Feasible cognition requires latent state

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Control Theory

Background:

• Intelligent behavior operates under inherent uncertainty, including noisy sensory feedback, imprecise actions, and ambiguous causes.
• When multiple hidden states produce similar observations, behavior-based regulation, learning, and evaluation become ill-defined.

Purpose of the Study:

• To propose that inference over latent states is a fundamental requirement for cognition under partial observability.
• To unify diverse psychological and control-theoretic literatures under a common framework of regulation under uncertainty.

Main Methods:

• Conceptual analysis integrating control theory with cognitive science.
• Framing observable signals as noisy evidence for hidden causes (e.g., motor state, intentions).

Main Results:

• Latent state inference provides a feasibility condition for cognition, enabling regulation, learning, and evaluation.
• This perspective explains robustness, context sensitivity, and credit assignment in various cognitive domains.
• Cognitive properties arise from regulating inferred states, not just observable behavior.

Conclusions:

• Latent state representations are essential for well-defined behavior, learning, and evaluation under uncertainty.
• Cognition fundamentally depends on estimating task-relevant hidden states that are detectable and actionable from observations.