The representation of value in the prefrontal cortex during novel choices

决策需要了解和学习选择的价值。在非人灵长类动物（NHP）的前额叶皮层（PFC）中，编码决策变量值属性的神经元已被广泛报道。然而，迄今为止，大多数证据都源于动物执行过度训练的任务，这些任务很难再现自然主义的决策场景，通常涉及新的选择。这个项目的中心目标是研究PFC的基本编码原则，在新的或灵活的选择价值。最近的数据表明，“分布式”价值代码类似于海马体中众所周知的空间计算，神经元编码以“网格代码”方式连接的价值空间中的特定“位置”。这一点，再加上PFC的概率推理和抽象规则编码的能力，可以支持在新的上下文中的值的推理。此外，我们将通过比较分布式编码方案和概率群体编码来探索值的概率表示，概率群体编码依赖于每个神经元群体都调整到给定变量的一个特定值来编码变量的整个概率分布。可以应用于PFC的不同值和预测误差信号的一个这样的示例是分布式强化学习，其提出神经元学习期望奖励概率分布的不同部分。我们将通过从PFC，海马和内嗅皮层进行高密度记录来探索这些价值编码框架，而NHP在大概率值空间中进行过度训练和新颖的选择。

Decision-making requires knowing and learning the value of choices. Neurons encoding value attributes of decision variables have been widely reported in the prefrontal cortex (PFC) of non-human primates (NHPs). However, most evidence to date has stemmed from animals performing overtrained tasks which poorly reproduce naturalistic decision-making scenarios often involving novel options. The central aim of this project will be to investigate PFC's basic coding principles for value during novel or flexible choices. Recent data points to a 'distributed' value code akin to the well-understood spatial computations in the hippocampus, with neurons encoding specific 'places' in value space linked in a 'grid-code' fashion. This, together with PFC's capacity for probabilistic inference and encoding of abstract rules, may support the inference of value in novel contexts. Additionally, we will explore probabilistic representations of value by drawing comparisons between distributed coding schemes and probabilistic population codes, which rely on populations of neurons each tuned to one specific value of a given variable to encode the variable's entire probability distribution. One such example which may apply to PFC's diverse value and prediction error signals is distributional reinforcement learning, which proposes neurons learn to expect different parts of a reward probability distribution. We will probe these value-coding frameworks by performing high density recordings from the PFC, hippocampus and entorhinal cortex while NHPs make overtrained and novel choices in large probabilistic-value spaces.