A computational model of infant sensorimotor exploration in the mobile paradigm

TL;DR

This paper tackles how infants learn sensorimotor contingencies in the mobile paradigm by introducing a predictive neural-network model that integrates action-outcome prediction, exploration via an activity-interest mechanism, motor noise, and biologically-inspired motor control across many muscle commands. The model reproduces key infant findings: faster, limb-specific activation of the connected limb in contingent conditions, higher activity in contingent versus non-contingent groups, and occasional extinction bursts, with stronger effects in binary than non-binary setups. Ablation analyses reveal that prediction, exploration, motor noise, and a rich muscle-command repertoire are essential to capturing the observed behavior, suggesting these components underpin early sensorimotor learning. The work provides a mechanistic bridge between developmental psychology findings and computational modeling, with implications for understanding learning and informing robotics that rely on intrinsic motivation and prediction-based exploration.

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.

Figures (16)

• Figure 1: Architecture of our model. The architecture shows each component of our model. The red arrows show the aspects of the model that have a direct influence on the network’s weights. The green arrow depicts the flow of the new limb activations. The purple arrow shows the flow of the new sensory feedback prediction and the blue arrows show the flow of the input data of the neural network. We separate the model into four components, on the top left is the neural network, on the bottom left is the deviation from baseline module, on the bottom right is the exploration module and on the the top right is the sensory feedback module.
• Figure 2: Plots of example distributions of the used beta function. (Upper) Muscle commands unordered by weight, color-coded separately for each limb (see legend). (Lower) An example of muscle commands ordered by weight for the left arm.
• Figure 3: An example process of three timesteps and corresponding changes in the activity interest map of limb 1. The activity is divided into 10 ranges: 0 to 0.1, 0.1 to 0.2 and so forth. Each range can have a different amount of interest from 0 to 1. At Timestep 1, the activity range 0.4-0.5 has an interest value of 0.6 that is the highest value (shown on green background). The model produces an activity in this range, 0.46 and the sensory feedback is unsurprising. Therefore, the interest value of the range 0.4-0.5 is reduced from 0.6 to 0.5 in Timestep 2. Now, in Timestep 2 the highest interest value among all activity ranges is 0.5 and there are two ranges with that value, 0.4-0.5 and 0.3-0.4. When there is more than one activity range with the highest interest value, one of them is randomly chosen. In this example that is the range 0.3-0.4 which is highlighted in green. The model produces an activity in this range, 0.39, and is surprised by the sensory feedback. Due to being surprised the interest in this range is raised to 1 in Timestep 3.
• Figure 4: Comparison of limb activity in simulations and infants (mean activity per limb and per 10-s bin). Activity in simulations is expressed in arbitrary units and activity in infants is expressed in gravitation acceleration units. The thick curves show the mean activities across individual data, of the connected limb (red) and the unconnected limb(s) (gray). Thinner pale curves show individual data (individual simulation runs or infants). (A) Model in the Binary condition (20 simulation runs). (B) Infants in the Binary condition (data on 20 six-month-old infants in Popescu et al. popescu_6-month-old_2021). (C) Model in the Non-Binary condition (20 simulation runs). (D) Infants in the Non-Binary condition (data on 18 six-month-old infants in Jacquey et al. jacquey_popescu_2020; note that this experiment lasted only 4 minutes and that the data on the attention-getter corresponding to the first bin of every minute was removed).
• Figure 5: Evolution of the activity interest map for each limb, example of a run in the non-binary condition. For each limb, the figure shows how the values of interest for each activity range evolve over time. The X-axis shows the time and the Y-axis shows the activity ranges. The color of each (x,y) cell shows the interest of that activity range at that specific time. The blue-yellow spectrum represents the scalar value corresponding to the interest, blue color corresponds to lowest interest values, greenish color--to intermediary and yellow color--to highest interest values.