How neuronal selectivity distributions change with experience

Most measured visual-cortical plasticity was reproduced without reward

The paper reports 89 recordings in 19 TetO-GCaMP6s × CaMK2a-tTA mice (Methods: animals ↗) and 20,547–89,577 Suite2p-derived traces per recording across V1 and higher visual areas (Results: supervised and unsupervised plasticity ↗; Figure 1e ↗). The rewarded task and matched unrewarded-exposure cohorts are defined in the same Results section and timeline (Results section ↗; Figure 1a–b ↗).

After Train 1, the paper reports increased familiar-stimulus selectivity in task and unrewarded natural-texture cohorts, with no comparable increase after grating exposure (Results: supervised and unsupervised plasticity ↗; Figure 1i–j ↗). Separately, it reports a late-versus-early cue response concentrated in anterior HVAs of task mice (Results: reward prediction in anterior HVAs ↗; Figure 4e–g ↗).

Open Figure 1 on Nature ↗

The paper reports plasticity mainly as a before-versus-after endpoint comparison after roughly two weeks of exposure (Fig. 1b,i,jpaper ↗). It establishes a change in the fraction and cortical distribution of selective neurons, but does not ask whether the shape of the full d′ distribution changes across trial order. That source-grounded extension is defined in §11.

What the complete paper establishes

Each numbered figure is a multi-panel argument, not a single result. The cards below keep the decisive subfigures beside the explanation, identify what every panel group measures, and distinguish the authors’ published conclusion from the new analysis that this reference proposes. Select any screenshot to jump to its larger annotated rendering elsewhere in the document.

01

Nature Figure 1 · panels a–j

Plasticity after supervised and unsupervised training

The figure moves from task design and behavioral verification to the cellular selectivity definition, then to cortical maps and regional group statistics.

Figure 1a–b ↗ · design

Panel a defines the circle1 and leaf1 corridors, randomized sound cue, and reward availability. Panel b separates task, unrewarded natural-texture, and unrewarded grating cohorts across Train 1/Test 1/Train 2/Test 2/Test 3 landmarks.

Figure 1c–d ↗ · behavior

Example lick rasters and anticipatory-lick summaries verify that the task cohort learned the rewarded corridor; these are behavioral checks, not the neural effect itself.

Figure 1e–f ↗ · measurement

Panel e establishes mesoscope coverage and cellular resolution. Panel f defines signed leaf1-versus-circle1 d′ from response distributions inside the two corridors.

Figure 1g–h ↗ · response structure

Single-neuron trials and sorted population sequences show the positive leaf1 and negative circle1 selectivity poles rather than collapsing them into |d′|.

Figure 1i–j ↗ · endpoint

Aligned cortical density maps and regional selective-neuron fractions compare before versus after across visual areas and cohorts; the mouse/session summaries, not the thousands of neurons, carry group replication.

02

Nature Figure 2 · panels a–j

Visual identity is retained across corridor position

The figure tests whether neural sequences encode the visual stimulus itself, rather than merely repeating a position-locked trajectory.

Figure 2a–c ↗ · stimulus and behavior

Four learned/new corridors are presented in Test 1; lick rasters and anticipatory licking establish which stimulus identities the mice behaviorally distinguish.

Figure 2d–f ↗ · sequence reliability

Medial-area responses are sorted by leaf1 preferred position on held-out trials. Preferred-position correlations are then summarized within task mice and across regions and cohorts, including leaf2 versus leaf1.

Figure 2g–h ↗ · trial-resolved populations

Leaf1- and circle1-selective neurons are shown as a population average for each trial, preserving both selectivity poles and the trial-by-position response structure.

Figure 2i–j ↗ · geometry

A coding direction is defined from the leaf1- and circle1-selective populations. Held-out projections yield a similarity index for new stimuli, summarized by region and cohort.

03

Nature Figure 3 · panels a–h

Fine discrimination reshapes representational geometry

The figure asks what happens when mice must distinguish the familiar leaf1 corridor from the visually similar leaf2 corridor.

Figure 3a–b ↗ · leaf2 selectivity

Maps and regional summaries compare neurons selective for leaf2 versus circle1 when leaf2 is new and after learning, alongside unrewarded and naive cohorts.

Figure 3c ↗ · behavioral learning

Licking to leaf2 before and after training establishes the behavioral fine-discrimination transition.

Figure 3d–e ↗ · leaf1 versus leaf2

Signed selective-neuron distributions and regional fractions directly compare the two similar natural textures across task, unrewarded, naive, and grating-control observations.

Figure 3f–g ↗ · coding projections

Leaf2 population responses are projected onto the familiar leaf1–circle1 coding direction in V1 and medial HVA.

Figure 3h ↗ · orthogonalization

The similarity index summarizes how much leaf2 remains aligned with the familiar axis; the schematic expresses the observed rotation away from leaf1.

04

Nature Figure 4 · panels a–n + Extended Data 8

A separate anterior reward-prediction signal

This figure defines a late-cue-versus-early-cue estimand, localizes it, and then tests cue, lick, corridor, and movement interpretations.

Figure 4a–c ↗ · discovery and location

Rastermap exposes a task-linked population sequence, a selected segment is enlarged, and the selected cells are mapped back onto cortex.

Figure 4d–g ↗ · estimand and regional effect

Trial averages sorted by cue position motivate a late-versus-early cue d′. Density maps and fractions then show the strongest task-specific increase in anterior HVA.

