Spatial predictive coding in visual cortical neurons

Spatial predictive coding in visual cortical neurons | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Spatial predictive coding in visual cortical neurons View ORCID Profile Qingqing Zhang , View ORCID Profile Sverre Grødem , View ORCID Profile Alexa Gracias , View ORCID Profile Kristian K. Lensjø , View ORCID Profile Marianne Fyhn , View ORCID Profile Carsen Stringer , View ORCID Profile Marius Pachitariu doi: https://doi.org/10.1101/2025.09.17.676794 Qingqing Zhang 1 HHMI Janelia Research Campus , Ashburn, VA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Qingqing Zhang For correspondence: zhangq{at}hhmi.org pachitarium{at}hhmi.org Sverre Grødem 2 Center for Integrative Neuroplasticity, Department of Bioscience, University of Oslo , Oslo, Norway Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Sverre Grødem Alexa Gracias 1 HHMI Janelia Research Campus , Ashburn, VA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Alexa Gracias Kristian K. Lensjø 2 Center for Integrative Neuroplasticity, Department of Bioscience, University of Oslo , Oslo, Norway Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Kristian K. Lensjø Marianne Fyhn 2 Center for Integrative Neuroplasticity, Department of Bioscience, University of Oslo , Oslo, Norway Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Marianne Fyhn Carsen Stringer 1 HHMI Janelia Research Campus , Ashburn, VA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Carsen Stringer Marius Pachitariu 1 HHMI Janelia Research Campus , Ashburn, VA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Marius Pachitariu For correspondence: zhangq{at}hhmi.org pachitarium{at}hhmi.org Abstract Full Text Info/History Metrics Preview PDF Abstract Predictive coding is a theoretical framework that can explain how animals build internal models of their sensory environments by predicting sensory inputs. Predictive coding may capture either spatial or temporal relationships between sensory objects. While the original theory by Rao and Ballard, 1999 described spatial predictive coding, much of the recent experimental data has been interpreted as evidence for temporal predictive coding. Here we directly tested whether the “mismatch” neural responses in sensory cortex are due to a spatial or a temporal internal model. We adopted two common paradigms to study predictive coding: one based on virtual-reality and one based on static images. After training mice with repeated visual stimulation for several days, we performed multiple manipulations, including: 1) we introduced a novel stimulus, 2) we replaced a stimulus with a novel gray wall, 3) we duplicated a trained stimulus, or 4) we altered the order of the stimuli. The first two manipulations induced a substantial mismatch response in neural populations of up to 20,000 neurons recorded across primary and higher-order visual cortex, while the third and fourth ones did not. Thus, a mismatch response only occurred if a new spatial – not temporal – pattern was introduced. Introduction Predictive coding has emerged as a popular all-encompassing neural theory which postulates that neural representations in sensory brain areas are optimized to predict sensory inputs [ 1 – 3 ]. To do this prediction well, neuronal networks must contain an intrinsic model of the natural world which can be used to detect unpredicted changes in the environment. Predictive coding theories can be divided into temporal and spatial predictive coding. Temporal predictive coding is the most prevalent form, and often implied implicitly [ 4 – 6 ]. It postulates that brains are optimized to predict future sensory inputs from past sensory inputs and/or from self-movements. Spatial predictive coding, while less popular as a general theory of brain function, was the basis of the original predictive coding framework [ 7 , 8 ]. It postulates that brains predict localized sensory inputs from their larger spatial context, with no requirement for temporal information. Predictive coding is a learning theory with substantial computational capabilities. Some of the most highly-capable artificial systems are trained using cost functions that are similar predictive coding [ 9 ]. These artificial intelligence (AI) methods vary in how they treat the temporal and spatial domains. For example, large language models are trained to predict the next word in a sequence, a temporal prediction [ 10 ]. In contrast, large vision models are often trained to predict left-out patches in images from surrounding patches, a spatial prediction [ 11 , 12 ]. Vision models for video analysis are pretrained using a combination of spatial and temporal prediction [ 13 , 14 ]. While the cost functions for temporal and spatial predictive coding share similar principles, their underlying mechanistic implementations and intended functions differ substantially. Thus, it is important to distinguish between different types of predictive coding, in order to understand the inner workings of a system like the mammalian sensory cortex. The main evidence for predictive coding in the brain has come from observations of the “mismatch response” or sensory “prediction error” in sensory cortex [ 5 , 6 , 15 ]. These are large neural responses that are observed under sensory manipulations that deviate from the expected, or typical sensory inputs. In most cases, the expectations are established spatially, temporally, and/or by the novelty of a stimulus [ 16 ]. Such large neural responses are observed in both humans and animals, using a variety of recording methods from single-unit recordings [ 4 , 17 – 22 ], to LFP [ 23 – 26 ], fMRI [ 27 – 31 ], and EEG/MEG [ 32 – 36 ]. Focusing on single-unit recordings in animals, the mismatch response has been observed in response to sensory manipulations after a specific stimulus sequence had been shown repeatedly over many trials and across several consecutive days [ 6 , 15 ]. In some studies, the stimulus sequence was coupled to motor drive, for example in closed-loop virtual reality where mice control the forward progression through a corridor [ 37 , 38 ]. The motor coupling is sometimes thought to be critical, as it could enable animals to predict the consequences of their own actions [ 39 – 42 ]. Here we perform a critical evaluation of the predictions of spatial and temporal predictive coding, using two closed-loop sensory paradigms designed to be similar to previous studies. We recorded large-scale neural populations from primary and higher-order visual cortices. Using various sensory manipulations with appropriate controls, we find that the mismatch response in primary and higher-order visual areas represents a spatial, rather than a temporal prediction error. Results In our first set of experiments, mice ran for five days in a virtual reality environment in single sessions of up to 1.5 hours ( Figure 1A ). The mice ran forward through linear corridors at an average speed of 19.4cm/s, experiencing in each trial a succession of four visual landmarks. This was similar to [ 6 ], except we used crops from natural images rather than gratings for increased saliency and ecological relevance. We denote individual stimuli by capitalized letters, followed by a subscripted number indicating the position of the landmark in the trial. Thus, standard trials are denoted A 1 B 2 C 3 D 4 , while manipulated test trials during test sessions are A 1 B 2 X 3 D 4 , with X = N, D etc. In-between landmarks, the corridors had a fixed pattern of sparse noise squares which never changed. All individual visual landmarks (i.e. C 3 ) were composed of a pair of repeated sub-images (i.e. cc). The role of these images was to allow the introduction of half-landmarks (i.e. c) or repeated landmarks (i.e. ccc) without introducing new spatial patterns locally (see next section). Download figure Open in new tab Figure 1: Mismatch neural responses to novel sensory inputs. (A) Experimental setup illustrating the progression through the VR corridor (top), the 270 degree virtual reality setup (bottom left) and the progression of training and testing sessions over days. (B) Field of view of the neural recordings with zoom-in. (C) Trial-averaged neural responses after training. Neurons are sorted by