A computational model of infant sensorimotor exploration in the mobile paradigm

Technical Summary: A Computational Model of Infant Sensorimotor Exploration in the Mobile Paradigm

Problem Statement
The paper addresses the computational mechanisms underlying infant sensorimotor learning, specifically within the "mobile paradigm." In this developmental psychology paradigm, an infant's limb is connected to a mobile hanging above a crib; infants learn to detect the contingency between their limb movements and the resulting sensory feedback (movement of the mobile). While empirical studies have established that infants (particularly at 5–6 months) can rapidly distinguish the "connected" limb and increase its activity, and that this learning is sensitive to the contingency itself rather than mere stimulation, the underlying cognitive and neural mechanisms remain debated. Previous computational attempts have either relied on simple operant conditioning without predictive capabilities (failing to explain phenomena like "extinction bursts") or utilized complex information-theoretic approaches deemed impractical for biological brains. This paper proposes a detailed computational model to replicate infant behavior, test specific learning mechanisms, and investigate controversial findings such as the extinction burst and the difference between binary and conjugate (non-binary) stimulation conditions.

Methodology
The authors developed a neural network-based model simulating an infant interacting with a mobile. The model is grounded in two guiding principles: (1) the infant attempts to predict the sensory outcomes of its actions (action-outcome prediction), and (2) the motor system is biologically realistic, involving redundancy and noise.

• Architecture: The model consists of four main modules:

  1. Neural Network: A fully connected network with 5 input neurons (4 limb activities + 1 sensory feedback), 8 hidden neurons, and 601 output neurons (600 for muscle commands + 1 for sensory prediction).
  2. Motor Control & Body: The 600 muscle command neurons are mapped to four limbs using a population code inspired by human anatomy. This mapping uses a Beta function distribution to create overlapping control, where some neurons strongly influence a specific limb, others have intermediate effects, and many have negligible effects. Motor noise (random values between -0.3 and 0.3) is added to limb activities to simulate biological variability.
  3. Sensory Feedback: The environment simulates a mobile that responds to the "connected" limb. Two conditions are tested: Binary (mobile activates for a fixed duration if limb activity exceeds a threshold) and Conjugate/Non-binary (mobile intensity is proportional to limb activity).
  4. Learning Mechanism: The network is trained online via backpropagation to minimize a composite Loss function comprising three terms:
     • Sensory Prediction Error: The Mean Squared Error (MSE) between predicted and actual sensory feedback.
     • Exploration Error: An MSE term derived from an "Activity Interest Map." This map tracks the "interest" in different limb activity ranges. If an action yields a sensory prediction error exceeding a "Novelty Threshold" (0.1), the interest in that range is set to 1 (high interest). If the prediction is correct, interest decays. The network is penalized for not moving toward high-interest ranges.
     • Deviation from Baseline: An MSE term ensuring limb activity stays near a preferred baseline level, which gradually increases over time to simulate growing fussiness.

• Experimental Setup: The model was run for 20 simulations per condition (Binary and Non-Binary), matching the participant counts of recent empirical studies (Popescu et al., 2022; Jacquey et al., 2021). Each run included a 1000-step initialization, a 1200-step contingent phase, and a 240-step extinction phase (disconnection).

Key Contributions

1. Replication of Core Phenomena: The model successfully replicates the primary empirical finding: the rapid differentiation and selective increase in activity of the "connected" limb compared to unconnected limbs.
2. Contingency Sensitivity: The model demonstrates higher activity in contingent conditions compared to non-contingent (random) control groups, confirming that the learning mechanism is sensitive to the causal link, not just the presence of stimulation.
3. Binary vs. Conjugate Effects: The model reproduces the empirical finding that contingency detection is more robust and faster in binary conditions than in conjugate conditions. The authors attribute this to the larger sensory change per unit of movement in binary conditions, making the "connected" limb easier to identify.
4. Extinction Burst Simulation: The model occasionally exhibits an "extinction burst" (a transient increase in activity upon disconnection), mirroring the inconsistent observation of this phenomenon in human infants. The authors suggest this arises from the violation of predictive expectations.
5. Ablation Studies: The paper provides rigorous evidence for the necessity of specific components:
   • Prediction: Removing the prediction error loss prevents the model from distinguishing connected limbs.
   • Exploration: Removing the novelty-based reinforcement calculator eliminates the drive to explore, preventing learning.
   • Motor Redundancy: Reducing the number of muscle command neurons (from 600 to 50 or 100) significantly hinders the ability to learn the connection, suggesting that high-dimensional motor control is crucial for discovery.
   • Motor Noise: Both insufficient and excessive motor noise inhibit learning; an intermediate level (0.3) is optimal for enabling the exploration necessary to break symmetry between limbs.

Results
The simulations showed that the complete model learns to activate the connected limb within the first few minutes of the simulation, closely matching the time-course of 6-month-old infants.

• Specificity: In the contingent group, the connected limb's activity rose significantly above unconnected limbs. In the non-contingent group, no such differentiation occurred.
• Extinction: While not systematic, extinction bursts appeared in individual simulation runs, particularly in the non-binary condition, supporting the hypothesis that this behavior stems from predictive mechanisms.
• Ablation Outcomes: Models without prediction error or exploration mechanisms failed to learn. Models with reduced muscle command neurons or inappropriate motor noise levels failed to consistently differentiate limbs or showed high variability in learning success.

Significance and Claims
The authors claim that their model provides a "first attempt at a detailed computational model" of the mobile paradigm that integrates action-outcome prediction with biologically inspired motor control. The significance of the work lies in its demonstration that predictive mechanisms (specifically action-outcome prediction and novelty-driven exploration) are essential for sensorimotor contingency learning, rather than simple reinforcement or operant conditioning alone.

The paper modestly concludes that while the model replicates the main empirical findings, it does not capture the full variability of human infants (both between and within subjects), likely due to the exclusion of attentional, social, and environmental distractions. However, the ablation studies strongly suggest that the components included in the model—prediction, exploration, motor noise, and redundant motor control—are likely necessary ingredients in the actual developmental mechanisms of infants. The model serves as a starting point for further investigation into sensorimotor learning mechanisms and their application to robotics.