Figure 4h–j ↗ · stimulus and timing

Responses across Test 1 corridors are followed by cue-aligned and first-lick-aligned averages, separating cue timing from simple lick timing.

Figure 4k–l ↗ · lick controls

Reward-prediction and medial leaf-selective populations are compared on trials with versus without licks.

Figure 4m–n ↗ · transfer tests

Population responses are repeated in Test 2 and Test 3 to test how the signal behaves under later stimulus arrangements.

Extended Data Figure 8 ↗ · specificity

Panel a compares reward-prediction fractions across regions; b–c define non-reward prediction controls; d–f place neural traces beside corridor-specific running and licking.

05

Nature Figure 5 · panels a–h + Extended Data 9

Natural-texture pretraining accelerates later behavior

A separate behavior-only experiment compares natural-texture, grating, and no-pretraining cohorts across five rewarded training days.

Figure 5a–b ↗ · independent design

Three new cohorts receive natural-texture, grating, or no unrewarded pretraining, then five rewarded task days. Rewards are deterministic in the second corridor half and the sound cue is absent, unlike the imaging task.

Figure 5c–d ↗ · example behavior

Lick rasters show representative mice on the first active-reward day and the last training day.

Figure 5e–f ↗ · learning trajectory

Mean lick responses and the rewarded-minus-non-rewarded performance difference are tracked across days for all three cohorts.

Figure 5g–h ↗ · spatial and support checks

First-lick locations test where behavioral anticipation emerges; daily trial counts show the observation support behind each curve.

Extended Data Figure 9 ↗ · within day

Each day is split into first and second halves, making the early within-session behavioral improvement visible rather than inferred from day averages alone.

M

Nature Methods · acquisition and processing

What the plotted neural values actually are

The figures use processed cellular activity aligned to behavior and visual areas—not raw movies and not a continuous identity-tracked recording across days.

Signal

Suite2p-derived deconvolved activity with a 0.75 s decay parameter is the paper’s neural layer (Methods: processing of calcium imaging data ↗). The shared Figshare files are processed arrays, not raw fluorescence movies (Data availability ↗; Figshare v2 inventory ↗).

Valid frames

The paper’s selectivity estimator retains original, non-interpolated running frames inside the 0–4 m textured corridor (Methods: neural selectivity ↗). Cue, reward, and lick timing come from the imaging behavioral protocol (Methods: behavioural training ↗); cortical regions come from retinotopic assignment (Methods: retinotopy ↗).

Scale

The paper reports 20,547–89,577 Suite2p traces per recording (Results: supervised and unsupervised plasticity ↗) and 89 recordings in 19 mice (Methods: animals ↗). Those are separate descriptive and experimental sample sizes.

Published observation unit

Figure 1f forms d′ from frame-response distributions (Figure 1f ↗; Methods: neural selectivity ↗). The proposed trial-window analysis preserves the signed algebra but replaces pooled frames with balanced per-trial summaries.

The animals and the training schedule

The paper’s Animals Methods report 89 recordings in 19 TetO-GCaMP6s × CaMK2a-tTA mice (13 male, 6 female; 2–11 months) (Methods: animals ↗). The filename key recording_id = {mouse}_{date}_{block} and cohort prefixes come from the deposited Imaging_Exp_info.npy ↗, not from the prose paper. Counts are panel-specific: Figure 1j ↗ reports n=4 task, 9 unrewarded natural-texture, and 3 grating mice; other panels use the counts shown explicitly below.

Stage schedule

The task and unrewarded imaging cohorts follow the staged Train/Test sequence below, with before-learning and after-learning imaging blocks roughly two weeks apart (timeline, Fig. 1bpaper ↗). Naive and grating controls contribute comparison sessions rather than passing through every task stage. The release index contains 23 experiment labels spanning cohort, stage, and learning status.

Stimulus roles are session-specific

The paper refers to stimulus roles in its naturalistic-corridor design (Nature Fig. 1a,bpaper ↗); the released files store physical wall names. The mapping between them is set per session, and conflating the two is a frequent source of mislabeling.

Contents of the release, and the tractable access path

The Nature Data availability statement deposits the study on Figshare; the workspace pins the updated release (doi:10.25378/janelia.28811129.v2, CC BY 4.0), and the paper recipes are in the paper-tagged analysis release. Files are keyed by recording_id = {mouse}_{date}_{block}. The reduced representation reconstructs spk = UTV and supports efficient exploratory population analysis. Exact paper-style per-neuron selectivity uses the full deconvolved traces; agreement between representations must be measured rather than assumed.

How much neural data exists, and whether it is day by day

The imaging release described by the Nature article and its Data availability statement contains 89 physical recordings from 19 mice. The workspace pins the updated Figshare v2 release: 297 files and 452,233,500,962 bytes (421.175 GiB). The values below are computed from that immutable inventory rather than inferred from folder names.

The publication originally cited Figshare v1. Version 2 retains those data and adds 89 reduced SVD files plus one neural example; it adds a tractable representation, not 89 new recording sessions.

What “day by day” means in this release

There are 89 mouse–date recordings on 77 calendar dates. Eighteen of 19 mice were imaged more than once, with a median of five recording dates per mouse. Across the 70 within-mouse intervals, the median gap is three days (range 1–67); only 21 intervals are one day apart. Those adjacent dates cluster around test-stage boundaries and do not form a complete daily training series. This sampling follows the milestone design shown in Nature Fig. 1bpaper ↗.

Download footprint for the 13-mouse candidate analysis