Machine Learning of Spatiotemporal Bursting Behavior in Developing Neural Networks

As with other modern sciences (and their computational counterparts), neuroscience experiments can now produce data that, in terms of both quantity and complexity, challenge our interpretative abilities. It is relatively common to be faced with datasets containing many millions of neural spikes collected from tens of thousands of neurons. Traditional data analysis methods can, in a relatively straightforward manner, identify large-scale features in such data (e.g., on the scale of entire networks). What these approaches often cannot do is to connect macroscopic activity to the relevant small-scale behaviors of individual cells, especially in the face of ongoing background activity that is not relevant. This communication presents an application of machine learning techniques to bridge the gap between microscopic and macroscopic behaviors and identify the small-scale activity that leads to large-scale behavior, reducing data complexity to a level that can be amenable to further analysis. A small number of spatiotemporal spikes (among many millions) were found to provide reliable information about if and where a burst will occur.