preferred spatial position and responses are shown on held-out trials. (D) (left) Same as (C) for trials where a new stimulus N 3 was introduced. Preferred spatial positions were computed on A 1 B 2 C 3 D 4 trials excluding the C 3 positions; (right) Population-averaged responses (n=7 mice). (E) Average responses from (D) in the shaded blue area. (F) (left) Average running speeds for A 1 B 2 C 3 D 4 and A 1 B 2 N 3 D 4 trials, and (right) quantification in the blue shaded area. (G-I) Same as (D-F) for A 1 B 2 D 3 D 4 trials (n=5 mice). (J-L) Same as (D-F) for open-loop A 1 B 2 C 3 D 4 trials where the treadmill was blocked (n=3 mice). All data are mean ± s.e.m. across mice. Statistical significance in all figures was assessed using paired two-sided t-tests across mice with *, **, and *** representing p < 0.05, p < 0.01, and p < 0.001 respectively. During test sessions after training, we recorded neural activity simultaneously from populations of 5,846 - 22,711 neurons from primary visual cortex as well as higher-order visual areas using a two-photon mesoscope ( Figure 1B ). Neural activity was deconvolved and interpolated to the position of the corridor, then averaged over trials. Neurons were sorted according to their position tuning in the A 1 B 2 C 3 D 4 corridor, and their averaged activity was computed on held-out trials, that were not used to calculate the peak tuning position ( Figure 1C ). Finally, neurons were assigned into groups based on the location of their peak firing (A, B, C, D cells in Figure 1C ). Note that on some subsequent analyses, A, B, D cells were defined without including the C 3 corridor portion, because that portion was subject to manipulations and quantifications (see Methods). We also chose to focus on B and D cells exclusively, since these have preferred responses close to the manipulated C 3 region and would be more likely to show effects. Responses to novel spatial and temporal patterns We first tested whether a novel spatial pattern N3 would drive a mismatch response, defined as a substantially larger neural response compared to the trained stimuli ( Figure 1D ). Both B and D cells responded substantially more to the novel stimulus N 3 compared to the old pattern C 3 , nearly matching their firing rates to their own preferred stimuli (B 2 and D 4 respectively, Figure 1DE ). This increased response was accompanied by a reduction in running speed in nearly all mice, as well as modest pupil dilation ( Figure 1F , Figure S1A ). The increased response was not present in anesthesia ( Figure S2 ). The novel spatial pattern may thus have driven a mismatch response, either due to a failed prediction of stimulus C 3 (temporal predictive coding), or due to the novelty of the spatial pattern in stimulus N 3 (spatial predictive coding). Download figure Open in new tab Figure S1: Pupil fluctuations. (A) (left) Average pupil size over trials. (right) Average pupil size in the new stimulus time window compared between C 3 and N 3 trials (n=7 mice). (B-F) Same as (A) for test sessions where different stimuli are introduced: (B): D 3 , D stimulus at position 3 (n=5 mice), (C): G 3 , gray screen at position 3 (n=6 mice), (D): c 3 , half of C stimulus at position 3 (n=6 mice), (E): ccc 3 , repeated half of C stimulus at position 3 (n=7 mice), (F): 8xc 3 , 8 times repeated half of C stimulus at position 3 (n=7 mice). All data are mean ± s.e.m. across mice. Paired two-sided t-tests were performed across mice. Download figure Open in new tab Figure S2: Responses to novel stimuli in anesthesia. (A) Sorted neural responses to the trained stimulus (C 3 ) and the novel stimulus (N 3 ). (B) Average responses of B and D cells on C 3 and N 3 trials (n=4 mice). (C) Average responses during the C 3 and N 3 stimulus presentations compared between B and D cells. All data are mean ± s.e.m. across mice. Paired two-sided t-tests were performed across mice. To disentangle these two hypotheses, we next tested whether replacing C 3 with a trained pattern D 3 also results in a mismatch response. If temporal predictive coding applies, we would still expect a failed prediction of stimulus C 3 . If mismatch responses instead result from spatial predictive coding, they should not occur here, because the spatial patterns in D 3 also occur in D 4 , a trained stimulus. We observed no mismatch response on A 1 B 2 D 3 D 4 trials, with neural responses to D 3 being indistinguishable from those to D 4 ( Figure 1GH ). We also did not observe changes in mouse behaviors such as running speed or pupil fluctuations ( Figure 1I , Figure S1B ). Since mice did not indicate behaviorally that a “failed prediction” occurred for the D 3 stimulus, it is possible that the mismatch responses to N 3 are in fact directly due to behavioral modulations, such as decreased running speeds, which did occur in A 1 B 2 N 3 D 4 trials ( Figure 1F ). To test this hypothesis, we performed test sessions with N 3 trials in which we clamped the treadmill on all trials, thereby preventing the mice from running, while automatically moving the VR forward in open-loop. Similar to unclamped, closed-loop trials, we observed mismatch responses to N 3 in the B and D cells that were larger than the responses to C 3 and nearly as large as the responses to their preferred stimuli (B and D, respectively, Figure 1JK ). Thus, mismatch responses were observed even without running ( Figure 1L ). In our next test, we shortened the stimulus C 3 to just its first half “c” ( Figure 1A ). No new spatial patterns were introduced in this way, since the spatial transitions from sparse noise to “c” and from “c” to sparse noise were already contained in the trained stimulus containing C 3 =“cc” ( Figure 1A ). However, a new temporal pattern was introduced this way, which could lead to a failed prediction of a second “c”, should the temporal theory hold. We did not find a mismatch response to the missing second “c” ( Figure 2AB ), again arguing against the temporal predictive coding hypothesis. We also did not observe changes in running and pupil fluctuations ( Figure 2C , Figure S1C ). Download figure Open in new tab Figure 2: Additional manipulations and omissions for test stimuli. (A) Same as Figure 1D-F for a shortened stimulus manipulation (c 3 = half of C 3 , n=6 mice). (D-F) Same as Figure 1D-F for trials where C 3 is omitted and replaced with gray screen G 3 (n=6 mice). (G-I) Same as (D-F), but before training in naive animals (n=6 mice). All data are mean ± s.e.m. across mice. Paired two-sided t-tests were performed across mice. Understanding omission responses In the next test we adopted a paradigm from previous studies [ 6 ], by omitting a trained stimulus C 3 from the sequence and replacing it with a gray wall G 3 ( Figure 2D ). Similar to previous studies, this omission resulted in elevated neural responses, which have previously been interpreted as a temporal mismatch error ( Figure 2DE ). In our data, these omission responses were substantially smaller than the responses to the novel stimulus N 3 (compare to Figure 1DE ). We also observed a behavioral response, similar to the N 3 trials ( Figure 2F ). The elevated neural responses and behavioral changes to the gray wall were not observed in naive mice ( Figure 2G-I ). In previous studies, such findings were interpreted to suggest that omission responses are due to temporal prediction errors [ 6 ]. However, notice that the gray wall also introduces novel spatial patterns. Before addressing omission responses further, we introduce another set of experiments, designed to be similar to other paradigms where temporal predictive coding was invoked, and where omission mismatch responses were similarly observed [ 5 ]. In these experiments, we presented sequences of static images A 1 B 2 C 3 G 4 where G is a gray screen ( Figure 3A ). To maintain the closed-loop nature of the experiment, which is thought to be required for temporal predictive coding [ 40 ], we linked the image transitions to the running speed of the mice (averaging 29.8 cm/s), requiring mice to run for a set period before transitioning to the next stimulus. We used different durations for each stimulus to encourage temporal structure learning (2s, 1s, 3s and 1s respectively for A 1 , B 2 , C 3 and