Abstract
The tuning properties of neurons are often variable, and can be modulated by internal and external factors. Various schemes have been proposed to capture such variations, e.g., gain changes, baseline shifts, tuning shifts and tuning width changes, all of which have profound coding implications. However, neural systems may exhibit a mix of these different effects. Furthermore, neural data in the form of spikes require averaging over trials to obtain the tuning curves. These issues can complicate the interpretation of experimental data. To better characterize the variations of the neural tuning, we developed a new technique based on functional principal component analysis (fPCA). Specifically, we augmented the standard fPCA method by incorporating a Poisson observation noise model. Our new method, Poisson fPCA, takes the spike count data as the inputs, and generates a set of basis functions as outputs, providing a compact summary of the tuning variations. Importantly, our method can capture latent variations of tuning for a pre-specified stimulus dimension, making it different from pure regression-based methods such as generalized linear model (GLM), or latent-based methods such as PCA and its nonlinear extensions. We apply this method to previous published datasets collected from primary visual cortex (V1) of anesthetized primates. While most neurons exhibit gain fluctuations as previously reported, we also found that gain change is not a constant, and is highly correlated with the orientation tuning curve for many neurons. Furthermore, the fluctuations of the whole neural population collectively exhibit low dimensional structure, as well as specific spatial structure. In sum, we developed a new method to extract the variations of the neural tuning. While here we only use it to analyze the variability of orientation tuning in V1, our method is general and could be applied to other brain areas or sensory modalities as well.
Acknowledgement: NSF NeuroNex Award DBI-1707398, The Gatsby Charitable Foundation