State-switching models of human brain activity using recurrent neural networks

Thesis

It has been shown that spatiotemporal dynamics of neuronal activity can be well described using state-related behaviour, comprising a discrete set of reoccurring quasi-stable states associated with distinct patterns of spatial and functional connectivity. Most methods of analysis will either assume stationarity of these states, as in ICA; or constrain the dynamics to be Markovian, as in hidden Markov models (HMMs). These tools lack the capability to explicitly model the higher order temporal dependencies that can occur over timescales of various scales.

In this thesis, we introduce a model that combines probabilistic state-space models with recurrent neural networks (RNNs), enabling us to relax the Markovian constraint of HMMs and learn temporal features of the data occurring over longer timescales. The model takes the form of a recurrent state-switching network, which models the uncertainty in time-varying state labels via discrete random variables. We introduce a variational Bayesian framework for computationally efficient inference of the model that also generalises to a variety of time series models.

Using simulations, data taken from the resting state magnetoencephalography (MEG) scans of 55 participants, and data taken from the MEG recordings of a face-viewing task undertaken by 19 participants, we demonstrate that we can reliably infer a set of states that fits the data better than where the Markovian constraint is enforced, however we do not see significantly different temporal behaviour emerging. We additionally demonstrate that unlike the Markovian model, the recurrent model can internally represent the temporal dynamics of the data.

Authors: Skates, A

• DOI: 10.5287/ora-pmy7jzj4q
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: RNN; variational inference; functional connectivity; MEG
• Subjects: Neural networks (Computer science); Time-series analysis; Statistics