G 4 ). This created an abstract linear “corridor” in each trial, that we used to interpolate neural responses to, similar to the VR corridor. To focus on the fast neural responses to stimulus transitions, we recorded at faster frame rates (30Hz) and focused on individual recording locations in the primary visual cortex ( Figure 3A ). After training, neural responses were primarily concentrated at stimulus transitions, including at gray screen onset ( Figure 3B ). Download figure Open in new tab Figure 3: Mismatch responses with self-initiated static stimuli. (A) Experiment setup, similar to the VR experiments, but the forward running initiates transitions between static images, and the recorded field of view is smaller and focused on V1. Transitions between different stimulus pairs require different durations of constant forward running as shown in left panels. (B) Trial-averaged neural responses after training. Neurons were sorted by their peak response times and held-out trials are shown. (C) (left) Same as (B) for trials where a novel stimulus N 2 was introduced. Cell sorting was done without the B 2 stimulus. (right) Population-average for A cells and C cells (n=4 mice). (D) Quantification of baselined average responses inside the blue window in (C). (E) (left) Average running speeds for A 1 B 2 C 3 D 4 and A 1 B 2 N 3 D 4 trials, and (right) quantification in the blue shaded area. (F-H) Same as (C-E) for trials where B 2 is omitted and replaced with a gray screen G 2 (n=5 mice). All data are mean ± s.e.m. across mice. Paired two-sided t-tests were performed across mice. Similar to the VR experiments, we found that a novel stimulus N 2 introduces elevated firing rates in multiple neural populations ( Figure 3CD ), although without a significant behavioral change ( Figure 3E ). Also similar to the VR experiments, “omission” trials where we replaced B 2 with a gray screen G 2 also resulted in elevated firing rates, potentially indicating mismatch errors ( Figure 3FG ). These were also not accompanied by clear behavioral changes ( Figure 3H ). Thus, omission trial responses in both the VR and static image experiments appear to suggest temporal prediction errors. To illustrate why this is not necessarily the case, consider the bottom-up receptive field (RF) responses of visual cortical neurons ( Figure 4A ). The virtual corridors drive spatial RFs at every position, including at the boundary between the sparse noise and the gray screen (omitted C 3 ). In this case, G 3 introduces a new spatial pattern at this boundary, which would drive different sets of neurons based on their spatial RFs. Similarly, at image transitions in the static image experiment, spatiotemporal RFs (stRF) need to take into consideration the transition between one spatial pattern and another, since neurons as early as the retina respond to step-changes in illumination in an on/off fashion rather than to sustained illumination patterns [ 43 ]. A new stRF pattern is introduced at the transition between A 1 and G 2 , that was not present in the trained stimuli. Thus, mismatch responses on omission trials can be explained by novel bottom-up inputs or novel spatial patterns ( Figure 4A ). Download figure Open in new tab Figure 4: Novel spatial but not temporal patterns drive mismatch responses. (A) (left) Illustration of bottom-up receptive field mismatch in “omission” trials, where an adapted stimulus is replaced with a gray screen. Note that for static images, receptive fields at image transition need to take into account the relative changes from the previous stimulus to the next, similar to how transient retinal cells respond to illumination changes. (middle) Temporal and spatial predictive coding theories suggest different causal relations in the prediction. (right) New experiment design to disentangle the predictions of spatial and temporal predictive coding on static image experiments. (B) (left) Trial-averaged neural responses to A 1 B 2 A 3 N 4 G 5 trials after training with A 1 B 2 A 3 C 4 G 5 sequences. Neurons were sorted by their peak response times and held-out trials are shown. (right) Population-average for A cells and B cells (n=6 mice). (C) Quantification of baselined average responses inside the blue window in (B). (D) (left) Average running speeds for A 1 B 2 A 3 C 4 G 5 and A 1 B 2 A 3 N 4 G 5 trials, and (right) quantification in the blue shaded area. (E-G) Similar to (B-D) when the C 4 stimulus is replaced with a B 4 stimulus (n=6 mice). (F) Comparing responses to (left) B 2 vs B 4 in A 1 B 2 A 3 B 4 G 5 trials or (right) to G 5 in A 1 B 2 A 3 C 4 G 5 vs A 1 B 2 A 3 B 4 G 5 trials. All data are mean ± s.e.m. across mice. Paired two-sided t-tests were performed across mice. To further distinguish between the temporal and spatial predictive coding, we designed a new experiment, in which we trained mice on the sequence A 1 B 2 A 3 C 4 G 5 ( Figure 4A ). After training, we introduced a test sequence where we replaced C 4 with B 4 , resulting in A 1 B 2 A 3 B 4 G 5 trials. Temporal predictive coding would postulate that a prediction error would occur in this case, because the expected transition A 3 → C 4 did not happen. The bottom-up or spatial version of predictive coding would not postulate a prediction error, because the transition that is introduced in this way A 3 → B 4 was already present in the training stimuli as A 1 → B 2 . Furthermore, the spatial theory predicts a mismatch error for the transition B 4 → G 5 , since this was not present in the training stimuli ( Figure 4A ). We first verified that under this protocol a mismatch response can indeed still be generated by the introduction of a novel stimulus N 4 replacing B 4 ( Figure 4B-D ). We then focused on trials where C 4 was replaced with B 4 ( Figure 4E ). In these trials, responses to B 4 were not different from the responses to B 2 , indicating the absence of a prediction error ( Figure 4E-G ). However, responses to G 5 on B 4 trials were greater than on the trained C 4 trials, in accordance with the prediction of the spatial theory. Thus, elevated neural responses in the static image experiments are also likely due to spatial rather than temporal predictive errors. Extended temporal changes drive increased neural activity Finally, we present a set of experiments that suggest a role for temporal mismatch modulation under more exaggerated manipulations. These experiments were performed using the VR paradigm, and took advantage of the definition of C 3 as two consecutive “c” images. We first extended the stimulus by another “c”, thus replacing C 3 with ccC 3 ( Figure 5A ). Similar to the experiments with a single “c”, no new spatial patterns were introduced in this way. However, there was slightly elevated neural firing in the extended neural response to the third “c” in the sequence, higher than the response to the second “c”, but similar to the response to the first “c” ( Figure 5AB ), with no detectable behavioral change ( Figure 5C ). Next, we repeated the “c” stimulus eight times in a sequence ( Figure 5D ). For both C and D cells, we observed ramping of neural activity throughout this extended stimulus ( Figure 5DE ). The response to the last “c” in the sequence was substantially elevated compared to the second “c” ( Figure 5E ), with no behavioral change ( Figure 5F ). This ramping response was somewhat surprising, because one may predict adaptation rather than facilitation when the same stimulus is repeated many times. This facilitation was not present in naive animals ( Figure 5G-I ), ruling out potential accumulation of the jGCaMP8s sensor. We interpret this facilitation as a gain modulation due to a gradual increase in an internal state of alertness or “surprise”. This change in state likely originates elsewhere in the brain, but may lead to differential sensory processing, similar to how other internal states modulate sensory responses [ 44 ]. Download figure Open in new tab Figure 5: Responses to extended stimuli. (A-C) Same as Figure 1D-F for a test stimulus manipulation where the C = cc stimulus was extended with another copy of c (n=7 mice). Note that this does not create new receptive field patterns (of the kind in Figure 4A). (D-F) Same as (A-C) for 8 consecutive copies of c (n=7 