CoChat

Passive exposure to task-relevant stimuli enhances categorization learning

Open access
Schmid, Christian, Haziq, Muhammad, Baese-Berk, Melissa M, Murray, James M, Jaramillo, Santiago · 2024 · eLife

Published version of a preprint

The DOI matches its Crossref record; 41 of the 42 references it cites were found in a registry or an index; none of the 11 sources its claims cite is checked for support yet.

Abstract

Learning to perform a perceptual decision task is generally achieved through sessions of effortful practice with feedback. Here, we investigated how passive exposure to task-relevant stimuli, which is relatively effortless and does not require feedback, influences active learning. First, we trained mice in a sound-categorization task with various schedules combining passive exposure and active training. Mice that received passive exposure exhibited faster learning, regardless of whether this exposure occurred entirely before active training or was interleaved between active sessions. We next trained neural-network models with different architectures and learning rules to perform the task. Networks that use the statistical properties of stimuli to enhance separability of the data via unsupervised learning during passive exposure provided the best account of the behavioral observations. We further found that, during interleaved schedules, there is an increased alignment between weight updates from passive exposure and active training, such that a few interleaved sessions can be as effective as schedules with long periods of passive exposure before active training, consistent with our behavioral observations. These results provide key insights for the design of efficient training schedules that combine active learning and passive exposure in both natural and artificial systems.

Overview

CoChat Summary

TL;DR

Passive exposure to task-relevant frequency-modulated sounds sped up sound-categorization learning in adult mice whether it occurred before or interleaved with active training. Neural-network models suggest this benefit arises when unsupervised learning in early layers builds latent representations that supervised readout can use, with interleaved schedules aligning active and passive weight updates.

Key findings

  • Mice with passive exposure learned faster than active-only mice: time to 70% correct was shorter (p=0.01 for A only vs A+P; p=0.006 for A only vs P:A; p=0.71 for A+P vs P:A, Wilcoxon rank-sum test), and estimated performance at 21 days was higher (p=0.01, p=0.006, p=0.47 respectively). Last-4-day S3 performance matched these comparisons (p=0.01, p=0.006, p=0.71).
  • Passive exposure also improved psychometric performance in S4, including intermediate sounds not used in S3: overall performance p=0.01 for A only vs A+P, p=0.01 for A only vs P:A, p=0.27 for A+P vs P:A; intermediate stimuli p=0.037, p=0.028, p=0.36; psychometric slopes did not differ significantly (p=0.07, p=1, p=0.86). After 21 days of S4, A-only mice improved (p=0.031) and all cohorts became indistinguishable (p>0.26).
  • In models, a one-layer model did not benefit from passive pre-exposure; two-layer models with unsupervised learning in the first layer and supervised learning at readout accounted for faster learning with passive exposure. Model 5, combining supervised and unsupervised learning at the input layer, reproduced similar A+P and P:A benefits and psychometric curves; active and passive weight updates were more aligned for A+P than P:A (p<10⁻³, Wilcoxon rank-sum test).

Methods

  • 27 wild-type C57BL/6J adult mice (both sexes, 2.5–4 months) were water-restricted and trained on a two-alternative choice task to discriminate positive vs negative slopes of 200-ms frequency-modulated sounds; cohorts were A only (8 mice), A+P (9), P:A (9), with one mouse excluded. Passive exposure consisted of ~3600 trials/day (600 presentations of each of six stimuli, every 4.5 s) in home cages.
  • Behavioral analysis fit a straight line to daily S3 performance to estimate 21-day performance and days to 70% correct; Gaussian mixture models tested bimodality; psychometric curves were sigmoidal fits; cohort comparisons used Wilcoxon rank-sum tests.
  • Neural-network models were implemented in JAX with five architectures/learning rules; schedules were A only (5000 active trials), A+P (5000 active trials each followed by 9 passive exposures), and P:A (45,000 passive exposures then 5000 active trials), using isotropic, non-isotropic, and non-aligned input distributions.

