Predictive Computational Models of Olfactory Networks

摘要（项目5） 知觉对象的本质和导致它们在大脑中再现的神经机制是 神经科学的基本问题之一。感知对象的表征是否填充了空间 或者它们反映了刺激空间的复杂性？刺激的特征是什么 由感知空间的维度所代表的在大脑中呈现的物体是如何 不变性相对于刺激特征，时间和背景的变化？ 尽管我们对嗅觉的分子基础的理解取得了实质性的进展， 嗅觉系统，这些问题仍然没有答案。例如，在眼睛中， 视锥光感受器的类型对应于由人类色觉感知的三维。 了解颜色空间的低维本质是我们理解颜色的基础 视野在嗅觉系统中，缺少类似的概念性理解。 这个项目是协同努力的一部分，以了解嗅觉编码的性质。基于 由同一U19项目的其他项目收集的实验数据集，以及 数据集，我们将研究嗅觉刺激的空间结构，嗅觉神经元的反应， 感知质量，建立一个神经网络模型，建立空间之间的连接，并解决 让这个网络在生物学上变得现实。使用最先进的机器学习 方法，我们将产生一个嗅觉系统的预测计算模型作为交付。 我们的目标是开发，在计算模型中实现，并测试至少两个理论想法， 嗅觉密码的本质首先，我们将测试假设，嗅觉空间包含实质上较少的 比气味受体（OR）的类型的数量。我们的初步数据表明， 主维度的数量可以低至10，而OR类型的主维度数量为103。我们将定义这些 数学上的维度，并将它们与气味剂的分子性质联系起来。第二，我们将测试 首因编码假设，根据该身份的一小群最敏感的嗅觉 受体类型以浓度不变的方式代表气味剂身份。这种表述使 气味物体对噪音具有鲁棒性。我们的计算/理论研究将在密切合作下进行， 实验组。我们的项目包括三个具体目标：目标1：建立预测计算 嗅觉刺激，反应和感知空间的模型;目标2：开发预测网络模型 嗅觉空间之间的映射;和目标3：建立嗅觉网络的生物现实模型。 由于感官对象的表征是神经科学中的一个基本问题， 我们的研究所揭示的将阐明在其他感觉形式的感觉表征的原则。

SUMMARY (Project 5) The nature of perceptual objects and of the neuronal mechanisms leading to their representation in the brain is one of the fundamental questions in neuroscience. Do representations of perceptual objects populate spaces of low dimensionality or do they mirror the complexity of the stimulus space? What features of the stimulus are represented by the dimensions of the perceptual space? How can objects represented in the brain retain invariance with respect to variations in stimulus features, timing, and background? Despite substantial progress in our understanding of the molecular basis of the sense of smell, for the olfactory system, these questions remain unanswered. In the eye, for example, the responses of the three types of cone photoreceptors correspond to the three dimensions sensed by human color vision. Understanding the low dimensional nature of color space was fundamental to our understanding of color vision. In the olfactory system, a similar conceptual understanding is missing. This project is a part of synergistic effort to understand the nature of olfactory coding. Based on experimental datasets collected by other projects of the same U19 program as well as publically available datasets, we will study the structure of the spaces of olfactory stimuli, responses of olfactory neurons, and perceptual qualities, build a neural network model that establishes connections between spaces, and resolve conceptual questions to make this network biologically realistic. Using state-of-the-art machine learning approaches, we will generate a predictive computational model of the olfactory system as a deliverable. Our goal is to develop, implement in a computational model, and test at least two theoretical ideas about the nature of olfactory code. First, we will test the hypothesis that olfactory spaces contain substantially fewer dimensions than the number of types of odorant receptors (OR). Our preliminary data indicates that the number of principal dimensions may be as low at 10, compared to ~103 of OR types. We will define these dimensions mathematically and relate them to the molecular properties of odorants. Second, we will test the primacy coding hypothesis, according to which identities of a small cohort of the most sensitive olfactory receptor types represent odorant identity in a concentration-invariant manner. Such representations render odor objects robust to noise. Our computational/theoretical studies will be carried out in close collaboration with the experimental groups. Our project includes three Specific Aims: Aim 1: To build predictive computational models for spaces of olfactory stimuli, responses, and percepts; Aim 2: To develop a predictive network model for mapping between olfactory spaces; and Aim 3: To build biologically realistic models of olfactory networks. Since representation of sensory objects is a fundamental problem in neuroscience, mathematical principles uncovered by our studies will elucidate the principles of sensory representation in other sensory modalities.