Human inference reflects a normative balance of complexity and accuracy

Area of Science:

• Cognitive Science
• Decision Making
• Machine Learning Theory

Background:

• Humans often infer hidden world properties from imperfect sensory data.
• Complex probabilistic methods (e.g., Bayesian inference) are accurate but cognitively taxing.
• Simple heuristics are less demanding but potentially less accurate.

Purpose of the Study:

• To model the trade-off between cognitive strategy complexity and inferential accuracy.
• To determine the optimal balance between strategy complexity and accuracy under varying environmental uncertainty.
• To investigate how working memory and adaptivity are modulated by uncertainty.

Main Methods:

• Developed a computational model of a hierarchy of strategies with varying complexity.
• Analyzed the relationship between strategy complexity and accuracy using a power law.
• Conducted two psychophysical experiments to test model predictions on human behavior.

Main Results:

• A power law of diminishing returns was observed: increased complexity yielded progressively smaller accuracy gains.
• The benefit of complex strategies was contingent on environmental statistical uncertainty.
• Complex strategies offered no substantial advantage when uncertainty was very high or very low; a 'complexity dividend' existed at intermediate uncertainty levels.

Conclusions:

• Cognitive strategy complexity and accuracy exhibit diminishing returns, governed by environmental uncertainty.
• Human working memory and adaptivity are modulated by uncertainty, aligning with model predictions.
• Optimal cognitive performance involves selecting strategy complexity matched to environmental statistical properties.