mice). (G-I) Same as (D-F) before training (n=6 mice). All data are mean ± s.e.m. across mice. Paired two-sided t-tests were performed across mice. Discussion Our results show that continued exposure to a visual stimulus sequence does not result in a temporal mismatch response when that sequence is changed, whether in a closed-loop virtual reality paradigm, or in a closed-loop static stimulus paradigm. Rather, the elevated neural responses signal novel spatial patterns and spatial transitions, previously not viewed during training. Our observations provide evidence to the original formulation of predictive coding as a spatial prediction framework [ 7 ]. Other studies have argued against temporal predictive coding in the primary visual cortex of mice, monkeys and humans using oddball paradigms [ 45 , 46 ]. However, these studies either did not include training over days [ 45 ], or did not demonstrate a behavioral or neural effect of training over days [ 46 ], leaving open the possibility that the animals did not pay attention to the stimuli and thus could not form an internal model of the sequence. In contrast, we used paradigms that are similar to those in which mismatch responses were previously reported [ 5 , 6 , 15 ]. We demonstrated clear evidence of mismatch to novel spatial patterns, accompanied by behavioral responses during mismatch, and showed that these effects were not present in naive mice. We also used closed-loop experiments, which have previously been suggested to be important for predictive coding [ 39 – 42 ], and we used salient, ecologically-relevant natural images that were not repeated in an oddball paradigm. Briefly, we distinguish here between studies like ours where a stimulus sequence is habituated over many sessions, thus becoming “learned”, and those where sensorimotor coupling itself is being tested with non-”learned” stimuli, such as in virtual reality environments in closed/open loop. We note that under those paradigms too, it is unclear if temporal predictive processing needs to be invoked to explain mismatch responses, because the neural sensory response properties may be sufficient to explain elevated responses [ 47 ]. Together with our study, this suggests that temporal prediction is not a fundamental aspect of sensory processing, either in primary or in higher-order sensory areas. The spatial prediction framework, however, may be an important mechanism by which the brain forms internal models of the visual world. Practical machine learning methods, such as masked autoencoders, demonstrate that a spatial prediction criterion can indeed be leveraged to form such internal models [ 12 ]. The biological implementation of such models remains highly unspecified. The mismatch response may be “bottom-up”, for example due to adaptation to familiar images in the feedforward thalamocortical projection, though this is somewhat argued against by the lack of mismatch responses in anesthesia ( Figure S2 ). It could also be “top-down”, such as through a high-level “memory bank” which detects the novelty of a stimulus, and modulates the processing of sensory information in sensory areas through feedback mechanisms [ 48 ]. Finally, lateral processing through horizontal connections [ 49 ] may implement a version of the masked autoencoder approach, where representations of local image patches are predicted through lateral “attention” mechanisms similar to those in the transformer architecture. This top-down or lateral feedback could be implemented via loops with higher-order thalamic nuclei [ 15 ], dendritic modulation [ 50 , 51 ], and/or specific inhibitory neural targeting [ 22 , 52 , 53 ]. While many possibilities remain open, the precise mechanisms of the spatial pattern learning algorithm remain to be uncovered. Author contributions QZ and MP designed the study. AG performed the surgeries. QZ performed the experiments. QZ performed data analysis with advice from MP and CS. SG, KKL and MF provided gcamp8 viruses. QZ, CS, and MP wrote the manuscript with input from all authors. Data availability The data will be made available upon publication in a journal. Code availability Analysis code will be made available upon publication in a journal. Methods All experimental procedures were conducted according to IACUC, and received ethical approval from the IACUC board at HHMI Janelia Research Campus. Experimental methods Animals Eight mice were used in the virtual reality (VR) experiments (4 females, 4 males), ranging in age from 5 to 15 months. Of these, five mice were bred to express GCaMP8s (TetO-jGCaMP8s X CamK2a-tTa, JAX 037717, JAX 007004) and three expressed GCaMP6s (TetO-GCaMP6s × CamK2a-tTa, JAX 024742, JAX 007004) [ 54 ]. Nine mice were used in the static image experiments (5 females, 4 males), ranging in age from 4 to 15 months. Of these, five mice were bred to express GCaMP8s (TetO-jGCaMP8s X CamK2a-tTa, JAX 037717, JAX 007004); three transgenic mice (VGAT-CRE x Ai14, JAX 016962, JAX 007914) and one C 57 Bl/6Crl received virus injections to express soma-targeted RiboL1-jGCaMP8s tethered to the ribosome [ 55 ]. Mice were housed in reverse light cycle (12 h light on, 12 h light off). Surgical procedures Mice were anesthetized briefly using 3% isoflurane until breathing slowed to approximately 1-2 breaths per second. The animal was transferred to the nosecone on the stereotax. The isoflurane was maintained between 1.5-2% to ensure the animal was within a surgical plane of anesthesia and breathing was approximately 1-2 breaths per second. Marcaine (2.5 mg/kg) was injected subcutaneously beneath the incision area, warmed LRS fluids (+ 5% dextrose) and buprenorphine 0.1 mg/kg (systemic analgesic) were administered subcutaneously along with dexamethasone 7 mg/kg via intramuscular route. Eye lubrication was applied. An incision was made. Vetbond was applied to the skin around the edges of the incision area. Measurements were taken to determine bregma-lambda distance. The center of the right V1 window was measured from lambda: anterior +0.9 mm, lateral +2.6 mm. The edges of the 4 mm craniotomy were measured from this center point, 2 mm in each direction, and the coordinates were lightly etched in the skull using a dental drill. A thin layer of Optibond Solo Plus was applied to the entire surface of the skull and UV cured for 40 s. The headbar was placed over the marked craniotomy location while following the curvature of the skull, then secured using Calibra Universal Resin Cement. The craniotomy was drilled slowly. The skull was also cooled with chilled cortex buffer every few minutes. Once the edges were sufficiently thin, cold sterile-filtered cortex buffer was placed over the well of the headbar and allowed to sit for several minutes. Once the edges softened, the entire bone piece was carefully lifted away using Dumont forceps. A 4+5 mm double window was placed into the craniotomy so that the 4 mm window replaced the previously removed bone piece and the 5 mm window lay over the edge of the bone. Vetbond was applied around the edges of the window. The window was held in place for 5 minutes until the Vetbond fully dried. Then, Calibra Universal Resin Cement was applied over the edges of the 5 mm window and covering the inner sides of the headbar and UV cured. The mouse was removed from the stereotax. Carprofen 5 mg/kg was administered subcutaneously and the animal was allowed to recover on heat. The post-operative care consisted of 2 additional days of monitoring for pain or distress, warm LRS fluids (+5% dextrose), and Carprofen (5 mg/kg) delivered subcutaneously. For the mice that received viral injections, in addition to the procedures described above, we delivered a 1:1 mixture of Thy1s:TTA (AAV9, 1.64 × 1013 V.G./ml) and TRE3G:RiboI1-jGCaMP8s (AAV9, 2.45 × 1013 V.G./ml). A total volume of 150 nL was injected across three sites at a depth of –0.3 mm to ensure even expression throughout the cranial window. The injection coordinates relative to lambda were: (1) AP +1.9 mm, ML +2.6 mm; (2) AP 0 mm, ML +1.6 mm; and (3) AP 0 mm, ML +3.6 mm. Imaging acquisition A 2p-RAM mesoscope [ 56 ] was used for two-photon imaging. We did online Z correction while recording with ScanImage [ 57 ]. Recordings for VR experiments were performed at 3.17 