A Bayesian Need for Simplicity: A study of how the brain selects and implements internal models, using a normative approach and Bayesian data analysis methods

A prominent hypothesis in cognitive neuroscience is that the brain operates as an inference engine, relying on internal models to interpret sensory input and guide decision-making. While a great variety of behaviors and biases are explained through this approach, it is less clear how the brain builds and, most importantly, chooses among different models to interpret its environment. A general bias toward simpler interpretations is well-established, but critical gaps remain regarding the computational mechanisms driving this preference. Specifically, it is unclear whether the brain assesses the simplicity (or complexity) of an interpretation through optimal, theory-aligned computations or if it relies on heuristic approximation that can be skewed by perceptual biases. Furthermore, we don't know if the measure of complexity is fixed or adapted to the detailed features of the models at hand. The first part of this work examines the criteria humans use to select between alternative interpretations of noisy data. First, investigating the effect of a variable amount of data, we show that human interpretations, while qualitatively following the principle of parsimony, diverge from optimal behavior. Our results suggest that intuitive model selection is based on a perceptual evaluation of the sample size, rather than a principled computation of the model fit. Second, we address which properties of the model affect the perception of its complexity. We show that the brain does not rely on a generic complexity metric, as some model selection criteria do. Instead, characteristics unique to the models considered have a direct effect on the choice between interpretations. Broadening the scope from intuitive model selection to biological implementation, the latter half of this work investigates how internal models shape sensory perception and how they are instantiated in neural populations, leveraging the same probabilistic frameworks used in the first half. In the domain of sensory processing, we apply Bayesian inference to explain perceptual biases in rats. We show that the influence of task-irrelevant sounds on visual discrimination is best explained by an internal model affected by a compressive warping of visual representations caused by auditory inputs, providing behavioral evidence for direct sensory interaction. Finally, we develop a novel Bayesian method for analyzing neural population activity. Applied to data from mice performing a value-based decision-making task, this technique allows for the decoding of latent variables required to construct and update an internal model of reward contingency. Taken together, these findings offer a multi-level perspective on internal models, from cognition to perception to neural implementation, shedding light on the algorithms the brain uses to manage uncertainty and providing methodological tools to investigate them further.