An Information-Theoretic Framework for Understanding Learning and Choice Under Uncertainty

Although information theory is widely used in neuroscience, its application has primarily been limited to the analysis of neural activity, with much less emphasis on behavioral data. This is despite the fact that the discrete nature of behavioral variables in many experimental settings-such as choice and reward outcomes-makes them particularly well-suited to information-theoretic analysis. In this study, we provide a framework for how behavioral metrics based on conditional entropy and mutual information can be used to infer an agent's decision-making and learning strategies under uncertainty. Using simulated reinforcement-learning models as ground truth, we illustrate how information-theoretic metrics can reveal the underlying learning and choice mechanisms. Specifically, we show that these metrics can uncover (1) a positivity bias, reflected in higher learning rates for rewarded compared to unrewarded outcomes; (2) gradual, history-dependent changes in the learning rates indicative of metaplasticity; (3) adjustments in choice strategies driven by reward harvest rate; and (4) the presence of alternative learning strategies and their interaction. Overall, our study highlights how information theory can leverage the discrete, trial-by-trial structure of many cognitive tasks, with the added advantage of being parameter-free as opposed to more traditional methods such as logistic regression. Information theory thus offers a versatile framework for investigating neural and computational mechanisms of learning and choice under uncertainty-with potential for further extension.