Prediction of psychophysical responses from spike recordings in rat sensorimotor cortex by using Bayesian models

In this thesis, we studied the fundamental question in neuroscience: how perception is built based on the sensory stimuli from the physical world and turned into motor actions in the face of uncertain neural representations. The vast body of literature contains models using neural activity to decode stimulus parameters, motor responses, and behavioral patterns. In particular, this line of research became more important as sensorimotor neuroprostheses and brain-computer interfaces (BCI) were made possible by recent advances in technology. The real-time algorithms used in those applications have many limitations. The main goal of the thesis is to use Bayesian models to understand sensorimotor processing and develop a novel approach for future BCIs. Specifically, spike data were collected from awake behaving rats during psychophysical yes/no detection task. Within a Bayesian framework, task-related priors, posterior beliefs, and the objective function to match the observed choice of the animal were calculated. The random variables for stimulus presentation, population neural activity, and motor responses were combined in a probabilistic graph network. First, a somatosensory neuroprosthesis application is demonstrated. Next, the Bayesian model was used to predict trial-by-trial responses offline. It was found that psychophysically low-performing rats could be modelled better with the Bayesian approach. The simulation results were compared to predictions of other supervised learning algorithms (such as linear discriminant analysis, decision trees, etc.). The Bayesian prediction was one of best among those algorithms for low-performing rats. Finally, behavioral responses from previous trials and neural activity from the current trial were included in various Bayesian models, which studied the effects of incremental information to predict the behavioral response in the current trial. The results showed that the average firing rates in a population of neurons are mostly adequate to predict lever presses in the psychophysical task with high sensitivity and low bias. This thesis provides new insights into computational modeling to understand sensorimotor processing and development of future BCIs. Bayesian modeling can be particularly useful in rehabilitation and during the training period of neuroprostheses.