{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,5,15]],"date-time":"2026-05-15T06:11:53Z","timestamp":1778825513024,"version":"3.51.4"},"reference-count":40,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,5,15]],"date-time":"2026-05-15T00:00:00Z","timestamp":1778803200000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"DOI":"10.13039\/501100020884","name":"Agencia Nacional de Investigaci\u00f3n y Desarrollo","doi-asserted-by":"publisher","id":[{"id":"10.13039\/501100020884","id-type":"DOI","asserted-by":"publisher"}]}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Comput. Neurosci."],"abstract":"<jats:sec>\n                    <jats:title>Introduction<\/jats:title>\n                    <jats:p>Understanding how deep learning models map neural population activity to stimuli requires both high predictive accuracy and interpretable internal mechanisms.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Methods<\/jats:title>\n                    <jats:p>In this work, we employ the POYO framework, a scalable transformer architecture based on spike tokenization and latent modeling, to decode large-scale retinal ganglion cell recordings. We ask whether the model's attention mechanisms can provide biologically meaningful insight by evaluating two contrasting conditions: a uniform flash stimulus and a spatiotemporally structured moving ball stimulus.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>We show that the model decodes both stimuli reliably and adapts rapidly to new preparations via fine-tuning, suggesting the capture of transferable population codes. We then analyze the model's internal organization, revealing that encoder attention patterns adapt to stimulus complexity: attention heads appear synchronized and broadly distributed for the flash stimulus, whereas they exhibit heterogeneous, specialized allocation strategies for the moving ball. By aggregating attention weights to identify the most relevant neurons for each task, we demonstrate that these high-attention units possess distinct physiological signatures\u2014concentrating sustained, high-firing rates responses for the flash vs. diverse kinetics for the structured input. We confirm the causal validity of these findings via attention-guided ablations, where the progressive removal of these top-ranked units yields systematic losses in decoding performance. Furthermore, we expand the analysis to the decoder's attention, uncovering stimulus-specific retrieval strategies where individual heads exhibit distinct directional tuning preferences.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Discussion<\/jats:title>\n                    <jats:p>We conclude that generic attention mechanisms can spontaneously recover biological coding strategies, identifying functionally distinct neural subpopulations without supervision, thus validating the utility of transformer-based architectures for neuroscientific discovery.<\/jats:p>\n                  <\/jats:sec>","DOI":"10.3389\/fncom.2026.1769478","type":"journal-article","created":{"date-parts":[[2026,5,15]],"date-time":"2026-05-15T05:42:54Z","timestamp":1778823774000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":0,"title":["Attention maps reveal stimulus-dependent retinal population codes"],"prefix":"10.3389","volume":"20","author":[{"given":"Francisco","family":"Miqueles","sequence":"first","affiliation":[{"name":"Departamento de Electr\u00f3nica, Universidad T\u00e9cnico Federico Santa Mar\u00eda","place":["Valpara\u00edso, Chile"]}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Adri\u00e1n G.","family":"Palacios","sequence":"additional","affiliation":[{"name":"Centro Interdisciplinario de Neurociencia de Valpara\u00edso, Facultad de Ciencias, Universidad de Valpara\u00edso","place":["Valpara\u00edso, Chile"]}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"John","family":"Atkinson","sequence":"additional","affiliation":[{"name":"Universidad Andr\u00e9s Bello","place":["Las Condes, Chile"]}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Mar\u00eda-Jos\u00e9","family":"Escobar","sequence":"additional","affiliation":[{"name":"Departamento de Electr\u00f3nica, Universidad T\u00e9cnico Federico Santa Mar\u00eda","place":["Valpara\u00edso, Chile"]}],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"1965","published-online":{"date-parts":[[2026,5,15]]},"reference":[{"key":"B1","doi-asserted-by":"publisher","first-page":"358","DOI":"10.1038\/nrn1888","article-title":"Neural correlations, population coding and computation","volume":"7","author":"Averbeck","year":"2006","journal-title":"Nat. Rev. Neurosci"},{"key":"B2","article-title":"\u201cMulti-session, multi-task neural decoding from distinct cell-types and brain regions,\u201d","author":"Azabou","year":"2025"},{"key":"B3","article-title":"\u201cA unified, scalable framework for neural population decoding,\u201d","author":"Azabou","year":"2023"},{"key":"B4","doi-asserted-by":"publisher","first-page":"345","DOI":"10.1038\/nature16468","article-title":"The functional specialization of retinal ganglion cells in the mouse","volume":"529","author":"Baden","year":"2016","journal-title":"Nature"},{"key":"B5","doi-asserted-by":"publisher","first-page":"345","DOI":"10.1038\/nature16468","article-title":"The functional diversity of retinal ganglion cells in the mouse","volume":"529","author":"Baden","year":"2016","journal-title":"Nature"},{"key":"B6","unstructured":"Bengio\n              Y.