Towards re-unifying the theoretical perspectives on conditioning, category learning, probability learning and economic decisions from experience.

Abstract

Stimulus classification is an everyday feat (e.g., in medical diagnoses by differentiating ultrasound images). Category feedback, however, is often non-deterministic (e.g., by 25% chance untrue a.k.a. probabilistic feedback) rendering experiences as somewhat unreliable. In probability learning and economic decisions (when humans try to predict which of two outcomes is more rewarding), it is often observed that humans decide non-rational (probability matching, Gamblers Fallacy). Despite shared origins, the research areas of category learning, probability learning, conditioning and economic decisions (experience based) did not arrive at a consensus of what drives such probability matching strategies. Here, we offer a domain-general integrative model that predicts the mixture of behavioral trends when participants learn probabilistic stimulus-outcome regularities. We use the Category Abstraction Learning framework (CAL), implementing the hypothesis, that humans count the streak of events, and generate simple rules and conditional hypotheses with them. We present simulations of four studies, one from each domain, showing that CAL’s learning mechanisms accurately predict systematic and individual differences in category (correlation) learning, reward learning, risky gambles, and fear conditioning, concerning the phenomena of Gamblers Fallacy, positive and negative recency, or win-stay-lose-shift strategies, in a single model. One central novel CAL hypothesis to link learning phenomena under gains, losses and neutral feedback, is that gains and losses differently affect cognitive control during learning (corrective feedback processing), thereby also providing a novel perspective on risk-preferences in experience-based risky gambles. We discuss CAL’s potential as a domain-general account of human learning in experience-based decisions in light of a broader range of theories from multiple domains.