Learning spatio-temporal statistics from the environment in recurrent networks

项目摘要 学习新任务和接触新环境会导致大脑回路的动态变化， 在最近的各种实验中观察到。将环境统计数据嵌入大脑的能力 电路对于动物在不断变化的环境中茁壮成长和生存的能力至关重要。但 电路动力学的实现和学习机制还没有得到很好的理解， 重大理论挑战。最近在理论和实验室的工作都强调了 电路动态的重要性。然而，在大多数理论模型中，网络连接要么不是可塑的，要么 遵守生物学上难以置信的学习规则。在这里，我们将发展一个理论，大脑回路如何学习他们的大脑回路。 从环境的统计数据中获得动力学。我们将在一系列实验中锚这项工作，以便 使其具有生物学上的现实性并限制其范围。在aim 1中，我们将尝试理解网络是如何学习的 刺激奖励时空统计学这一目标将基于电路级实验，显示如何 神经元动力学的变化是由于刺激后的延迟奖励，以及细胞实验， 阐明强化学习的机制。这是我们比较了解的一个问题，也是 本质上比以无监督的方式学习环境的统计数据更简单。在目标2中， 专注于皮层回路学习时空的顺序而不是时间的实验。 顺序在这样的网络中，学习序列的定时由内在网络动态确定; 使得该问题比学习序列的顺序和定时两者更简单。在目标3中，我们开发了 网络和学习规则，可以学习时空序列的顺序和时序。这 努力将建立在目标2的结果之上，在目标2中，以无监督的方式学习事件的顺序， 目标1，其中使用强化学习来学习事件的定时。

Project Summary Abstract Learning new tasks and exposure to new environments lead to changes in the dynamics of brain circuits, as observed in various recent experiments. The ability to embed the statistics of the environment within brain circuits is essential for animals ability to thrive and survive in changing environments. However, the mechanisms by which circuits dynamics are implemented and learned are not well understood, and pose significant theoretical challenges. Recent work in both theoretical and experimental labs has highlighted the importance of circuit dynamics. Yet in most theoretical models the network connectivity is either not plastic, or obeys biologically implausible learning rules. Here we will develop a theory of how brain circuits can learn their dynamics from the statistics of the environment. We will anchor this work in a set of experiments, in order to make it biologically realistic and limited in scope. In aim 1 we will try to understand how networks can learn stimulus-reward spatiotemporal statistics. This aim will be based on circuit level experiments that show how neuronal dynamics change due to a stimulus followed by a delayed reward, and by cellular experiments that shed light on the mechanisms of reinforcement learning. This is a problem we know more about, and it is also inherently simpler than learning the statistics of the environment in an unsupervised manner. In aim 2 we will concentrate on experiments on which cortical circuits learn the order, but not the timing, of a spatiotemporal sequence. In such networks the timing of the learned sequence are determined by intrinsic network dynamics; making this problem simpler than learning both the order and the timing of a sequence. In aim 3 we develop networks and learning rules that can learn both the order and the timing of a spatiotemporal sequence. This effort will build on results in aim 2 in which the order of events is learned in an unsupervised manner, and of aim 1 in which the timing of events is learned using reinforcement learning.