Learning Enhances Sensory and Multiple Non-sensory Representations in Primary Visual Cortex

Abstract

We determined how learning modifies neural representations in primary visual cortex (V1) during acquisition of a visually guided behavioral task. We imaged the activity of the same layer 2/3 neuronal populations as mice learned to discriminate two visual patterns while running through a virtual corridor, where one pattern was rewarded. Improvements in behavioral performance were closely associated with increasingly distinguishable population-level representations of task-relevant stimuli, as a result of stabilization of existing and recruitment of new neurons selective for these stimuli. These effects correlated with the appearance of multiple task-dependent signals during learning: those that increased neuronal selectivity across the population when expert animals engaged in the task, and those reflecting anticipation or behavioral choices specifically in neuronal subsets preferring the rewarded stimulus. Therefore, learning engages diverse mechanisms that modify sensory and non-sensory representations in V1 to adjust its processing to task requirements and the behavioral relevance of visual stimuli.

Figure 1

Rapid Learning of a V1-Dependent Visual Discrimination Task in Virtual Reality (A) Schematic of the virtual reality setup. (B) Task schematic with virtual corridor wall patterns. CR, correct rejection, FA, false alarm. (C) Changes in licking over learning in an example mouse. Licks (dots) aligned to grating onset in vertical grating (left, blue shading) and angled grating (middle, pink shading) trials. Red dots, reward delivery; yellow dots, licking after reward delivery. Right, average running speed for session shown on left, aligned to grating onset for vertical (blue) and angled (red) trials. Shading, SEM. (D) Behavioral performance (behavioral d-prime; see Experimental Procedures) of five mice imaged on consecutive training sessions. See also Figure S1. (E) Behavioral performance in the visual and an equivalent odor discrimination task (see Experimental Procedures, average across sessions) as a function of light intensity during bilateral optogenetic silencing of visual cortex. PV-ChR2, transgenic mice expressing Channelrhodopsin-2 in parvalbumin-positive interneurons (n = 4 mice, 10 visual and 4 odor discrimination sessions). WT, wild-type mice (n = 3 mice, 7 sessions). ^∗p < 0.05, ^∗∗∗p < 0.001 after Bonferroni correction, Wilcoxon rank-sum test comparing PV-ChR2 to WT in the visual task.

Figure 2

Chronic Two-Photon Imaging of Single Cells across Learning (A) Example calcium traces of four V1 neurons during the task in an expert mouse, aligned to running speed (gray trace on top), licking (black lines), and reward delivery (red lines). Blue and red shading indicate time spent in the vertical and angled grating corridor, respectively. (B) Average responses and corresponding images of four additional example cells in four training sessions aligned to grating onset (dashed vertical line). Values above each trace on day 6 denote neuronal selectivity for grating corridors, computed from responses 0–1 s after grating onset (see Experimental Procedures). (C) Histograms of neuronal selectivity (positive values: cells prefer vertical, rewarded gratings; negative values: cells prefer angled, non-rewarded gratings) for different behavioral discrimination performance levels. Colors denote bins of behavioral d-prime from chance performance (blue) to expert performance (orange). (D) Proportions of neurons significantly preferring the vertical or the angled grating or those without preference, before (sessions with behavioral d-prime < 1) and after learning (behavioral d-prime > 2); session mean ± SEM computed from responses 0–1 s after grating onset. (E) Grating selectivity of the same neurons (rows) across sessions (columns) in the first three and last three sessions; cells were ordered based on the selectivity averaged across the middle four sessions; n = 8 mice. (F) Persistence of response selectivity across consecutive training sessions during different stages of learning. Values are the probability of a neuron with a grating preference on one day to maintain this preference on the next day within each learning stage (response 0–1 s after grating onset; vertical grating, N_pre = 51, N_dur = 121, N_pst = 279; angled grating, N_pre = 90, N_dur = 95, N_pst = 200 cells). Error bars depict SEM (determined by bootstrapping with replacement). Pre learning, behavioral d-prime (d′) of both sessions < 1, and Δd′ < 0.5 (14 session pairs); during learning: d′ first session < 2, d′ second session > 0.5, Δd′ > 0.5 (14 session pairs); after learning: d′ both sessions > 2, absolute Δd′ < 0.5 (19 session pairs). (G) The fraction of non-selective cells becoming selective for task-relevant stimuli across consecutive training sessions during different stages of learning (as in F). Values are the probability cells non-selective on one day (N_pre = 549, N_dur = 417, N_pst = 422) to develop a preference for one of the two gratings the next day within each learning stage. n = 11 mice for all panels, except where indicated. See also Figures S2–S5.

