Abstract
We present a computational model of the mechanisms that may determine infants' behavior in the "mobile paradigm". This paradigm has been used in developmental psychology to explore how infants learn the sensory effects of their actions. In this paradigm, a mobile (an articulated and movable object hanging above an infant's crib) is connected to one of the infant's limbs, prompting the infant to preferentially move that "connected" limb. This ability to detect a "sensorimotor contingency" is considered to be a foundational cognitive ability in development. To understand how infants learn sensorimotor contingencies, we built a model that attempts to replicate infant behavior. Our model incorporates a neural network, action-outcome prediction, exploration, motor noise, preferred activity level, and biologically-inspired motor control. We find that simulations with our model replicate the classic findings in the literature showing preferential movement of the connected limb. An interesting observation is that the model sometimes exhibits a burst of movement after the mobile is disconnected, casting light on a similar occasional finding in infants. In addition to these general findings, the simulations also replicate data from two recent more detailed studies using a connection with the mobile that was either gradual or all-or-none. A series of ablation studies further shows that the inclusion of mechanisms of action-outcome prediction, exploration, motor noise, and biologically-inspired motor control was essential for the model to correctly replicate infant behavior. This suggests that these components are also involved in infants' sensorimotor learning.
Introduction
In developmental psychology, the "mobile paradigm" is a classic experiment: a ribbon connects an infant’s limb to a mobile hanging above their crib. Infants quickly learn that moving that specific limb makes the mobile dance. While this seems simple, it represents a foundational cognitive milestone: detecting sensorimotor contingencies.
Despite its importance, the underlying mechanisms remain poorly understood. Previous computational models were often too "minimalist"—using simple state machines that lacked internal representations or predictive capabilities. These simpler models couldn't explain certain nuances, such as why infants sometimes show a sudden "burst" of activity (the extinction burst) the moment the mobile is disconnected. This research, conducted by a team from the University of Hamburg and the Czech Technical University, introduces a sophisticated model that bridges the gap between developmental psychology and computational neuroscience.
Methods and Architecture
The researchers developed a neural network-based agent designed to replicate how infants learn to control their limbs. The model integrates several key components to simulate complex infant behaviors:
- Biologically Inspired Motor Control: Instead of four simple "limb" outputs, the model uses 600 "muscle commands" that map to four limbs in an overlapping, redundant fashion. This mimics the complexity of a real human body where multiple muscles contribute to a single movement.
- Action-Outcome Prediction: The model predicts the sensory feedback (mobile movement) its actions will cause.
- Exploration Module (Curiosity): The model maintains an "Activity Interest Map." If an action results in a "surprise" (a high prediction error), the interest in that action increases. If the outcome is predictable, interest wanes.
- Multi-Part Loss Function: The network is trained by minimizing three distinct losses: Prediction Error (expected vs. actual feedback), Exploration Error (current vs. "most interesting" activity), and Baseline Deviation (keeping activity within a realistic "fussy" baseline).
Curiosity-Driven Exploration
A central feature of the model is its ability to explore based on "surprise." The Exploration Module tracks which limb activities lead to novel outcomes. When the mobile is connected, moving the connected limb creates a prediction error (surprise) because the model does not yet expect the mobile to move. This surprise increases the "interest" value for that specific activity range in the Activity Interest Map, driving the agent to repeat and refine those movements.
Key Results and Insights
The model successfully replicated several key empirical findings from infant studies:
- Limb Specificity: The model learned to move the "connected" limb significantly more than the three "unconnected" limbs, matching the timing and magnitude seen in 6-month-old infants.
- Contingency Sensitivity: In "yoked" control tests (where the mobile moves randomly), the model correctly showed no limb-specific increase in activity, proving it was learning the causal link, not just reacting to visual stimulation.
- The Extinction Burst: Because the model relies on prediction, the sudden removal of the mobile connection causes a massive spike in "surprise." This drives the exploration module to aggressively re-test the limb, creating the "extinction burst" seen in real infants—a behavior previous models couldn't replicate.
- Ablation Studies: Removing any core component—the 600 muscle commands, the motor noise, or the exploration mechanism—significantly degraded performance, suggesting all these elements are necessary for infant-like learning.
Discussion and Future Work
The study demonstrates that prediction is the engine of learning. Without the prediction module, the model fails to differentiate between limbs. Furthermore, the researchers found a "Goldilocks" zone for motor noise: too little noise prevents the model from discovering the connection, while too much makes the system too chaotic to learn.
While successful, the model has limitations. Real infants are subject to attention shifts, social factors, and physical constraints like gravity. Future work could incorporate a more robust habituation mechanism or a full physics engine to further increase the fidelity of the simulation.
Summary updated