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 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.
Introduction
In both biological and artificial systems, acquiring a new skill typically involves "active learning"—effortful practice guided by feedback such as rewards or error signals. However, organisms are constantly immersed in sensory environments even when not actively performing a task. This study explores whether this "passive exposure" can be leveraged to accelerate active learning and identifies the computational mechanisms that enable passive data to enhance task performance.
Behavioral Evidence in Mice
To test the impact of passive exposure, adult mice were trained on a two-alternative choice task to categorize the slope of frequency-modulated (FM) sounds. The researchers compared three distinct training cohorts:
- Active only (A): Standard training with feedback.
- Active + Passive (A+P): Passive exposure sessions interleaved between active training days.
- Passive then Active (P:A): A block of passive exposure before active training began.
The results demonstrated that mice receiving passive exposure (either before or interleaved) learned significantly faster than those in the active-only group. Mice with passive exposure reached a 70% accuracy threshold in approximately 12–15 days, whereas the active-only group required over 20 days. Notably, a few interleaved passive sessions proved as effective as a long block of pre-exposure.
Computational Modeling of Learning Mechanisms
The team developed five Artificial Neural Network (ANN) models to investigate which architectures could replicate the observed mouse behavior. These models explored combinations of supervised learning (SL) from feedback and unsupervised learning (UL) from the statistical structure of the data.
The study found that simple single-layer models failed to benefit from passive exposure. The critical architectural requirement was a multi-layer structure where unsupervised learning occurs in the early layers. This process builds a "latent representation" that enhances the separability of the data before the final layer applies task-specific labels. Only "Model 5," which combined supervised and unsupervised learning at the input layer, successfully replicated the behavioral curves of all three mouse cohorts.
Synergy and Generalization
Analysis of the interleaved (A+P) schedule revealed a high degree of alignment between weight updates from passive (unsupervised) and active (supervised) sessions. Passive sessions essentially "prime" the network weights in a direction that supports the active task, increasing overall training efficiency. Furthermore, passive exposure improved the categorization of intermediate sounds for which the mice had not yet received rewards, suggesting a fundamental and generalized shift in sensory representation.
Conclusion
This study provides evidence that passive exposure is a powerful, low-effort tool for accelerating categorization learning. By building more separable internal representations of stimuli, passive exposure reduces the amount of effortful training required to reach peak performance. These findings offer valuable insights for designing more efficient training protocols in both biological research and the development of artificial intelligence.
Summary updated