Hz and for static image experiments they were performed at 29-30 Hz. Three mice used in the static image experiments were recorded with a Bergamo II microscope at 30 Hz. Data from these animals were combined with those from mice recorded with the mesoscope for subsequent analyses. During awake recordings, mice were free to run on a styrofoam cylinder. Mice were acclimatized to running on the cylinder in the dark for several sessions before stimulus presentation and imaging. During anesthetized recordings, mice were briefly anesthetized with 3% isoflurane until respiration decreased to 1–2 breaths/s; isoflurane was subsequently maintained at 1% to stabilize breathing at 1 breath/s. A warming pad was used to maintain body temperature. Visual stimuli The same visual stimulation system was used as in [ 58 ]. Three LED tablet screens (aspect ratio 4:3) surrounded the head-fixed mouse covering 270 degrees in azimuth. The mouse was placed in the center of the screen, at a slightly lower elevation from the vertical center. VR presentation The VR environment was rendered with OpenGL in MATLAB. The stimuli were presented with Matlab Psychtoolbox [ 59 ]. Natural images A, B, C, D formed the core scene. The images were concatenated, interleaved with random-dot gray walls, and placed on the walls of a linear corridor ( Figure 1A ). The complete concatenated image contained 2,700 horizontal and 150 vertical pixels. We used pixels as a unit of distance travelled in VR. For closed-loop VR, the VR progressed with the locomotion of the animal: if the mouse speed exceeded a threshold (1.5 cm/s), the corridor advanced at a constant speed (180 pixels/s). For open-loop VR, the VR progressed at a constant speed of 180 pixels/s independent of the mouse running. For the standard A 1 B 2 C 3 D 4 sequence, it took 15s to finish one trial if the mouse continuously exceeded the running speed threshold. Static image presentation Static images were presented with Matlab Psychtoolbox [ 59 ]. Natural images A, B, C, D formed the core scene. In the A 1 B 2 C 3 G 4 paradigm, stimuli stayed on the screen for 2, 1, 3, 1 seconds, respectively. In the A 1 B 2 A 3 C 4 G 5 paradigm, stimuli stayed on the screen for 1, 2, 1, 2, 3 seconds, respectively. The progression of stimulus timing was counted only if the mouse was running. Stationary periods were defined as epochs in which the running speed remained below 1 mm/s for at least five consecutive monitor frames (running at 60 Hz). Therefore, if the mouse was stationary, the same stimulus would stay on the screen. Behavioral training Across both the VR and static image paradigms, mice underwent a 1-week training period. During the first 5–7 days they experienced only the standard sequences. Beginning thereafter, we introduced manipulations on a subset of trials. Before training, we collected a baseline recording that included the standard sequence and all manipulations used later, presented in randomized order. During the post-training test period, we typically ran 1–2 manipulation types per day. Manipulation trials accounted for 7–12% of the total number of trials. For the VR experiments, we performed a different manipulation on each day, in the following orders over days: batch 1 (3 mice): A 1 B 2 D3D 4 , A 1 B 2 G 3 D 4 , A 1 B 2 C 3 D 4 , A 1 B 2 N 3 D 4 , A 1 B 2 ccC 3 D 4 , A 1 B 2 8xC 3 D 4 batch 2 (4 mice): A 1 B 2 D3D 4 , A 1 B 2 G 3 D 4 , A 1 B 2 N 3 D 4 , A 1 B 2 C 3 D 4 , A 1 B 2 ccC 3 D 4 , A 1 B 2 8xC 3 D 4 Two different sets of natural images (A, B, C etc.) were used in those two VR batches. The mice ran 100-300 trials per day with a session duration of 1-1.5 h. If the mouse ran less than 100 trials, the session was excluded from analysis. Some mice stopped running and thus did not complete all manipulation sessions. For the static image experiments, we did manipulations over days in the following order: A 1 B 2 C 3 G 4 (5 mice): A 1 G 2 C 3 G 4 , A 1 N 2 C 3 G 4 A 1 B 2 A 3 C 4 G 5 (6 mice): A 1 B 2 A 3 B 4 G 5 , A 1 B 2 A 3 N 4 G 5 Two different sets of natural images were used in these two static-image paradigms. The mice ran 100-400 trials per day with a session duration of 1-1.5 h. Data analysis For analysis we used Python 3 [ 60 ], using numpy and scipy [ 61 , 62 ]. The figures were made using matplotlib and jupyter-notebook [ 63 , 64 ]. Processing of calcium imaging data Calcium raw data were processed with Suite2p [ 65 ]. We used Suite2p for motion correction, ROI and neuropil extraction. The neuropil signal was subtracted from the ROI signal with a coefficient of 1. All ROIs with an SNR greater than 0.3 were included for subsequent analysis. The SNR definition was: Spike deconvolution was carried out using a new algorithm implemented in Suite2p [ 65 ], which employs a neural network trained to predict ground-truth spikes from calcium traces [ 66 ]. VR analysis We defined trial types based on the manipulation type and location of the manipulation, resulting in the following naming conventions: C 3 trials: A 1 B 2 C 3 D 4 sequence N 3 trials: A 1 B 2 N 3 D 4 sequence D3 trials: A 1 B 2 D3D 4 sequence G 3 trials: A 1 B 2 G 3 D 4 sequence C 3 trials: A 1 B 2 C 3 D 4 sequence 8xC 3 trials: A 1 B 2 8xC 3 D 4 sequence Train/test split and cell ordering Normal C 3 trials were split into training and test sets with 50% of trials in each set. Using the training set, we aligned and sorted cells by the location of their peak response along the linear corridor. This fixed ordering was then applied to the test set ( Figure 1C ). We defined cell classes by the stimulus window containing their peak. Cells with peaks during the A, B, C, or D image windows were classified as A, B, C, or D cells, respectively, whereas cells with peaks during the random-dot gray periods were classified as G cells. Baseline and response windows Unless otherwise specified, responses for comparisons between standard and manipulated trials were taken from the manipulation window and baselined by subtracting the mean response in a window [−150, −50] pixels preceding the manipulation window. Responses were quantified on the test trials as defined above. In the second VR batch, We used two novel stimuli in the novelty session and analyzed the data by combining trials from both (N 3 trials). To compare B and D cell responses between C 3 and N 3 trials, we determined the cell classes using training C 3 trials and assigned cells to the image where their peak firing occurred, excluding the C 3 image itself. This ensured that neither C 3 nor N 3 influenced the cell class assignment, thus allowing for an unbiased comparison between responses to these two stimuli. Responses were computed as the mean activity in the 300 pixel window following C 3 /N 3 onset (approx. 1,667 ms), after baseline subtraction in the window [−150, −50] before C 3 /N 3 onset ( Figure 1E ); running speed and pupil size were evaluated over the same window ( Figure 1F , Figure S1A ). Similarly for comparing between D 3 and D 4 responses, we sorted the D 3 trials based on the sorting order from C 3 train trials without including the C 3 image responses. Neuronal response within the D 4 window (300 pixels after stimulus onset, approx. 1,667 ms) from C 3 trials was compared with that within D 3 window from D 3 trials after baseline subtraction ( Figure 1H ); running speed and pupil size were measured in the corresponding windows ( Figure 1I , Figure S1B ). For G 3 analyses, we again used the C 3 -based ordering without C 3 responses. We quantified the boundary response when the animal entered the C 3 image versus G 3 pure gray screen [-150, 50] around stimulus transition after baseline subtraction ( Figure 2EH ); running speed and pupil size were compared over the same window ( Figure 2FI , Figure S1D ). When analyses included C cells (e.g., C 3 , ccC 3 , 8×C 3 ), sorting was derived from C 3 train trials with C 3 responses considered. For the shortened stimulus C 3 condition, the offset response to the second c in C 3 and the only c in C 3 (150 pixels after the corresponding stimulus offset, approx. 