Limitations

  • The paper states that its modeling approach omits features that may be important for fully capturing unsupervised learning during passive exposure in the brain; more sophisticated approaches beyond linear dimensionality reduction will be required when relevant stimulus features are encoded nonlinearly, as likely for natural sounds.
  • The paper states that the comparable benefits of P:A and A+P could be a consequence of the large number of daily passive-exposure trials, and that a comprehensive evaluation of learning performance as a function of passive-exposure number is needed; models most often found pre-exposure led to larger gains (Models 3 and 4), though Model 5 did not.
  • Evident from the text: the study is limited to one auditory categorization task in mice, with no neural recordings, and the slow-learner comparisons were not statistically significant (p-values 0.064–0.16); the models use simplified linear networks and specific stimulus distributions.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.7554/elife.88406

Cites

41 of 41 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1. Two-alternative choice sound-categorization task for mice. (A) Mice initiated a trial by poking their nose into the center port of a three-port chamber, triggering the presentation of a frequency-modulated (FM) sound. To obtain reward, animals had to choose the correct side port according to the slope of the frequency modulation (left for positive slopes, right for negative slopes). (B) Training schedule: mice underwent several shaping stages (S0–S2) before learning the full task; the main learning stage (stage S3) used only the highest and lowest FM slopes; psychometric performance was evaluated using 6 different FM slopes (stage S4). (C) Daily performance for one mouse during S3. Arrows indicate estimates of the time to reach 70% correct and the performance at 21 days given a linear fit (black line). (D) Average leftward choices for each FM slope during one session of S4 for the mouse in C. Error bars indicate 95% confidence intervals.
Figure 1. Two-alternative choice sound-categorization task for mice. (A) Mice initiated a trial by poking their nose into the center port of a three-port chamber, triggering the presentation of a frequency-modulated (FM) sound. To obtain reward, animals had to choose the correct side port according to the slope of the frequency modulation (left for positive slopes, right for negative slopes). (B) Training schedule: mice underwent several shaping stages (S0–S2) before learning the full task; the main learning stage (stage S3) used only the highest and lowest FM slopes; psychometric performance was evaluated using 6 different FM slopes (stage S4). (C) Daily performance for one mouse during S3. Arrows indicate estimates of the time to reach 70% correct and the performance at 21 days given a linear fit (black line). (D) Average leftward choices for each FM slope during one session of S4 for the mouse in C. Error bars indicate 95% confidence intervals.CC BY · PubMed Central
Figure 2. Passive exposure to sounds improves learning speed. (A) Training schedule for each mouse cohort: A-only mice received no passive exposure; A + P mice received passiveexposure sessions during stage S3; P:A mice received a similar number of passive-exposure sessions before S2. (B) Distributions of performance at 21 days of S3 given estimates from linear fits to the learning curve for each mouse from each cohort. Solid lines represent the results from a Gaussian mixture model with two components, separating ‘fast’ from ‘slow’ learners. (C) Distributions of times to reach 70% correct given estimates from linear fits. (D) Average learning curves across fast learners from each cohort. Shading represent the standard error of the mean across mice. (E) Estimates of the time to reach 70% for fast learners from each group. Each circle is one mouse. Horizontal bar represents the median. (F) Estimates of the performance at 21 days for fast learners from each group. (G) Actual performance averaged across the last 4 days of S3 for fast learners from each group. Stars indicate p<0.05.
Figure 2. Passive exposure to sounds improves learning speed. (A) Training schedule for each mouse cohort: A-only mice received no passive exposure; A + P mice received passiveexposure sessions during stage S3; P:A mice received a similar number of passive-exposure sessions before S2. (B) Distributions of performance at 21 days of S3 given estimates from linear fits to the learning curve for each mouse from each cohort. Solid lines represent the results from a Gaussian mixture model with two components, separating ‘fast’ from ‘slow’ learners. (C) Distributions of times to reach 70% correct given estimates from linear fits. (D) Average learning curves across fast learners from each cohort. Shading represent the standard error of the mean across mice. (E) Estimates of the time to reach 70% for fast learners from each group. Each circle is one mouse. Horizontal bar represents the median. (F) Estimates of the performance at 21 days for fast learners from each group. (G) Actual performance averaged across the last 4 days of S3 for fast learners from each group. Stars indicate p<0.05.CC BY · PubMed Central
Figure 3. Passive exposure improves categorization of intermediate stimuli. (A) Average psychometric performance for the first 4 days of stage S4 across fast learners from each group. Error bars show the standard error of the mean across mice. (B) Performance averaged across all stimuli is better for mice with passive exposure. Horizontal lines indicate median across mice. (C) Performance for extreme stimuli (included in S3) is better for mice with passive exposure. (D) Performance for intermediate stimuli (which were not used in the task before S4) is better for mice with passive exposure. (E) Psychometric slope is not different across groups. (F–J) All groups achieve similar levels of performance after 3 weeks of additional training. Stars indicate p<0.05.
