Structuring the World: Statistical Learning From the Top and From the Bottom.

Statistical learning is the brain's primary strategy for transforming continuous, noisy, and ambiguous sensory inputs into structured internal representations. This process is studied through a diverse array of perspectives spanning different species, sensory systems, levels of analysis, and methodological frameworks. Here, we present four such perspectives. Two of them approach the problem from the bottom, examining the neural substrates of the biological implementation. The first line describes implicit contextual learning in the rodent auditory midbrain, showing how neuronal frequency tuning adapts to predictable statistical regularities over days. The second line shows that explicit knowledge modulates belief updating as early as the human auditory midbrain and thalamus, revealing a functionally specialized subcortical inferential hierarchy. The other two research lines approach it from the top, examining the computations the brain performs. The third line compares statistical computations across species and human development, uncovering fundamental differences in probabilistic chunking strategies between honeybees, human infants, and adults. The fourth line provides a computational framework for continual learning in which experience is organized into structured internal models through ongoing contextual inference. We then examine what each line can contribute to the others. We argue that, if the brain evolved statistical learning under pressure to behave optimally, this optimality must have been achieved globally through the constraints and synergies that emerge from the interplay between subsystems and levels of analysis. The neuronal mechanisms implementing a computation are shaped by the computation they perform, while that computation is itself constrained by the available physiology: a mutual dependence that can only be understood by bridging across perspectives.