833 ms) were compared after subtracting the baseline ([ − 150, −50] pixels before the first c onset) response ( Figure 2B ). For the extended stimulus ccC 3 condition, the response to the second c and some offset in C 3 and the third c and some offset in ccC 3 (200 pixels after the corresponding stimulus onset) were compared after subtracting the baseline ([−150, −50] pixels before the first c onset) response ( Figure 5B ). For the extended stimulus 8xC 3 condition, the response to the second c in C 3 and the last c in 8xC 3 (−150 pixels after the corresponding stimulus onset) were compared after subtracting the baseline ([−150, 50] pixels before the first c onset) response ( Figure 5EH ). Running speed and pupil size were evaluated in the corresponding windows. Static image analysis We defined trial types based on the manipulation type and location of the manipulation, resulting in the following naming conventions: B 2 trials: A 1 B 2 C 3 G 4 sequence N 2 trials: A 1 N 2 C 3 G 4 sequence G 2 trials: A 1 G 2 C 3 G 4 sequence C 4 trials: A 1 B 2 A 3 C 4 G 5 sequence B 4 trials: A 1 B 2 A 3 B 4 G 5 sequence N 4 trials: A 1 B 2 A 3 N 4 G 5 sequence Cells were sorted in the same way as in the VR experiment. The normal trials (either A 1 B 2 C 3 G 4 or A 1 B 2 A 3 C 4 G 5 sequence) were split into training and test data. We aligned cells based on their peak response time in the static image sequence within the training data. This sorting order was applied to the test data ( Figure 3B ). We defined cells by their peak response position in the A 1 stimulus window as the A cells (and similarly for B, C, G cells). We used two novel stimuli in the novelty session and analyzed the data by combining trials from both in the N 2 and N 4 trial analyses. For the A 1 B 2 C 3 G 4 sequence, to compare the responses of A, C cells in B 2 and N 2 (or G 2 ) trials, in the training set with the B 2 trials, we aligned cells without using B 2 image responses. Neuronal response within the B 2 /N 2 (or G 2 ) window (0-300 ms after stimulus onset) was averaged and baselined using the final 100 ms of the gray screen period ( Figure 3DG ). When comparing the running speed in B 2 and N 2 (or G 2 ) trials, traces were extracted from 0-400 ms after stimulus onset and baselined by the final 200 ms of the gray screen period. Similarly, for the A 1 B 2 A 3 C 4 G 5 sequence, to compare the responses of A, B cells in C 4 and N 4 trials, responses were extracted from C 4 and N 4 window (0-300 ms after stimulus onset) and baselined with the final 100 ms of the gray screen period ( Figure 4C ). For the A 1 B 2 A 3 B 4 G 5 sequence, when comparing responses in B 2 and B 4 , responses were extracted from B 2 and B 4 window (0-300 ms after stimulus onset) and baselined using the final 100 ms of the gray screen period of B 4 trials ( Figure 4F , left); when comparing responses in G 5 from C 4 and B 4 trials, responses were extracted from G 5 window (0-300 ms after the onset of gray screen) and baselined (the final 100 ms of the gray screen period) of C 4 and B 4 trials, respectively ( Figure 4F , right). The running speed during B 2 and B 4 stimuli in B 4 trials was defined as the average running speed 0-400 ms after stimulus onset, and was baselined by the final 200 ms of the gray screen period ( Figure 4G ). Statistical tests All the statistics were paired two-sided t-tests across mice. Detailed P values are reported below. 1E. B cells: 2.68 × 10 − 4 / 1.41 × 10 − 3 / 9.17 × 10 − 3 / 1.76 × 10 − 3; D cells: 4.83 × 10 − 4 / 4.96 × 10 − 3 / 5.90 × 10 − 3 / 0.0731 1F. 0.0348 1H. B cells: 0.370 / 0.0815 / 0.478 / 0.599; D cells: 0.348 / 0.165 / 0.599 / 0.920 1I. 0.249 1K. B cells: 2.54 × 10 − 3 / 0.0990 / 0.107 / 0.116; D cells: 0.0419 / 0.0440 / 0.499/ 0.0723 2B. C cells: 0.133 / 0.184 / 0.827 / 0.586; D cells: 0.0588 / 0.121 / 0.861 / 0.0742 2C. 0.199 2E. B cells: 0.685 / 0.0543 / 0.0574 / 0.0365; D cells: 5.30 × 10 − 3 / 0.0191 / 0.0162 / 0.0713; G cells: 3.27 × 10 − 3 / 9.49 × 10 − 3 / 0.0891 / 0.105 2F. 0.0179 2H. B cells: 5.08 × 10 − 3 / 0.538 / 0.168 / 0.188; D cells: 0.0281 / 0.431 / 0.872 / 0.182; G cells: 0.0284 / 2.58 × 10 − 3 / 0.0649 / 0.510 2I. 0.721 3D. A cells: 0.0250; C cells: 8.88 × 10 − 3 3E. 0.497 3G. A cells: 7.79 × 10 − 3; C cells: 4.94 × 10 − 3 3H. 0.312 4C. A cells: 1.23 × 10 − 3; B cells: 5.42 × 10 − 5 4D. 0.078 4F. (left) A cells: 0.174; B cells: 0.214; (right) A cells: 0.0245; B cells: 3.42 × 10 − 3 4G. 0.297 5B. C cells: 3.40 × 10 − 3 / 8.26 × 10 − 3 / 0.623 / 0.0948; D cells: 0.0489 / 0.0986 / 0.618 / 0.0335 5C. 0.425 5E. C cells: 5.80 × 10 − 4 / 0.0126 / 0.838 / 3.00 × 10 − 3; D cells: 1.82 × 10 − 4 / 0.0437 / 0.934 / 0.548 5F. 0.649 5H. C cells: 0.282 / 0.181 / 0.0199 / 0.257; D cells: 0.146 / 0.0109 / 0.0965 / 0.852 5I. 0.864 1S1C. B cells: 0.255 / 0.245 / 0.0690 / 0.760; D cells: 0.637 / 0.203 / 0.839 / 0.193 1S2. 0.0237 / 0.730 / 0.235 / 4.50 × 10 − 3 / 0.268 / 0.174 Retinotopy We used the same procedure as in [ 58 ] to estimate spatial receptive locations. A set of 500 natural images, each presented three times, was used for mapping. First, responses from a reference mouse were fitted with 200 spatial kernels. Recordings from all other mice were then fitted using the same model to determine the preferred kernel and corresponding spatial location of each cell. These spatial locations were aligned to the spatial map of the reference mouse, in which the recorded region was segmented according to the computed sign map [ 67 ]. Acknowledgments This research was funded by the Howard Hughes Medical Institute at the Janelia Research Campus. From the Vivarium, we thank Jim Cox, Crystall Lopez, Anne Kuzspit, Miriam Rose, Michalis Michaelos, Gillian Harris, Sarah Lindo, Rachel Gattoni and their respective teams for animal breeding, husbandry, surgeries, and behavioral training support. From JeT, we thank Daniel Flickinger, Vasily Goncharov, Alex Sohn, Tobias Goulet, and Steven Sawtelle for help with rig maintenance and upgrades. From MBF Bioscience we thank Georg Jaindl, Mitchell Sandoe, and Boris Djiguemde for scanimage support. We thank Nelson Spruston, Sandro Romani, and Vivek Jayaraman for helpful discussions about the work. Funder Information Declared Howard Hughes Medical Institute, https://ror.org/006w34k90 References [1]. ↵ Georg B Keller and Thomas D Mrsic-Flogel . Predictive Processing: A Canonical Cortical Computation . Neuron , 100 ( 2 ): 424 – 435 , October 2018 . OpenUrl CrossRef PubMed [2]. Karl Friston . The free-energy principle: A unified brain theory? Nat. Rev. Neurosci ., 11 : 127 – 138 , 2010 . OpenUrl CrossRef PubMed Web of Science [3]. ↵ Rajesh PN Rao . A sensory–motor theory of the neocortex . Nature Neuroscience , 27 ( 7 ): 1221 – 1235 , July 2024 . OpenUrl CrossRef PubMed [4]. ↵ Travis Meyer and Carl R. Olson . Statistical learning of visual transitions in monkey inferotemporal cortex . Proceedings of the National Academy of Sciences , 108 ( 48 ): 19401 – 19406 , November 2011 . OpenUrl Abstract / FREE Full Text [5]. ↵ Jeffrey P Gavornik and Mark F Bear . Learned spatiotemporal sequence recognition and prediction in primary visual cortex . Nature Neuroscience , 17 ( 5 ): 732 – 737 , May 2014 . OpenUrl CrossRef PubMed [6]. ↵ Aris Fiser , David Mahringer , Hassana K Oyibo , Anders V Petersen , Marcus Leinweber , and Georg B Keller . Experience-dependent spatial expectations in mouse visual cortex . Nature neuroscience , 19 ( 12 ): 1658 – 1664 , 2016 . OpenUrl CrossRef PubMed [7]. ↵ Rajesh PN Rao and Dana H Ballard . Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptivefield effects . Nature neuroscience , 2 ( 1 ): 79 – 87 , 1999 . OpenUrl CrossRef PubMed Web of Science [8]. ↵ Victor AF Lamme and Pieter R. Roelfsema . The distinct modes of vision offered by feedforward and recurrent processing . Trends in Neurosciences , 23 ( 11 ): 571 – 579 , November 2000 . OpenUrl CrossRef PubMed Web of Science [9]. ↵ Josh Achiam , Steven Adler , Sandhini Agarwal , Lama Ahmad , Ilge Akkaya , Florencia Leoni Aleman , Diogo Almeida , Janko Altenschmidt , Sam Altman , Shyamal Anadkat , et al. Gpt-4 technical report . arXiv preprint arxiv: 2303.08774 , 2023 . [10]. ↵ Alec Radford , Karthik Narasimhan , Tim Salimans , and Ilya Sutskever . Improving Language Understanding by Generative Pre-Training . 