Representation in natural and artificial agents: an embodied cognitive science perspective

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Robotics

Background:

• Traditional approaches to representation often overlook the physical embodiment of agents.
• Category learning is a fundamental cognitive task crucial for understanding representation.

Purpose of the Study:

• To present an embodied cognitive science perspective on representation formation.
• To highlight the role of embodiment in generating learnable data for agents.
• To reframe the challenges of representation acquisition and learning.

Main Methods:

• Introduced novel perspectives: frame-of-reference and complete agent.
• Elaborated on embodiment, distinguishing dynamic and information-theoretic aspects.
• Analyzed implications for representation, focusing on sensorimotor integration and data generation.

Main Results:

• Embodiment is a key prerequisite for representation, enabling manipulation of sensory input to generate invariances.
• Agents actively generate data through interaction, transforming complex learning problems into simpler ones.
• Representation is intrinsically linked to agent-environment interactions, not an abstract concept.

Conclusions:

• Representation acquisition is facilitated by embodiment, which shapes sensory data for learnability.
• The embodied view shifts focus from improving learning algorithms to optimizing data generation through interaction.
• Understanding representation requires studying agents within their environmental and physical contexts.