Conference on Cognitive Computational Neuroscience

In our lives, we routinely encounter two kinds of quantities: magnitudes (e.g., reward magnitude, size, orientation or another dimension of the stimulus) and probabilities (e.g., stimulus occurrence). While we learn both through experiences, a fundamental difference is that magnitudes can be learned from one instance, whereas probabilities require repeated exposure. This difference becomes particularly significant for learning in dynamic environments where both magnitudes and probabilities can change. While the dynamics of magnitude learning have been well studied, those of probability learning remain relatively unknown, especially on a moment-to-moment basis. Here, we measure them through a new continuous learning paradigm, where subjects continuously update their estimate of a changing magnitude or probability. We show that in response to a change, subjects adapt their learning rate not only for magnitudes, but also for probabilities, with different dynamics as a normative learner. In both cases, two determinants underlie the adaptation of subjects’ learning rate, change-point probability and prior uncertainty, but in markedly different proportions in the two cases, again as predicted by normative theory. Those different determinants may correspond to different neural mechanisms.