Identification of Invariant Sensorimotor Structures as a Prerequisite for the Discovery of Objects

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Robotics

Background:

• Object perception is crucial for intelligent agents.
• Identifying spatio-temporally invariant structures is a fundamental mechanism for object perception.
• Sensorimotor Contingencies Theory provides a framework for understanding agent-environment interaction.

Purpose of the Study:

• To develop a computational model for identifying spatio-temporally invariant structures from sensorimotor experience.
• To investigate an unsupervised, predictive approach inspired by Sensorimotor Contingencies Theory.
• To demonstrate how agents can build internal models of invariant environmental structures.

Main Methods:

• Utilizing a sensorimotor, unsupervised, and predictive computational model.
• Processing unsupervised interactions of an artificial agent with its environment.
• Representing invariant structures as densely connected subgraphs in a sensory state graph.
• Employing Spectral Clustering on the sensorimotor experience graph.

Main Results:

• Demonstrated that invariant environmental structures induce regularities in sensorimotor experience.
• Showcased the agent's ability to capture these regularities as subgraphs.
• Successfully built an internal model of moving, invariant structures using Spectral Clustering.
• Highlighted the model's distinctiveness from supervised methods using static images.

Conclusions:

• The proposed model effectively captures spatio-temporally invariant structures through unsupervised sensorimotor learning.
• This approach offers a fundamental mechanism for object perception in artificial agents.
• The method provides insights into building internal world models from continuous interaction, contrasting with static image-based recognition.