Machine learning for analysis and interpretation of neuronal population activity

The aim of the project currently is to gain useful insight into data sets comprised of multi dimensional neuronal spiking trains using deep generative models. By training neural network based models such as variational autoencoders on neural spiking data, we can produce models, which can not only reproduce neural spiking data, but also yield insights about the information encoded in the population activity. The aim of this work will be to develop approaches for interpretative modelling of neural activity, with potential applications in prosthetics.

In parallel, we will also investigate if neural networks can explain computations performed by the brain. It is currently unclear if the same principles are implemented in biological neurones, but the way neural networks encode data such as images is very similar to that found in the brain.. Spiking neural networks which mirror aspects of brain function will be explored with the aim of reproducing or even outperforming the current deep learning paradigm. To this end, different methods of training these spiking based models will be explored, as in theory these models are more expressive for a given size than their scalar value based counterparts, but training them in an effective way is currently an issue.