\n            \n            \n              Deleu\n              T.\n            \n            \n              Rahaman\n              N.\n            \n            \n              Ke\n              N. R.\n            \n            \n              Lachapelle\n              S.\n            \n            \n              Bilaniuk\n              O.\n            \n          \n          \n          2019"},{"key":"B7","doi-asserted-by":"publisher","DOI":"10.1371\/journal.pcbi.1006057","article-title":"Nonlinear decoding of a complex movie from the mammalian retina","author":"Botella-Soler","year":"2018","journal-title":"PLoS Comput. Biol"},{"key":"B8","doi-asserted-by":"publisher","first-page":"199","DOI":"10.1080\/713663221","volume":"12","author":"Chichilnisky","year":"2001"},{"key":"B9","doi-asserted-by":"publisher","first-page":"533","DOI":"10.1038\/323533a0","article-title":"Learning representations by back-propagating errors","volume":"323","author":"David","year":"1986","journal-title":"Nature"},{"key":"B10","doi-asserted-by":"publisher","first-page":"1424","DOI":"10.1152\/jn.1997.78.4.2048","article-title":"Mosaic arrangement of ganglion cell receptive fields in rabbit retina","volume":"87","author":"de Vries","year":"2002","journal-title":"J. Neurophysiol"},{"key":"B11","doi-asserted-by":"publisher","first-page":"4377","DOI":"10.1113\/jphysiol.2008.156638","article-title":"Functional circuitry of visual adaptation in the retina","volume":"586","author":"Demb","year":"2008","journal-title":"Journal of Physiology"},{"key":"B12","article-title":"\u201cAn image is worth 16?-16 words: transformers for image recognition at scale,\u201d","author":"Dosovitskiy","year":"2021"},{"key":"B13","doi-asserted-by":"publisher","author":"Dowling","year":"2012","DOI":"10.2307\/j.ctv31zqj2d"},{"key":"B14","doi-asserted-by":"publisher","first-page":"163","DOI":"10.1016\/j.conb.2021.10.014","article-title":"Dynamics on the manifold: identifying computational dynamical activity from neural population recordings","volume":"70","author":"Duncker","year":"2021","journal-title":"Curr. Opin. Neurobiol."},{"key":"B15","doi-asserted-by":"publisher","first-page":"444","DOI":"10.3389\/fncel.2018.00444","article-title":"Characterization of retinal functionality at different eccentricities in a diurnal rodent","volume":"12","author":"Escobar","year":"2018","journal-title":"Front. Cell. Neurosci"},{"key":"B16","doi-asserted-by":"publisher","first-page":"1331","DOI":"10.1152\/jn.00919.2014","article-title":"Visual coding with a population of direction-selective neurons in the retina","volume":"18","author":"Fiscella","year":"2015","journal-title":"Nat. Neurosci"},{"key":"B17","doi-asserted-by":"publisher","DOI":"10.1523\/ENEURO.0506-19.2020","article-title":"Machine learning for neural decoding","author":"Glaser","year":"2020","journal-title":"eNeuro"},{"key":"B18","doi-asserted-by":"publisher","first-page":"150","DOI":"10.1016\/j.neuron.2009.12.009","article-title":"Eye smarter than scientists believed: neural computations in circuits of the retina","volume":"65","author":"Gollisch","year":"2010","journal-title":"Neuron"},{"key":"B19","unstructured":"Limits of optimal decoding under synaptic coarse-tuning\n          \n          \n            \n              Hendler\n              O.\n            \n            \n              Segev\n              R.\n            \n            \n              Shamir\n              M.