Quantitative Tools for Investigating Sensory Systems

描述(申请人提供)：我们对人脑生理学的理解有赖于对感觉刺激和神经反应之间的映射的详细描述。这些映射是复杂的，因为它们涉及非线性变换、动力学、学习和适应、反馈和先验，由于自然或学习的限制。受机器学习领域最新进展的启发，我们将调查一系列算法的使用，这些算法将使我们能够表征这种复杂的映射。这些算法代表了对经典工程系统分析方法的现代改进，这种方法只能应用于感觉处理的最外围阶段。这些新算法利用现代计算机的能力，高效地找到描述动态和非线性映射的最简单的函数集。我们的前两个目标集中于参数和非参数算法的发展，以表征一般的刺激反应映射。参数模型包括自适应增益控制和反馈的显式表达式。非参数模型是使用几种算法从数据中估计出来的，包括最大信息维度、神经网络分析、Lasso回归和核回归。我们的第三个目标是开发各种方法来验证各种模型。我们还建议通过将这些工具整合到STRFPAK中来向广大神经科学界提供这些工具，STRFPAK是一个用于估计感觉神经元感受野的软件包(在上一次获奖期间发布并正在不断改进)。最后，我们建议开发一个数据库，作为社区开发的神经生理数据和分析工具的储存库。该数据库将鼓励数据分析方法的验证和分发。该数据库还将为大脑功能的理论家和模型师提供实验数据。根据这一提议开发的刺激-反应映射算法将为神经生物学家提供以前只有专家才能使用的量化工具。因此，它们将对基础研究产生直接的好处。彻底了解感觉系统的复杂刺激-反应图谱也将对医学的几个领域产生重大好处：评估和诊断疾病状态，如黄斑变性，以及改善神经假体，如助听器。大脑如何代表感觉世界的计算机制在许多方面与大脑如何将运动意图转换为运动动作的计算机制相似。因此，这些算法的应用也可能最终导致用于运动控制和动作的神经假体的改进。

DESCRIPTION (provided by applicant): Our understanding the physiology of human brain relies on the detailed description of the mapping between sensory stimuli and neural responses. These mappings are complex as they involve non-linear transformations, dynamics, learning and adaptation, feedback and priors due to natural or learned constraints. Inspired by recent advances in the field of machine learning, we will investigate the use of a series of algorithms that will allow us to characterize such complex mappings. The algorithms represent modern improvements to the classical engineering systems analysis approach that could only be applied to the most peripheral stages of sensory processing. These new algorithms use the power of modern computers to efficiently find the simplest set of functions that describe dynamical and non-linear mappings. Our first two aims focus on development of parametric and non-parametric algorithms to characterize general stimulus response mappings. The parametric models include explicit formulations of adaptive gain control and feedback. The non-parametric models are estimated from the data using several algorithms, including maximally informative dimensions, neural network analyses, Lasso regression and kernel regression. Our third aim is to develop methods to validate various models. We also propose to make these tools available to the neuroscience community at large by incorporating them into STRFPAK, a software package (released during the previous award period and undergoing continuous improvement) for estimating receptive fields of sensory neurons. Finally, we propose to develop a database that will serve as a repository for neuro-physiological data and analysis tools developed in the community. The database will encourage the validation and distributions of data analysis methods. The database will also provide experimental data to theorist and modelers of brain function. The stimulus-response mapping algorithms developed under this proposal will provide neurobiologists with quantitative tools previously only available to specialists. They therefore will have a direct benefit on basic research. A thorough understanding of the complex stimulus-response mapping of sensory systems will also have significant benefits for several areas of medicine: evaluation and diagnosis of disease states such as macular degeneration, and improvements in neural prosthetics such as hearing aids. The computational mechanisms governing how the brain represents the sensory world are in many respects similar to those governing how the brain translates motor intention to motor action. Therefore, application of these algorithms is also likely to lead to eventual improvements in neural prosthetics for motor control and action.