Analyzing the Computational Complexity of a Dynamic Model of the Time-varying Spatiotemporal Sensitivity of Neurons in the Visual Cortex

In order to precisely track the dynamics of visual information encoding during saccadic eye movements, we have recently developed a computational model applicable to the sparse, limited neural data available during the short timescale of saccades. The time-varying stimulus kernels estimated by the model can provide a precise description of the neurons' dynamic spatiotemporal sensitivity with high precision during a saccade. Achieving this precision however requires handling a trade-off between the computational complexity, generalizability, and prediction accuracy of the encoding model. In this paper, we derive a principled strategy for the model parameterization to handle this trade-off, which is particularly critical for estimating the high-dimensional models using sparse neuronal data. The results of this analysis can guide optimizing various dimensionality reduction and model parameterization approaches used for the estimation of the generalized linear model-based frameworks applied on spiking data beyond the data and approaches used in this study.