Developing and validating computational models of working memory in the human prefrontal cortex

我们感知周围世界的方式是通过感官的透镜进入我们的大脑的连续流。每时每刻，我们只能看到、听到和感受到周围环境的点点滴滴，但我们不断表现出的行为远远超出了简单的反射反应。毫无疑问，记忆是我们长时间绑定信息能力的重要组成部分。人类大脑能够通过几种类型的记忆系统来记忆经验，这些记忆系统根据行为的时间和上下文要求而有所不同。其中一个关键的记忆系统是工作记忆--在短时间内储存和处理信息的能力。我的团队的长期研究目标是明确人类大脑如何在记忆中保持和组织先前的经验，使它们易于快速决策。我的团队使用实验和计算技术相结合的方法来建立人脑记忆中神经计算的机械模型。为了做到这一点，我们专注于1）从人脑反应中开发基准，以量化构建此类模型的进展; 2）开发计算模型，目标是最大化其对神经基准的得分。 这个提议的总体目标是应用我们的建模方法来理解人类前额叶皮层中工作记忆的神经计算。具体来说，1）我们使用神经成像技术来表征各种记忆依赖性任务期间的大脑反应模式。然后，我们使用这些数据从前额叶皮层的神经反应中构建基准; 2）通过对模型架构，学习目标和环境组成进行特定选择，以计算模型的形式形成科学假设。3)我们通过对神经基准测试验证这些假设的正确性。 计算模型在理解生理系统方面变得越来越重要。这些机械模型允许我们将我们对系统的整体知识封装到一个框架中，这个框架可能不是口头表达的。虽然当前的前额叶皮层理论强调了它在跟踪目标和实现目标的方法方面的作用，但它们并没有提供关于这些信息如何在神经元反应群体中表示的机械解释。本申请提出了一个计算框架，用于开发前额叶皮层中工作记忆相关的神经计算的机械模型，并解决了这种模型的结构，发展和生态方面。由此产生的具有已知连接性的机械模型对于理解前额叶皮层中工作记忆的神经计算是非常宝贵的。

We perceive the world around us in the form of a continuous stream that enters our mind through the lens of our senses. At every moment we can only see, hear and feel bits and pieces of our surroundings, yet we constantly demonstrate behaviors that reach far past the simple reflexive reactions. Memory is inarguably an essential component of our ability to bind information over long time spans. The human brain is able to memorize experiences through several types of memory systems that differ depending on the temporal and contextual requirements of the behavior. One of the critical memory systems is working memory - the ability to store and manipulate information over short periods. The long-term goal of research in my group is to specify how the human brain maintains and organizes the previous experiences in memory to make them readily accessible for rapid decision making. My group uses a combination of experimental and computational techniques to build mechanistic models of the neural computations underlying memory in the human brain. To do this, we focus on 1) developing benchmarks from human brains responses to quantify the progress towards building such models; 2) developing computational models with the goal of maximizing their scores against the neural benchmarks. The overall goal of this proposal is to apply our modeling approach towards understanding the neural computations underlying working memory in the human prefrontal cortex. Specifically 1) we use neuroimaging techniques to characterize the brain response patterns during various memory-dependent tasks. We then use this data to construct benchmarks from neural responses in the prefrontal cortex; 2) we form scientific hypotheses in the form of computational models by making specific choices of model architecture, learning objective, and environment composition. 3) we validate the correctness of these hypotheses by validating them against the neural benchmarks. Computational models are becoming increasingly important in understanding physiological systems. These mechanistic models allow us to encapsulate our integral knowledge of a system into a framework that may not be verbally expressive. While current theories of the prefrontal cortex highlight its role in tracking goals and ways to achieve them, they do not offer mechanistic explanations as to how such information may be represented in the population of neuronal responses. This application proposes a computational framework for developing mechanistic models of the neural computations underlying working memory in the prefrontal cortex and tackles the structural, developmental, and ecological aspects of such models. The resulting mechanistic models with known connectivity are invaluable for understanding the neural computations underlying working memory in the prefrontal cortex.