\n            \n          \n          arXiv preprint\n          \n          2026"},{"key":"B20","doi-asserted-by":"publisher","author":"Kietzmann","year":"2018","DOI":"10.1101\/133504"},{"key":"B21","doi-asserted-by":"publisher","first-page":"1523","DOI":"10.1016\/j.cub.2004.08.060","article-title":"Efficiency of information transmission by retinal ganglion cells","volume":"14","author":"Koch","year":"2004","journal-title":"Curr. Biol"},{"key":"B22","author":"Kolb","year":"1995"},{"key":"B23","doi-asserted-by":"publisher","DOI":"10.1371\/journal.pcbi.1004304","article-title":"High accuracy decoding of dynamical motion from a large retinal population","author":"Marre","year":"2015","journal-title":"PLoS Comput. Biol"},{"key":"B24","doi-asserted-by":"publisher","first-page":"1207","DOI":"10.1126\/science.270.5239.1207","article-title":"Concerted signaling by retinal ganglion cells","volume":"270","author":"Meister","year":"1995","journal-title":"Science"},{"key":"B25","doi-asserted-by":"publisher","first-page":"14958","DOI":"10.1523\/JNEUROSCI.1004-13.2013","article-title":"Interacting linear and nonlinear characteristics produce population coding asymmetries between on and off cells in the retina","volume":"33","author":"Nichols","year":"2013","journal-title":"J. Neurosci"},{"key":"B26","doi-asserted-by":"publisher","first-page":"698","DOI":"10.1038\/35079612","article-title":"Retinal ganglion cells act largely as independent encoders","volume":"411","author":"Nirenberg","year":"2001","journal-title":"Nature"},{"key":"B27","doi-asserted-by":"publisher","first-page":"732","DOI":"10.1101\/196949","article-title":"Neural data science: accelerating the experiment-analysis-theory cycle in large-scale neuroscience","volume":"99","author":"Paninski","year":"2018","journal-title":"Neuron"},{"key":"B28","doi-asserted-by":"publisher","DOI":"10.1101\/153759","article-title":"\u201cNeural networks for efficient bayesian decoding of natural images from retinal neurons,\u201d","author":"Parthasarathy","year":"2017"},{"key":"B29","doi-asserted-by":"publisher","first-page":"995","DOI":"10.1038\/nature07140","article-title":"Spatio-temporal correlations and visual signalling in a complete neuronal population","volume":"454","author":"Pillow","year":"2008","journal-title":"Nature"},{"key":"B30","doi-asserted-by":"publisher","first-page":"456","DOI":"10.1038\/s41598-018-36861-8","article-title":"Speed-selectivity in retinal ganglion cells is sharpened by broad spatial frequency, naturalistic stimuli","volume":"9","author":"Ravello","year":"2019","journal-title":"Sci. Rep"},{"key":"B31","author":"Rieke","year":"1999"},{"key":"B32","doi-asserted-by":"publisher","first-page":"754","DOI":"10.1038\/s41586-023-06031-6","article-title":"Learnable latent embeddings for joint behavioural and neural analysis","volume":"617","author":"Schneider","year":"2023","journal-title":"Nature"},{"key":"B33","doi-asserted-by":"publisher","first-page":"215","DOI":"10.1016\/j.tics.2015.02.005","article-title":"Visual attention mitigates information loss in small- and large-scale neural codes","volume":"19","author":"Sprague","year":"2015","journal-title":"Trends Cogn. Sci"},{"key":"B34","doi-asserted-by":"publisher","first-page":"194","DOI":"10.1038\/nrn3165","article-title":"Direction selectivity in the retina: symmetry and asymmetry in structure and function","volume":"19","author":"Vaney","year":"2012","journal-title":"Nat. Rev. Neurosci"},{"key":"B35","article-title":"\u201cAttention is all you need,\u201d","author":"Vaswani","year":"2017","journal-title":"Advances in Neural Information Processing Systems (NeurIPS)"},{"key":"B36","doi-asserted-by":"publisher","first-page":"2336","DOI":"10.1152\/jn.1997.78.5.2336","article-title":"Decoding visual information from a population of retinal ganglion cells","volume":"78","author":"Warland","year":"1997","journal-title":"J. Neurophysiol"},{"key":"B37","doi-asserted-by":"publisher","first-page":"10365","DOI":"10.1101\/2021.01.16.426955","volume":"34","author":"Ye","year":"2021"},{"key":"B38","doi-asserted-by":"publisher","author":"Yger","year":"2018","DOI":"10.7554\/eLife.34518"},{"key":"B39","doi-asserted-by":"publisher","first-page":"614","DOI":"10.1152\/jn.90941.2008","article-title":"Gaussian-process factor analysis for low-dimensional learned latent dynamics of neural population activity","volume":"102","author":"Yu","year":"2009","journal-title":"J. Neurophysiol"},{"key":"B40","doi-asserted-by":"publisher","first-page":"3396","DOI":"10.1109\/TNNLS.2024.3351120","article-title":"Robust decoding of rich dynamical visual scenes with retinal spikes","volume":"36","author":"Yu","year":"2025","journal-title":"IEEE Trans. Neural Netw. Learn. Syst"}],"container-title":["Frontiers in Computational Neuroscience"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fncom.2026.1769478\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,5,15]],"date-time":"2026-05-15T05:42:56Z","timestamp":1778823776000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fncom.2026.1769478\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,5,15]]},"references-count":40,"alternative-id":["10.3389\/fncom.2026.1769478"],"URL":"https:\/\/doi.org\/10.3389\/fncom.2026.1769478","relation":{},"ISSN":["1662-5188"],"issn-type":[{"value":"1662-5188","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,5,15]]},"article-number":"1769478"}}