Current experimental techniques allow the simultaneous recording of hundreds of neurons, but extracting the behaviorally relevant and easily interpretable variables represented by the data is challenging.
We are using a recently developped machine learning technique, variational autoencoders, to identify the low-dimensional manifold in hippocampal population activity that accurately captures variability in the data and provides the interpretable latent factors.
The hippocampus is an ideal testbed for this analysis for at least two reasons:
- we know that the spatial location is amongst the few behaviorally relevant and easily measurable factors driving the population activity
- individual neurons have a highly non-linear tuning to spatial location.
Therefore, if we can accurately reconstruct the position of the animal from the population activity in an unsupervised way, we may hope that we can also identify other, yet unknown factors represented in the neuronal data.
For more details, see this conference abstract.