Figure 3. Passive exposure improves categorization of intermediate stimuli. (A) Average psychometric performance for the first 4 days of stage S4 across fast learners from each group. Error bars show the standard error of the mean across mice. (B) Performance averaged across all stimuli is better for mice with passive exposure. Horizontal lines indicate median across mice. (C) Performance for extreme stimuli (included in S3) is better for mice with passive exposure. (D) Performance for intermediate stimuli (which were not used in the task before S4) is better for mice with passive exposure. (E) Psychometric slope is not different across groups. (F–J) All groups achieve similar levels of performance after 3 weeks of additional training. Stars indicate p<0.05.CC BY · PubMed Central
Figure 4. A single-layer model (Model 1) does not benefit from passive pre-exposure. (A) Network architecture for the one-layer model. (B) The network is trained to find a hyperplane orthogonal to the decoding direction Δμ→0. (C) Learning performance for different training schedules. Curves show mean accuracy for n=50 network realizations, and shading shows standard deviation.
Figure 4. A single-layer model (Model 1) does not benefit from passive pre-exposure. (A) Network architecture for the one-layer model. (B) The network is trained to find a hyperplane orthogonal to the decoding direction Δμ→0. (C) Learning performance for different training schedules. Curves show mean accuracy for n=50 network realizations, and shading shows standard deviation.CC BY · PubMed Central
Figure 5. Passive-exposure benefits learning in a two-layer model with unsupervised learning in the first layer. (A) Network architecture for Model 2, which has supervised learning (SL) at the input layer, and unsupervised learning (UL) at the readout. (B) Learning dynamics of the hidden-layer representation of this model for the P:A schedule. (C) Learning performance for different training schedules for Model 2. Curves show mean accuracy for n=50 network realizations, and shading shows standard deviation. (D–F) Model 3, which has unsupervised learning (UL) at the input layer, and supervised learning (SL) at the readout.
Figure 5. Passive-exposure benefits learning in a two-layer model with unsupervised learning in the first layer. (A) Network architecture for Model 2, which has supervised learning (SL) at the input layer, and unsupervised learning (UL) at the readout. (B) Learning dynamics of the hidden-layer representation of this model for the P:A schedule. (C) Learning performance for different training schedules for Model 2. Curves show mean accuracy for n=50 network realizations, and shading shows standard deviation. (D–F) Model 3, which has unsupervised learning (UL) at the input layer, and supervised learning (SL) at the readout.CC BY · PubMed Central
Figure 6. An alternative unsupervised learning rule can be used to learn hidden representations of higher principal components. (A) A non-isotropic input distribution in which the coding dimension does not align with the direction of highest variance. (B) Learning performance for Model 3 on the non-isotropic input distribution. Curves show mean accuracy for n=50 network realizations, and shading shows standard deviation. (C) Network architecture for a two-layer model (Model 4) that uses the similarity matching algorithm for the input weights and supervised learning at the readout. (D) Learning performance for Model 4 on the non-isotropic input distributions.
Figure 6. An alternative unsupervised learning rule can be used to learn hidden representations of higher principal components. (A) A non-isotropic input distribution in which the coding dimension does not align with the direction of highest variance. (B) Learning performance for Model 3 on the non-isotropic input distribution. Curves show mean accuracy for n=50 network realizations, and shading shows standard deviation. (C) Network architecture for a two-layer model (Model 4) that uses the similarity matching algorithm for the input weights and supervised learning at the readout. (D) Learning performance for Model 4 on the non-isotropic input distributions.CC BY · PubMed Central
Figure 7. A more-general two-layer model accounts for similar benefits of A + P and P:A training schedules. (A) Schematic illustration of Model 5, which combines supervised and unsupervised learning at the input layer of weights. (B) Input distribution, in which the decoding direction is not aligned with any particular principal component. (C) Learning performance of Model 5. Curves show mean accuracy for n=50$ network realizations, and shading shows standard deviation. (D) Psychometric curves showing classification performance for all stimuli after 1500 trials, where the stimulus parameter ρ linearly interpolates between the two extreme stimulus values. (E) Psychometric curves showing classification performance after 5000 trials. (F) Schematic illustration of the angle between the summed weight updates during active training and passive exposure for A + P (top) and P:A (bottom) learners. (G) The alignment of active and passive weight updates for n=50 networks trained with either the A + P or P:A schedule after 1500 trials (star indicates p<10⁻³, Wilcoxon rank-sum test; box percentiles are 25/75).
Figure 7. A more-general two-layer model accounts for similar benefits of A + P and P:A training schedules. (A) Schematic illustration of Model 5, which combines supervised and unsupervised learning at the input layer of weights. (B) Input distribution, in which the decoding direction is not aligned with any particular principal component. (C) Learning performance of Model 5. Curves show mean accuracy for n=50$ network realizations, and shading shows standard deviation. (D) Psychometric curves showing classification performance for all stimuli after 1500 trials, where the stimulus parameter ρ linearly interpolates between the two extreme stimulus values. (E) Psychometric curves showing classification performance after 5000 trials. (F) Schematic illustration of the angle between the summed weight updates during active training and passive exposure for A + P (top) and P:A (bottom) learners. (G) The alignment of active and passive weight updates for n=50 networks trained with either the A + P or P:A schedule after 1500 trials (star indicates p<10⁻³, Wilcoxon rank-sum test; box percentiles are 25/75).CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 0 Not valid · 7 Not checked · 55 Own results and interpretations

