]},"subtype":"preprint"}}