Stevenson Ian H
Department of Psychological Sciences, University of Connecticut, Storrs, CT, USA.
Department of Biomedical Engineering, University of Connecticut, Storrs, CT, USA.
J Comput Neurosci. 2016 Aug;41(1):29-43. doi: 10.1007/s10827-016-0603-y. Epub 2016 Mar 23.
A key observation in systems neuroscience is that neural responses vary, even in controlled settings where stimuli are held constant. Many statistical models assume that trial-to-trial spike count variability is Poisson, but there is considerable evidence that neurons can be substantially more or less variable than Poisson depending on the stimuli, attentional state, and brain area. Here we examine a set of spike count models based on the Conway-Maxwell-Poisson (COM-Poisson) distribution that can flexibly account for both over- and under-dispersion in spike count data. We illustrate applications of this noise model for Bayesian estimation of tuning curves and peri-stimulus time histograms. We find that COM-Poisson models with group/observation-level dispersion, where spike count variability is a function of time or stimulus, produce more accurate descriptions of spike counts compared to Poisson models as well as negative-binomial models often used as alternatives. Since dispersion is one determinant of parameter standard errors, COM-Poisson models are also likely to yield more accurate model comparison. More generally, these methods provide a useful, model-based framework for inferring both the mean and variability of neural responses.
系统神经科学中的一个关键观察结果是，即使在刺激保持恒定的受控环境中，神经反应也会有所不同。许多统计模型假设逐次试验的尖峰计数变异性是泊松分布的，但有大量证据表明，根据刺激、注意力状态和脑区的不同，神经元的变异性可能比泊松分布大得多或小得多。在这里，我们研究了一组基于康威-麦克斯韦-泊松（COM-泊松）分布的尖峰计数模型，该模型可以灵活地解释尖峰计数数据中的过度离散和不足离散。我们说明了这种噪声模型在贝叶斯估计调谐曲线和刺激周围时间直方图中的应用。我们发现，具有组/观察水平离散度的COM-泊松模型（其中尖峰计数变异性是时间或刺激的函数）与泊松模型以及通常用作替代方案的负二项式模型相比，能更准确地描述尖峰计数。由于离散度是参数标准误差的一个决定因素，COM-泊松模型也可能产生更准确的模型比较。更一般地说，这些方法为推断神经反应的均值和变异性提供了一个有用的、基于模型的框架。
