Developing novel neural network tools for accurate and interpretable dynamical modeling of neural circuits

摘要 近年来，我们可以同时记录的神经元数量呈指数级增长 增长，提出了一个令人生畏的挑战：我们如何分析这些复杂和高维的 数据集，以深入了解神经电路如何执行计算？来自动力系统的工具 理论已经成功地解开了人工递归神经网络的计算机制 (RNN)经过训练以执行目标导向的任务。如果我们能将这些工具应用于生物神经电路， 它将提供对大脑内部运作的无与伦比的访问，并潜在地允许我们 将神经计算的理论与真实的生物数据联系起来。然而，要让这些工具发挥作用， 我们需要在硅胶中创建其动态忠实地代表底层动态的复制品 生物系统。 到目前为止，生物网络的最好的电子复制是经过训练的RNN，以产生 与记录的神经元放电模式相匹配。虽然这种方法正在迅速流行起来，但它已经 关键缺陷。目前的培训方法并不局限于产生准确的表示 事实上，RNN实际上是因为发明了多余的动力学而得到奖励的，所以 只要这些动力学有助于复制记录的神经数据。此外，这些型号通常 假设潜在活动和神经放电率之间的关系(“嵌入”)是线性的； 这一假设被证明是错误的，动力学准确性受到影响。多余的动力学问题 当试图对交互系统进行建模时，非线性嵌入尤其严重 神经回路。 该方案的目标是开发一种新的人工神经网络结构，该结构 解决上述挑战，并使我们的内置硅胶模型能够捕捉到 在赛道内和跨赛道建造。我的方法结合了两个关键组件：1)神经普通 微分方程式(节点)，我们已经演示学习的计算体系结构 动力学比RNN更准确和紧凑，以及2)可逆神经网络(INN)读出， 这消除了多余的动力学，并允许模型逼近非线性嵌入。这就做 验证了该模型的能力，称为常微分式可逆自动编码器 读出(ODIN)，使用合成神经数据和 之前从猴子身上收集的多电极记录。这个工具将有助于建造一座桥 神经数据与本地和分布式神经计算之间的关系。

Abstract In recent years, the number of neurons that we can record simultaneously has seen an exponential increase, presenting a daunting challenge: how do we analyze these complex and high-dimensional datasets to gain insight into how neural circuits perform computation? Tools from dynamical systems theory have successfully unraveled the computational machinery of artificial recurrent neural networks (RNNs) trained to perform goal-directed tasks. If we could apply these tools to biological neural circuits, it would provide unparalleled access to the inner workings of the brain and potentially allow us to connect theories of neural computation to real biological data. However, for these tools to be useful, we need to create in silico replicas whose dynamics faithfully represent the dynamics of the underlying biological system. To date, the best in silico replicas of biological networks are RNNs trained to produce output that matches recorded patterns of neuronal firing. While this approach is rapidly growing in popularity, it has critical flaws. Current training methodologies are not constrained to produce accurate representations of the underlying dynamics; in fact, RNNs are actually rewarded for inventing superfluous dynamics, so long as those dynamics help to reproduce recorded neural data. Additionally, these models often assume that the relationship (“embedding”) between latent activity and neural firing rates is linear; when this assumption proves false, the dynamical accuracy suffers. The problems of superfluous dynamics and non-linear embedding are especially severe when attempting to model a system of interacting neural circuits. The objective of this proposal is to develop a novel artificial neural network architecture that addresses the above challenges and allows our in-silico models to capture accurate dynamics that are built both within and across-circuits. My approach combines two key components: 1) neural ordinary differential equations (NODEs), a computational architecture that we have demonstrated learns dynamics more accurately and compactly than RNNs and 2) invertible neural network (INN) readouts, which eliminate superfluous dynamics and allow the model to approximate nonlinear embeddings. I will validate the ability of this model, called an Ordinary Differential equation auto-encoder with Invertible readout (ODIN), to find accurate within- and across-circuit dynamics using synthetic neural data and previously-collected multi-electrode recordings from monkeys. This tool will help to build a bridge between neural data and both local and distributed neural computations.