Inference of Network Dynamics and Architecture in Neural Systems with Data-Driven Methods

神经系统在执行各种连贯任务（如感觉处理、信息传递和信息存储）方面的能力是显著的。由于神经元回路的复杂性，理解神经元回路如何相互作用形成连贯的网络是具有挑战性的。在这个项目中，昆虫的嗅觉神经系统被研究为脊椎动物和人类感觉神经回路的模型，因为它将环境线索编码为连贯的感知对象，并且可以在无脊椎动物行为时获得电生理记录。该项目将开发数学工具，使用来自嗅觉处理单元的多神经记录，在几个月内对鲁棒模式进行分类，并推断出预测计算模型。计算模型将有助于揭示这些神经回路网络建立的设计原则。该模型生成的复杂和动态嗅觉刺激的模拟将用于指导电生理和行为实验，以量化嗅觉网络如何响应、分类和识别这些刺激。在这个项目中，电生理记录将被用来推断昆虫的主要嗅觉处理中心——触角叶的动态网络。为此，将获得一个大型神经记录数据集，并开发扩展最先进的数据简化和结构推理方法的数学工具。这些工具将被设计为最佳地表示多维时间序列数据并推断网络结构。特别是，研究将侧重于从数据构建决策空间，由每个代表刺激的种群向量（神经编码）跨越，其中气味可以被分类和识别。通过与每个神经元关联一个动态模型，在气味空间中产生类似动态的网络连接将通过优化来推断。通过预测模型和刺激分类方法，该项目将解决有关网络设计的问题，例如为什么编码动态即使对于嘈杂的刺激也表现出鲁棒性，或者网络结构，特别是抑制性连接和反馈如何产生神经活动的鲁棒模式。研究的结果将确定参数的最优值和引起动力学变化的参数。此外，模型的输出和电生理记录都将用于预测对动态输入、背景气味和混合气味的反应。

Neural systems are remarkable in their ability to perform a diversity of coherent tasks such as sensory processing, information transfer, and information storage. Understanding how neuronal circuits interact to form coherent networks is challenging due to their complexity. In this project, the olfactory neural system in insects is studied as a model for vertebrate and human sensory neural circuits because it encodes environmental cues into coherent perceptual objects and electrophysiological recordings can be obtained while the invertebrate animal is behaving. This project will develop mathematical tools to use multineural recordings from the olfactory processing unit in moths to classify robust patterns and to infer a predictive computational model. The computational model will help to reveal the design principles upon which these neuronal circuit networks are built. Simulations generated by the model for complex and dynamic olfactory stimuli will be used to guide electrophysiological and behavioral experiments to quantify how the olfactory network responds, classifies, and recognizes these stimuli.In this project electrophysiological recordings will be used to infer a dynamical network of the antennal lobe, the primary olfactory processing center in insects. For that purpose, a large dataset of neural recordings will be obtained, and mathematical tools that extend state-of-the-art data reduction and structure inference methods will be developed. These tools will be designed to optimally represent the multidimensional time-series data and infer the network structure. In particular, the research will be focused on constructing a decision space from data, spanned by population vectors (neural codes) each representing a stimulus, in which scents can be classified and recognized. By associating with each neuron a dynamical model, the network wiring that produces similar dynamics in the odor space will be inferred using optimization. With the predictive model and stimuli classification methods, this project will resolve questions regarding network design, such as why do the encoding dynamics appear to be robust even for noisy stimuli or how does the network structure, particularly inhibitory connections and feedback, produce robust patterns of neural activity. Optimal values of parameters and those that cause alteration of dynamics will be identified as a result of the study. Additionally, both the output of the model and electrophysiological recordings will be used to predict responses to dynamic inputs, background odor, and mixtures of odors.