Understanding decision criterion learning: From signal detection theory to neural implementation

动物的行为是由刺激和行为产生的结果控制的。通过学习，生物体调整其行为以产生期望的结果（例如，奖励）并避免不期望的结果（例如，惩罚）。然而，在自然环境中，并不总是清楚在遇到特定刺激时应该理想地采取哪种特定行动。此外，现实世界中的刺激-反应-结果偶然事件通常既不是确定性的，也不是静态的，因此行为需要通过经验来重塑，在信号检测理论（SDT）的框架下，这种在刺激不确定性下的学习可以被描述为适应一种灵活的决策准则。在计算层面的解释（“为什么要进行计算？），该理论描述了哪个决策标准是最优的，例如，在最大化奖励数量的意义上。虽然SDT已经成功地用于分析感知决策任务的行为数据几十年，但它几乎没有提到算法层面的过程，即，首先，观察者是如何学习标准的？当环境条件发生变化时，观察者又是如何适应标准的？在以前的工作集中在标准学习的算法，我们已经提出了一个基于收入的标准学习模型，结合SDT和动物学习理论的见解，并进行了实验研究。在这里，我们提出了一系列在大鼠和人类中进行的行为实验，以1）将我们基于奖励的模型与已成功应用于描述人类表现的错误学习帐户进行对比，2）通过观察不同刺激和结果条件下的行为来挑战并进一步完善模型。此外，在大鼠试验的基础上，我们将从算法层面的描述（“哪些变量被表示，它们如何相互抵消？）到实现层面（“计算是如何神经实现的？“）通过将理论建模和行为分析与关键脑结构的光遗传学介导的瞬时失活相结合。

Animal behavior is controlled by stimuli and by action-produced outcomes. Through learning, organisms adapt their behavior to produce desirable outcomes (e.g., rewards) and avoid undesirable outcomes (e.g., punishments). In natural environments, however, it is not always clear which specific action should ideally follow upon encounter with a specific stimulus. Moreover, real-world stimulus-response-outcome contingencies are usually neither deterministic nor static, so behavior needs to be reshaped by experience.In the framework of signal detection theory (SDT), such learning under stimulus uncertainty can be described as adapting a flexible decision criterion. At the computational level of explanation ("why is the computation performed?"), the theory describes which decision criterion is optimal, e.g., in the sense of maximizing the number of rewards. While SDT has successfully been used for decades to analyze behavioral data from perceptual decision-making tasks, it says next to nothing about processes at the algorithmic level, i.e., how is the criterion learned in the first place, and how do observers adapt it when environmental conditions change?In previous work focusing on the algorithms of criterion learning, we have proposed and experimentally scrutinized an income-based criterion learning model that combines insights from SDT and animal learning theory. Here, we propose a series of behavioral experiments in both rats and humans to 1) contrast our reward-based model with error-learning accounts that have successfully been applied to describe human performance, and 2) challenge and further refine the model by observing behavior under different stimulus and outcome conditions. Furthermore, building on pilot experiments in rats, we will move from the algorithmic level of description ("which variables are represented, and how are they set off against each other?") to the implementational level ("how are the computations realized neurally?") by combining theoretical modelling and behavior analysis with optogenetically mediated transient inactivation of key brain structures.