Event:

Modern recording methods in neuroscience allow the collection of increasingly large datasets, from acquiring data of thousands of neurons over long timescales, to behavioral recordings of animals which span weeks, months, or even years. In these recordings, we want to infer, probe and understand their "hidden causes", or latent variables and their dynamics. These dynamics can have different underlying structure, and unroll on different time scales. Up to date, there is no unifying statistical inference framework for non-linear system identification under these constraints.

To this end, my work has focused on developing new machine learning tools for inferring latent structure from observable data. In my talk, I will highlight recent and ongoing work on developing variants of identifiable contrastive learning algorithms suitable for neuroscientific inference, such as CEBRA. I will discuss these algorithms’ abilities to uncover consistent and robust neural latent dynamics for both calcium and electrophysiology datasets, across sensory and motor tasks, and in simple or complex behaviors across species. This algorithmic framework allows to test hypotheses on the structure of the underlying latents and to analyze their relation to the involved neurons and their activity.