Conceptual complexity and the bias/variance tradeoff.

Area of Science:

• Cognitive Science
• Machine Learning
• Psychology

Background:

• Concept learning models are typically categorized as either exemplar-based or prototype-based.
• These models represent opposite ends of the bias/variance tradeoff spectrum in statistical learning.
• Prototype models exhibit high bias and low variance, while exemplar models show low bias and high variance.

Purpose of the Study:

• To investigate human learners' position on the bias/variance continuum for concept learning.
• To compare the performance of exemplar, prototype, and a regularized model in explaining human category formation.
• To determine if human concept learning aligns with extreme models or an intermediate strategy.

Main Methods:

• Human participants were presented with category structures of varying complexity, from simple to multimodal.
• Behavioral data from participants learning these categories were collected and analyzed.
• A novel computational model incorporating hypothesis complexity regularization was developed and tested.

Main Results:

• Human learners did not consistently adopt either a high-bias (prototype) or low-bias (exemplar) strategy.
• Learners positioned themselves at an intermediate point on the bias/variance spectrum.
• The proposed regularized model demonstrated a better fit to human experimental data than traditional exemplar or prototype models.

Conclusions:

• Human concept learning is not adequately explained by purely exemplar-based or prototype-based models.
• Learners employ a flexible strategy that balances model complexity with data adherence, akin to regularization.
• This intermediate bias-variance approach offers a more accurate account of human category acquisition.