The algorithmic and neural logic of perceptual decision-making

Abstract

Decision-making based on noisy sensory stimuli is a fundamental part of everyday life. In this thesis, we present three studies that seek to advance the understanding of the algorithmic and neural logic underlying decision-making by reckoning with challenges posed by large-scale datasets and the complexity revealed by modern methodological techniques. First, we consider the challenge posed by nonstationarities in a key decision variable - the criterion used to pick between two options - that is of particular concern in large longitudinal datasets. We characterize the pitfalls of a method proposed to accommodate for such nonstationarity and recommend a model-based alternative. Second, we identify a unified mechanism that gives rise to two widely observed departures from the behavior of the optimal model - trial history biases and lapses. While these suboptimalities have traditionally been considered distinct, we demonstrate that normative decision-making under a misbelief about nonstationarity in the world gives rise to both history dependence and choices that appear to be evidence-independent or lapses. We test our model in choices of a large dataset of rats, and choices and reaction times of a novel reaction time task, and show that the constraints posited by the model are obeyed in these datasets. Finally, we investigate the neural logic underlying perceptual decision-making. Decision-making is often conceptualized as a sequence of two sub-computations - gradual accumulation of evidence followed by thresholding to commit to a choice. In rats, the anterior dorsal striatum (ADS) and the frontal orienting fields (FOF) have been mapped onto these two theoretically defined computations. This provides an intriguing - yet untested - neural implementation of the decision process in which ADS and FOF form a functional feedforward hierarchy. We present results from simultaneous neural recordings and projection-specific inactivations that challenge this previously proposed mapping. Our results show that both ADS and FOF carry redundant task-relevant information, are involved throughout the accumulation process, and yet respond differentially to perturbations. We reconcile these conflicting observations from physiology and perturbations using a multi-region recurrent neural network model.