22-BBSRC/NSF-BIO - Interpretable & Noise-robust Machine Learning for Neurophysiology

系统神经科学的主要目标之一是提供将感觉与感知和认知联系起来的网络级计算的机械性描述。通过分析单个神经元的编码特性，可以很好地理解感觉加工的早期阶段。但在支配感知和认知的较高阶段，关键计算出现在大型神经元群体的相互作用中。从实验和建模的角度描述这些相互作用都是一个挑战：实验必须能够以单个神经元的分辨率提供大规模的神经活动测量，并且模型必须能够在生物学上看似合理的框架内准确地捕捉单个神经元的行为和它们的大规模相互作用。幸运的是，最近在实验和计算方法方面的进展最终使迎接这一挑战成为可能。在这项提案中，我们概述了开发一套工具的工作计划，这些工具将使系统神经科学家能够在跨越多个空间尺度的分层框架内建立将感觉与感知和认知联系起来的模型。我们的工具将使神经科学家能够适应灵活的模型，这些模型可以直接从神经记录(包括颅内和非侵入性)执行关键的感知和认知任务。这一能力将通过提供一个系统假设生成和测试的平台，极大地促进对潜在系统的研究，并将允许在广泛的其他应用中对感知和认知进行类似于大脑的模拟。我们将使用超维计算(HDC)来开发一个从生物学上可信且可解释的建模框架，称为HDNeuro，它利用神经符号表示来提供对感知和认知的分层解释。HDNeuro模型由两个主要阶段组成：在编码阶段，HDNeuro模型通过模拟早期感觉通路的解剖学和生理学的尖峰神经元来转换数据；在认知阶段，HDNeuro通过HDC表示和模仿高级大脑动力学的算法来建立神经符号模型。重要的是，HDNeuro模型的认知阶段将被限制在生物学上可信的计算，允许在机械水平上解释模型现象。HDNeuro模型使用动态神经元来保持实验数据的内在结构，以在单个神经元水平上复制时空神经活动。然后，HDNeuro模式使用多维抽象操作来自然地记忆、联想和组合神经表征，同时保留认知任务所需的信息。大规模的空间和长期时间信息在一个分层网络中表示，该网络使用符号推理和分布式表示，根据需要进行组合，以建立与感知和认知功能的连接。

One of the primary goals of systems neuroscience is to provide mechanistic descriptions of the network-level computations that link sensation to perception and cognition. The early stages of sensory processing can be well understood through analysis of the encoding properties of individual neurons. But in the higher stages that govern perception and cognition, key computations emerge from interactions across large neuronal populations. Describing these interactions presents a challenge from both an experimental and a modelling perspective: experiments must be able to provide measures of neural activity on a large scale with single neuron resolution, and models must be able to accurately capture the behavior of both single neurons and their large-scale interactions within a biologically-plausible framework.Fortunately, recent advances in both experimental and computational methods have finally made it possible to meet this challenge. In this proposal, we outline a program of work to develop a set of tools that will enable systems neuroscientists to build models that link sensation to perception and cognition within a hierarchical framework that spans multiple spatial scales. Our tools will allow neuroscientists to fit flexible models that can perform key perceptual and cognitive tasks directly from neural recordings (both intracranial and non-invasive). This capability will greatly advance the study of the underlying systems by providing a platform for systematic hypothesis generation and testing, and will also allow for brain-like simulation of perception and cognition in wide range of other applications.We will use Hyperdimensional Computing (HDC) to develop a biologically-plausible and interpretable modelling framework, called HDNeuro, that leverages neuro-symbolic representation to provide a hierarchical explanation of perception and cognition. HDNeuro models are composed of two main stages: in the encoding stage, HDNeuro models transform data through spiking neurons that emulate the anatomy and physiology of early sensory pathways; in the cognitive stage, HDNeuro establishes neuro-symbolic models through HDC representations and algorithms that emulate higher-level brain dynamics. Importantly, the cognitive stage of HDNeuro models will be constrained to biologically-plausible computations, allowing for interpretation of the model phenomena at a mechanistic level.HDNeuro models use dynamic neurons that maintain the intrinsic structure of experimental data to replicate spatial-temporal neural activity at the single neuron level. HDNeuro modes then employ hyperdimensional abstract operations to naturally memorize, associate, and combine neural representations while preserving the information required for cognitive tasks. Large-scale spatial and long-term temporal information are represented within a hierarchical network that uses symbolic reasoning with distributed representations that are combined as needed to establish connections with perceptual and cognitive functions.