28 more claims

Create a free account to see all 62 claims and to verify the 10 references still open.

See all claims — free account
  • It may state two separate claims as one. 16
  • Probably not a substantial claim. 15

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Research works printed but not linked: 1

  1. Schmid C. 2023. Behaviour-models. Software Heritage. swh:1:rev:bdf55cd53c845fe00b7bc62476584f4eab1abc66. https://archive.softwareheritage.org/swh:1:dir:7fa425b87271bd779b361ea0a1d92ee699db85c7;origin=https://github.com/cschmidat/behaviour-models;visit=swh:1:snp:92c05cc5e8af3e633686e4f5c2417940a2417e14;anchor=swh:1:rev:bdf55cd53c845fe00b7bc62476584f4eab1abc66.

Cited by

See all 6

Metadata

Authors
Schmid, C., Haziq, M., Baese-Berk, M. M., Murray, J. M., & Jaramillo, S.
Affiliations
  • University of Oregon (US)
  • Department of Linguistics, University of Oregon
Corresponding author
  • Murray, James M
  • Jaramillo, Santiago
Venue
eLife
Published
2024-01-24
Volume
12
ISSN
2050-084X (ISSN-L)
OpenAlex
W4385066317
Cited by
6
References
42
Citations by year
2026: 1 · 2025: 2 · 2024: 2
FWCI
0.63
Type
journal article
Language
en
Open access
Locations
  • eLife (publishedVersion, cc-by)
  • eLife (submittedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (submittedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by-nc)
  • PubMed (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Science Foundation (2024926)
  • National Institute of Neurological Disorders and Stroke (Pathway to Independence Award R00NS114194, R01NS118461)
  • University of Oregon (Office of the Vice President for Research & Innovation I3 award)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Memory and Neural Mechanisms
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Algorithms
Keywords
passive exposure, category learning, perceptual decision making, unsupervised learning, audio classification, neural network model, weight updating, mice
MeSH terms
Behavior Observation Techniques, Animals, Sound, Neural Networks, Computer, Mice

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-19, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-19 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.