Figure 3

Learning Increases Neuronal Stimulus Selectivity in Populations of V1 Cells (A) Time course of neuronal population selectivity (see Experimental Procedures) aligned to grating onset (dashed line; 200 ms sliding time window) for different behavioral performance levels as in Figure 2C. Shading indicates SEM. (B) Time course of classification accuracy of a linear decoder (probability of correctly identifying vertical versus angled grating corridor trials), based on cumulative neuronal activity of simultaneously imaged cells from grating onset for different behavioral performance levels. Shading indicates SEM. (C) Relationship between population selectivity (average value 0–1 s after grating onset) and behavioral performance for individual sessions. Shades of gray indicate individual mice. n = 11 mice for all panels. See also Figures S6 and S7.

Figure 4

Learning-Related Increase in Neuronal Selectivity Is Partly Task-Dependent (A) Visual discrimination performance (behavioral d-prime) in interleaved blocks of the visual discrimination task and an analogous olfactory discrimination task in which the same grating stimuli were shown but ignored by the animals (black circles and lines; shades of gray indicate individual mice). Odor discrimination performance in the olfactory blocks is additionally shown (green circles). n = 4 mice. (B) Selectivity for grating stimuli of individual neurons during the visual and the olfactory task. The majority of neurons reduce their selectivity for the same grating stimuli during olfactory discrimination. (C) Proportions of neurons significantly preferring the vertical or the angled grating, or those without preference in the visual and the olfactory task (n = 11 sessions). (D) Time course of visual stimulus population selectivity aligned to grating corridor onset (dashed line) in the visual task (red, gratings relevant), in the olfactory task (black, gratings irrelevant), when the same stimuli are presented during anesthesia (gray, n = 11 mice), and before learning (blue, n = 8 mice). Error bars and shading are SEM. See also Figures S8 and S9.

Figure 5

Neuronal Changes during Learning Cannot Be Explained by Changes in Running Behavior (A) Population selectivity for different running speeds (thick traces, fast running trials; thin traces, slow running trials) matched across sessions before (behavioral d-prime < 1, gray traces) and after learning (behavioral d-prime > 2, black traces). Data are from 27 sessions pre learning and 52 sessions post learning. See also Figures S10–S13. (B) Average running speeds corresponding to conditions in (A). Solid lines indicate vertical grating trials and dashed lines angled grating trials. There was no difference in running speeds within the same speed bin across stimuli, nor before/after learning (all comparisons p > 0.08). (C) Time course of decoding performance from grating onset (probability of correct classification of grating corridor type, vertical versus angled). Decoder was based either on cumulative neuronal activity (solid lines) or cumulative running speed (dashed lines) for different behavioral discrimination performance levels during learning. n = 11 mice for (A)–(C). (D) Decoding performance as in (C), before learning (behavioral d-prime < 1, gray lines, 23 sessions, 11 mice), and after learning (behavioral d-prime > 2) for sessions with delayed divergence of running behavior in vertical and angled grating trials (see Experimental Procedures; purple, 8 sessions, 7 mice), and sessions with matched behavioral d-prime but early divergence of running behavior (black lines, 8 sessions, 6 mice). In all panels, 0 s = grating corridor onset.

Figure 6

Emergence of Signals Related to Behavioral Trial Outcome during Learning (A) Average running speed profile in different trial types: hit, correct rejection (CR) and false alarm (FA) (17 sessions). Dashed line indicates grating corridor onset. (B) Average grating responses of four example cells classified as behaviorally-modulated (cells 3 and 4; responses are different in FA and CR trials) and behaviorally not modulated (cells 1 and 2; responses are similar in FA and CR trials). (C) Proportions of neurons not modulated by behavior (only neurons with similar responses in FA and CR trials) which significantly preferred the vertical or the angled grating and without preference, before (behavioral d-prime < 1) and after learning (behavioral d-prime > 2); session mean ± SEM, 0–1 s after grating onset. (D) Time course of decoding performance, i.e., the probability of correct classification of CR and FA trials based on cumulative neuronal activity for three different behavioral discrimination performance levels during learning. (E) Average responses of cells preferring the vertical, rewarded grating (left) or angled, non-rewarded grating (middle) on Hit (blue), CR (red) and False alarm (black) trials. Right, change in response strength 0–1 s after angled grating onset during FA trials relative to CR trials. n = 11 mice for all panels.

Figure 7

Emergence of Anticipatory Signals during Learning (A) Increase in ramping activity prior to grating corridor onset (dashed line) in an example cell over the course of learning. (B) Average population responses to the vertical (blue) and angled grating (red) for vertical- (top) and angled-preferring (bottom) cells in the first training session (pre learning), in the first session after learning (post learning, behavioral d-prime > 2), and during anesthesia post learning. (C) Relative increase in pre-stimulus activity within the second preceding grating onset (see Experimental Procedures) for vertical and angled grating-preferring cells in the first training session, the first session post learning, and when virtual reality stimuli were played back to trained animals during anesthesia. p values from bootstrap test; error bars and shading are SEM. n = 11 mice.