Volatility-driven learning in human infants

Abstract

Adapting to change is a fundamental feature of human learning, yet its developmental origins remain elusive. We developed an experimental and computational approach to track infants' adaptive learning processes via pupil size, an indicator of tonic and phasic noradrenergic activity. We found that 8-month-old infants' tonic pupil size mirrored trial-by-trial fluctuations in environmental volatility, while phasic pupil responses revealed that infants used this information to dynamically optimize their learning. This adaptive strategy resulted in successful task performance, as evidenced by anticipatory looking toward correct target locations. The ability to estimate volatility varied significantly across infants, and these individual differences were related to infant temperament, indicating early links between cognitive adaptation and emotional responsivity. These findings demonstrate that infants actively adapt to environmental change, and that early differences in this capacity may have profound implications for long-term cognitive and psychosocial development.

Fig. 1.. A model-based approach to volatility estimation in infants.

(A) The reversal learning task involves multiple trials where a target stimulus appears from one of two locations. The predictability of the most likely target location was manipulated, such that it could either be stable (e.g., first 18 trials) or change more frequently (e.g., in trials between 19 and 36) thereby introducing volatility. (B) A schematic representation of the computational model and the dependent measures collected from infants’ gaze and pupil data. The model was a volatile Kalman filter (VKF), where $v_0$ indicates the initial volatility, $\mathrm{\lambda }$ is the volatility learning rate, $v_{\mathrm{t}}$ is the volatility of each trial, $p_{\mathrm{t}}$ is the belief about the target location, and $o_{\mathrm{t}}$ is the observed outcome of the target location. Observing the outcome would change the internal belief about the target location, as well as higher order beliefs about how volatile the environment is. In turn, these higher-order beliefs would guide future predictions about the target location. (C) The target locations order (red and blue dots) with the VKF’s probability estimates about the target location (red line) and the VKF’s volatility estimates (blue line) from an example sequence.

Fig. 2.. Infants’ tonic pupil size and preferential looking.

(A) Infants’ tonic pupil size correlated with the VKF model’s estimates of volatility, indicating that they successfully tracked fluctuations in environmental volatility across the task. (B) Infants’ preferential looking before target appearance correlated with the VKF model’s predictions about the target location in both stable and volatile periods. This shows that infants were able to integrate volatility information to update their expectations about the target location. Predictive means (y axes) were obtained generating trial-by-trial estimates of pupil size and proportion of anticipatory looking from the β coefficients of the fitted regression models.

Fig. 3.. Infants’ phasic pupil size.

(A) Raw data for phasic pupil size (baseline corrected) at the moment of the target presentation, divided in four groups depending on volatility (mean split in high or low) and the magnitude of the prediction error (big or small). The initial, apparent decrease in pupil size is due to saccades to the target locations on the lateral portions of the screen (0 to 1400 ms). Following this, the slower task-evoked pupil responses emerge (after 1400 ms). (B) Predictive means of phasic pupil size as a function of environmental volatility and the magnitude of the prediction errors. Infants considered volatility information when estimating the relevance of prediction errors, such that greater prediction errors were considered as more important when volatility was high (as they might signal a change in the environment). Conversely, smaller prediction errors were considered as more important when the environment was stable (as they corroborated existing beliefs about the target location).

Fig. 4.. Individual differences in volatility estimation.

(A) Three example infants displaying volatility overestimation, correct estimation, and underestimation as indexed by high, intermediate, and low $\mathrm{\delta }$ values, respectively. (B) $\mathrm{\delta }$ mean values and confidence intervals for all infants. The $\mathrm{\delta }$ values of the three example infants are highlighted in green. The dashed line indicates the group average value of $\mathrm{\delta }$ . (C) Individual differences in volatility estimation were found to be related to regulatory capacity and positive affect, but not negative emotionality. (D) Schematic representation of the potential developmental pathway from early volatility estimation abilities to psychological outcomes. Solid arrows indicate associations that were found in previous research, and dashed arrows indicate possible associations; purple arrow indicates the current findings. All arrows are bidirectional as causality cannot be established. Volatility estimation early in life is related to temperament, and it might relate to internalizing/externalizing difficulties in childhood, as well as anxious and depressive traits in adulthood (–38). Hence, volatility estimation might offer a transdiagnostic marker that captures the existing longitudinal relations between infants’ temperament, children’s internalizing and externalizing difficulties, and anxious and depressive symptoms and/or neurodevelopmental conditions such as autism and ADHD (, –43). This hypothesis awaits empirical testing through longitudinal research.