2019 . [11]. ↵ Jinghao Zhou , Chen Wei , Huiyu Wang , Wei Shen , Cihang Xie , Alan Yuille , and Tao Kong . iBot: Image BERT pre-training with online tokenizer . arXiv preprint arxiv: 2111.07832 , 2021 . [12]. ↵ Kaiming He , Xinlei Chen , Saining Xie , Yanghao Li , Piotr Dollár , and Ross Girshick . Masked autoencoders are scalable vision learners . arXiv preprint arxiv: 2111.06377 , 2021 . [13]. ↵ Zhan Tong , Yibing Song , Jue Wang , and Limin Wang . VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training . Advances in Neural Information Processing Systems , 35 : 10078 – 10093 , December 2022 . OpenUrl [14]. ↵ Christoph Feichtenhofer , Haoqi Fan , Yanghao Li , and Kaiming He . Masked Autoencoders As Spatiotemporal Learners . Advances in Neural Information Processing Systems , 35 : 35946 – 35958 , December 2022 . OpenUrl [15]. ↵ Shohei Furutachi , Alexis D. Franklin , Andreea M. Aldea , Thomas D. Mrsic-Flogel , and Sonja B. Hofer . Cooperative thalamocortical circuit mechanism for sensory prediction errors . Nature , 633 ( 8029 ): 398 – 406 , September 2024 . OpenUrl CrossRef PubMed [16]. ↵ Charan Ranganath and Gregor Rainer . Neural mechanisms for detecting and remembering novel events . Nature Reviews Neuroscience , 4 ( 3 ): 193 – 202 , March 2003 . OpenUrl CrossRef PubMed Web of Science [17]. ↵ Lin Li , Earl K Miller , and Robert Desimone . The representation of stimulus familiarity in anterior inferior temporal cortex . Journal of Neurophysiology , 69 ( 6 ): 1918 – 1929 , June 1993 . OpenUrl CrossRef PubMed Web of Science [18]. Nachum Ulanovsky , Liora Las , and Israel Nelken . Processing of low-probability sounds by cortical neurons . Nature Neuroscience , 6 ( 4 ): 391 – 398 , April 2003 . OpenUrl CrossRef PubMed Web of Science [19]. Shengjin Xu , Wanchen Jiang , Mu-ming Poo , and Yang Dan . Activity recall in a visual cortical ensemble . Nature Neuroscience , 15 ( 3 ): 449 – 455 , March 2012 . OpenUrl CrossRef PubMed [20]. Jan Homann , Sue Ann Koay , Kevin S. Chen , David W. Tank , and Michael J. Berry . Novel stimuli evoke excess activity in the mouse primary visual cortex . Proceedings of the National Academy of Sciences , 119 ( 5 ): e2108882119 , February 2022 . OpenUrl Abstract / FREE Full Text [21]. Matthew F Tang , Ehsan Kheradpezhouh , Conrad CY Lee , J Edwin Dickinson , Jason B Mattingley , and Ehsan Arabzadeh . Expectation violations enhance neuronal encoding of sensory information in mouse primary visual cortex . Nature Communications , 14 ( 1 ): 1196 , March 2023 . OpenUrl PubMed [22]. ↵ Marina Garrett , Peter Groblewski , Alex Piet , Doug Ollerenshaw , Farzaneh Najafi , Iryna Yavorska , Adam Amster , Corbett Bennett , Michael Buice , Shiella Caldejon , Linzy Casal , Florence D’Orazi , Scott Daniel , Saskia EJ de Vries , Daniel Kapner , Justin Kiggins , Jerome Lecoq , Peter Ledochowitsch , Sahar Manavi , Nicholas Mei , Christopher B. Morrison , Sarah Naylor , Natalia Orlova , Jed Perkins , Nick Ponvert , Clark Roll , Sam Seid , Derric Williams , Allison Williford , Ruweida Ahmed , Daniel Amine , Yazan Billeh , Chris Bowman , Nicholas Cain , Andrew Cho , Tim Dawe , Max Departee , Marie Desoto , David Feng , Sam Gale , Emily Gelfand , Nile Gradis , Conor Grasso , Nicole Hancock , Brian Hu , Ross Hytnen , Xiaoxuan Jia , Tye Johnson , India Kato , Sara Kivikas , Leonard Kuan , Quinn L’Heureux , Sophie Lambert , Arielle Leon , Elizabeth Liang , Fuhui Long , Kyla Mace , Ildefons Magrans de Abril , Chris Mochizuki , Chelsea Nayan , Katherine North , Lydia Ng , Gabriel Koch Ocker , Michael Oliver , Paul Rhoads , Kara Ronellenfitch , Kathryn Schelonka , Josh Sevigny , David Sullivan , Ben Sutton , Jackie Swapp , Thuyanh K. Nguyen , Xana Waughman , Joshua Wilkes , Michael Wang , Colin Farrell , Wayne Wakeman , Hongkui Zeng , John Phillips , Stefan Mihalas , Anton Arkhipov , Christof Koch , and Shawn R. Olsen . Stimulus novelty uncovers coding diversity in survey of visual cortex . bioRxiv , page 2023.02.14.528085, January 2025 . [23]. ↵ Francois D Szymanski , Jose A Garcia-Lazaro , and Jan WH Schnupp . Current Source Density Profiles of Stimulus-Specific Adaptation in Rat Auditory Cortex . Journal of Neurophysiology , 102 ( 3 ): 1483 – 1490 , September 2009 . OpenUrl CrossRef PubMed Web of Science [24]. Wolfger von der Behrens , Peter Bäuerle , Manfred Kössl , and Bernhard H. Gaese . Correlating Stimulus-Specific Adaptation of Cortical Neurons and Local Field Potentials in the Awake Rat . Journal of Neuroscience , 29 ( 44 ): 13837 – 13849 , November 2009 . OpenUrl Abstract / FREE Full Text [25]. André M Bastos , Mikael Lundqvist , Ayan S Waite , Nancy Kopell , and Earl K Miller . Layer and rhythm specificity for predictive routing . Proc. Natl. Acad. Sci. U. S. A ., 117 : 31459 – 31469 , 2020 . OpenUrl Abstract / FREE Full Text [26]. ↵ Georgia Bastos , Jacob T Holmes , Jordan M Ross , Anna M Rader , Connor G Gallimore , Joseph A Wargo , Darcy S Peterka , and Jordan P Hamm . Top-down input modulates visual context processing through an interneuron-specific circuit . Cell Rep ., 42 : 113133 , 2023 . OpenUrl PubMed [27]. ↵ Zoe Kourtzi and Nancy Kanwisher . Cortical Regions Involved in Perceiving Object Shape . Journal of Neuroscience , 20 ( 9 ): 3310 – 3318 , May 2000 . OpenUrl Abstract / FREE Full Text [28]. Tobias Egner , Jim M. Monti , and Christopher Summerfield . Expectation and Surprise Determine Neural Population Responses in the Ventral Visual Stream . Journal of Neuroscience , 30 ( 49 ): 16601 – 16608 , December 2010 . OpenUrl Abstract / FREE Full Text [29]. Arjen Alink , Caspar M. Schwiedrzik , Axel Kohler , Wolf Singer , and Lars Muckli . Stimulus Predictability Reduces Responses in Primary Visual Cortex . Journal of Neuroscience , 30 ( 8 ): 2960 – 2966 , February 2010 . OpenUrl Abstract / FREE Full Text [30]. Peter Kok , Janneke F. M. Jehee , and Floris P. de Lange . Less Is More: Expectation Sharpens Representations in the Primary Visual Cortex . Neuron , 75 ( 2 ): 265 – 270 , July 2012 . OpenUrl CrossRef PubMed Web of Science [31]. ↵ Helen Blank , Marlene Spangenberg , and Matthew H. Davis . Neural Prediction Errors Distinguish Perception and Misperception of Speech . The Journal of Neuroscience , 38 ( 27 ): 6076 – 6089 , July 2018 . OpenUrl Abstract / FREE Full Text [32]. ↵ Hannu Tiitinen , Patrick May , K Reinikainen , and R Näätänen . Attentive novelty detection in humans is governed by pre-attentive sensory memory . Nature , 372 ( 6501 ): 90 – 92 , November 1994 . OpenUrl CrossRef PubMed Web of Science [33]. Iiro P. Jääskeläinen , Jyrki Ahveninen , Giorgio Bonmassar , Anders M. Dale , Risto J. Ilmoniemi , Sari Levänen , Fa-Hsuan Lin , Patrick May , Jennifer Melcher , Steven Stufflebeam , Hannu Tiitinen , and John W. Belliveau . Human posterior auditory cortex gates novel sounds to consciousness . Proceedings of the National Academy of Sciences , 101 ( 17 ): 6809 – 6814 , April 2004 . OpenUrl Abstract / FREE Full Text [34]. Corinna Haenschel , David J. Vernon , Prabuddh Dwivedi , John H. Gruzelier , and Torsten Baldeweg . Event-Related Brain Potential Correlates of Human Auditory Sensory Memory-Trace Formation . Journal of Neuroscience , 25 ( 45 ): 10494 – 10501 , November 2005 . OpenUrl Abstract / FREE Full Text [35]. Marta I. Garrido , James M. Kilner , Stefan J. Kiebel , Klaas E. Stephan , Torsten Baldeweg , and Karl J. Friston . Repetition suppression and plasticity in the human brain . NeuroImage , 48 ( 1 ): 269 – 279 , October 2009 . OpenUrl CrossRef PubMed Web of Science [36]. ↵ Catherine Wacongne , Etienne Labyt , Virginie Van Wassenhove , Tristan Bekinschtein , Lionel Naccache , and Stanislas Dehaene . Evidence for a hierarchy of predictions and prediction errors in human cortex . Proc. Natl. Acad. Sci ., 108 : 20754 – 20759 , 2011 . OpenUrl Abstract / FREE Full Text [37]. ↵ Georg B Keller , Tobias Bonhoeffer , and Mark Hübener . Sensorimotor mismatch signals in primary visual cortex of the behaving mouse . Neuron , 74 : 809 – 815 , 2012 . OpenUrl CrossRef PubMed Web of Science [38]. ↵ Pawel Zmarz and Georg B. Keller . Mismatch Receptive Fields in Mouse Visual Cortex . Neuron , 92 ( 4 ): 766 – 772 , November 2016 . OpenUrl CrossRef PubMed [39]. ↵ Alexander Attinger , Bo Wang , and Georg B. Keller . Visuomotor Coupling Shapes the Functional Development of Mouse Visual Cortex . Cell , 169 ( 7 ): 1291 – 1302 .e14, June 2017 . OpenUrl CrossRef PubMed [40]. ↵ Marcus Leinweber , Daniel R. Ward , Jan M. Sobczak , Alexander Attinger , and Georg B. Keller . A Sensorimotor Circuit in Mouse Cortex for Visual Flow Predictions . Neuron , 95 ( 6 ): 1420 – 1432 .e5, September 2017 . OpenUrl CrossRef PubMed [41]. Rebecca Jordan and Georg B. Keller . Opposing Influence of Top-down and Bottom-up Input on Excitatory Layer 2/3 Neurons in Mouse Primary Visual Cortex . Neuron , 108 ( 6 ): 1194 – 1206 .e5, December 2020 . OpenUrl CrossRef PubMed [42]. ↵ Nicholas J Audette , WenXi Zhou , Alessandro La Chioma , and David M Schneider . Precise movement-based predictions in the mouse auditory cortex . Current Biology , 32 ( 22 ): 4925 – 4940.e6, November 2022 . OpenUrl CrossRef PubMed [43]. ↵ Stephen W Kuffler . Discharge patterns and functional organization of mammalian retina . Journal of neurophysiology , 16 ( 1 ): 37 – 68 , 1953 . OpenUrl CrossRef PubMed Web of Science [44]. ↵ Cristopher M Niell and Michael P Stryker . Modulation of visual responses by behavioral state in mouse visual cortex . Neuron , 65 ( 4 ): 472 – 479 , 2010 . OpenUrl CrossRef PubMed Web of Science [45]. ↵ Selina S Solomon , Huizhen Tang , Elyse Sussman , and Adam Kohn . Limited evidence for sensory prediction error responses in visual cortex of macaques and humans . Cereb. Cortex , 31 : 3136 – 3152 , 2021 . OpenUrl CrossRef PubMed [46]. ↵ Jacob A Westerberg , Yihan S Xiong , Eli Sennesh , Hamed Nejat , David Ricci , Séverine Durand , Ben Hardcastle , Hannah Cabasco , Hannah Belski , Ahad Bawany , Ryan Gillis , Henry Loeffler , Carter R Peene , Warren Han , Katrina Nguyen , Vivian Ha , Tye Johnson , Conor Grasso , Ahrial Young , Jackie Swapp , Ben Ouellette , Shiella Caldejon , Ali Williford , Peter A Groblewski , Shawn R Olsen , Carly Kiselycznyk , Christof Koch , Jerome A Lecoq , Alexander Maier , and André M Bastos . Sensory responses of visual cortical neurons are not prediction errors . bioRxiv , page 2024.10.02.616378, July 2025 . [47]. ↵ Tomaso Muzzu and Aman B. Saleem . Feature selectivity can explain mismatch signals in mouse visual cortex . Cell Reports , 37 ( 1 ), October 2021 . [48]. ↵ Catherine Wacongne , Jean-Pierre Changeux , and Stanislas Dehaene . A Neuronal Model of Predictive Coding Accounting for the Mismatch Negativity . J. Neurosci ., 32 : 3665 – 3678 , 2012 . OpenUrl Abstract / FREE Full Text [49]. ↵ Charles D Gilbert , Aniruddha Das , Minami Ito , Mitesh Kapadia , and Gerald Westheimer . Spatial integration and cortical dynamics . Proceedings of the National Academy of Sciences , 93 ( 2 ): 615 – 622 , 1996 . OpenUrl Abstract / FREE Full Text [50]. ↵ Fabian A. Mikulasch , Lucas Rudelt , Michael Wibral , and Viola Priesemann . Where is the error? Hierarchical predictive coding through dendritic error computation . Trends in Neurosciences , 46 ( 1 ): 45 – 59 , January 2023 . OpenUrl CrossRef PubMed [51]. ↵ Colleen J Gillon , Jason E Pina , Jérôme A. Lecoq , Ruweida Ahmed , Yazan N Billeh , Shiella Caldejon , Peter Groblewski , Timothy M Henley , India Kato , Eric Lee , Jennifer Luviano , Kyla Mace , Chelsea Nayan , Thuyanh V Nguyen , Kat North , Jed Perkins , Sam Seid , Matthew T. Valley , Ali Williford , Yoshua Bengio , Timothy P Lillicrap , Blake A Richards , and Joel Zylberberg . Responses to Pattern-Violating Visual Stimuli Evolve Differently Over Days in Somata and Distal Apical Dendrites . Journal of Neuroscience , 44 ( 5 ), January 2024 . [52]. ↵ Ryan G Natan , John J Briguglio , Laetitia Mwilambwe-Tshilobo , Sara I Jones , Mark Aizenberg , Ethan M Goldberg , and Maria Neimark Geffen . Complementary control of sensory adaptation by two types of cortical interneurons . eLife , 4 : e09868 , October 2015 . OpenUrl CrossRef PubMed [53]. ↵ Farzaneh Najafi , Simone Russo , and Jérôme Lecoq . Unexpected events trigger task-independent signaling in VIP and excitatory neurons of mouse visual cortex . iScience , 28 ( 2 ), February 2025 . [54]. ↵ Joseph B Wekselblatt , Erik D Flister , Denise M Piscopo , and Cristopher M Niell . Large-scale imaging of cortical dynamics during sensory perception and behavior . Journal of neurophysiology , 115 ( 6 ): 2852 – 2866 , 2016 . OpenUrl CrossRef PubMed [55]. ↵ Sverre Grødem , Ingeborg Nymoen , Guro Helén Vatne , Frederik Sebastian Rogge , Valgerður Björnsdóttir , Kristian Kinden Lensjø , and Marianne Fyhn . An updated suite of viral vectors for in vivo calcium imaging using intracerebral and retro-orbital injections in male mice . Nature Communications , 14 ( 1 ): 608 , 2023 . OpenUrl PubMed [56]. ↵ Nicholas James Sofroniew , Daniel Flickinger , Jonathan King , and Karel Svoboda . A large field of view two-photon mesoscope with subcellular resolution for in vivo imaging . Elife , 5 : e14472 , 2016 . OpenUrl CrossRef PubMed [57]. ↵ Thomas A Pologruto , Bernardo L Sabatini , and Karel Svoboda . Scanimage: flexible software for operating laser scanning microscopes . Biomedical engineering online , 2 ( 1 ): 13 , 2003 . OpenUrl PubMed [58]. ↵ Lin Zhong , Scott Baptista , Rachel Gattoni , Jon Arnold , Daniel Flickinger , Carsen Stringer , and Marius Pachitariu . nsupervised pretraining in biological neural networks . Nature , pages 1 – 8 , 2025 . [59]. ↵ Mario Kleiner , David Brainard , Denis Pelli , Allen Ingling , Richard Murray , Christopher Broussard , et al. What’s new in psychtoolbox-3 . Perception , 36 ( 14 ): 1 , 2007 . OpenUrl CrossRef PubMed [60]. ↵ Guido Van Rossum and Fred L. Drake . Python 3 Reference Manual. CreateSpace, Scotts Valley, CA , 2009 . [61]. ↵ Stefan Van Der Walt , S Chris Colbert , and Gael Varoquaux . The numpy array: a structure for efficient numerical computation . Computing in Science & Engineering , 13 ( 2 ): 22 , 2011 . OpenUrl [62]. ↵ Eric Jones , Travis Oliphant , Pearu Peterson , et al. SciPy: Open source scientific tools for Python , 2001 . [63]. ↵ John D Hunter . Matplotlib: A 2d graphics environment . Computing in science & engineering , 9 ( 3 ): 90 , 2007 . OpenUrl CrossRef [64]. ↵ Thomas Kluyver , Benjamin Ragan-Kelley , Fernando Pérez , Brian E Granger , Matthias Bussonnier , Jonathan Frederic , Kyle Kelley , Jessica B Hamrick , Jason Grout , Sylvain Corlay , et al. Jupyter notebooks-a publishing format for reproducible computational workflows . In ELPUB , pages 87 – 90 , 2016 . [65]. ↵ Marius Pachitariu , Carsen Stringer , Sylvia Schröder , Mario Dipoppa , L Federico Rossi , Matteo Carandini , and Kenneth D Harris . Suite2p: beyond 10,000 neurons with standard two-photon microscopy . bioRxiv , 2016 . [66]. ↵ Marius Pachitariu , Lin Zhong , Alexa Gracias , Amanda Minisi , Crystall Lopez , and Carsen Stringer . A critical initialization for biological neural networks . bioRxiv , pages 2025 – 01 , 2025 . [67]. ↵ Jun Zhuang , Lydia Ng , Derric Williams , Matthew Valley , Yang Li , Marina Garrett , and Jack Waters . An extended retinotopic map of mouse cortex . elife , 6 : e18372 , 2017 . OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted September 19, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Spatial predictive coding in visual cortical neurons Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Spatial predictive coding in visual cortical neurons Qingqing Zhang , Sverre Grødem , Alexa Gracias , Kristian K. Lensjø , Marianne Fyhn , Carsen Stringer , Marius Pachitariu bioRxiv 2025.09.17.676794; doi: https://doi.org/10.1101/2025.09.17.676794 Share This Article: Copy Citation Tools Spatial predictive coding in visual cortical neurons Qingqing Zhang , Sverre Grødem , Alexa Gracias , Kristian K. Lensjø , Marianne Fyhn , Carsen Stringer , Marius Pachitariu bioRxiv 2025.09.17.676794; doi: https://doi.org/10.1101/2025.09.17.676794 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7642) Biochemistry (17708) Bioengineering (13904) Bioinformatics (41992) Biophysics (21466) Cancer Biology (18618) Cell Biology (25531) Clinical Trials (138) Developmental Biology (13387) Ecology (19924) Epidemiology (2067) Evolutionary Biology (24337) Genetics (15615) Genomics (22521) Immunology (17749) Microbiology (40424) Molecular Biology (17194) Neuroscience (88673) Paleontology (667) Pathology (2839) Pharmacology and Toxicology (4827) Physiology (7650) Plant Biology (15160) Scientific Communication and Education (2046) Synthetic Biology (4302) Systems Biology (9826) Zoology (2271)