2D Capsule Networks Detect Perceived Changes in Infant∼Environment Relationship Reflected in 3D Movement Dynamics

Area of Science:

• Developmental Psychology
• Robotics
• Machine Learning

Background:

• Understanding infant exploration and causal discovery is crucial for developmental psychology.
• Purposeful action emerges in early life, but detecting its nuances is challenging.
• Artificial Intelligence (AI) offers novel methods for analyzing complex behavioral data.

Approach:

• Infants' functional connection to an object was manipulated (tethered foot to a mobile).
• Vicon motion capture data generated Histograms of Joint Displacements (HJDs) for 3D infant trajectories.
• Various machine and deep learning models (kNN, LDA, FCNet, 1D/2D-Conv, 1D/2D-CapsNet) classified experimental states using HJDs.

Key Points:

• Deep learning models, especially 2D-Capsule Networks (2D-CapsNet), excelled with full-body features.
• Simpler models like k-Nearest Neighbour (kNN) and Linear Discriminant Analysis (LDA) performed well with single joint features.
• Foot activity demonstrated the most significant and consistent pattern alterations across experimental stages.

Conclusions:

• AI, particularly deep learning, can effectively detect and analyze infant exploration and causal discovery.
• Foot movement analysis proved highly indicative of behavioral changes related to environmental interaction.
• Integrating AI with experimental paradigms offers a promising avenue for developing clinically relevant infant behavioral assessments.