Behaviour, the only approach for living creatures to interact with the environment, is the consequence of sensorimotor transformations in the central nervous system. Successful implementation of behaviour tasks requires high accuracy in sensory perception and motor execution, and more importantly, experience-based plasticities. We seek to investigate neuronal accomplishment of these sensorimotor functions in the brain. Due to technical obstacles, neural recording experiments fail to provide a complete understanding. Thus, we intend to bridge the gap by reverse-engineering the brain with machine learning algorithms as functional and quantitative frameworks. For this purpose, we develop a series of computational models and algorithms, covering different aspects of control and representational learning as sensorimotor transformations.
We propose a spiking neural circuit model of the cerebellum, learning from movement errors to generate motor correction against external perturbation to ensure accurate execution. This allows us to discover sensorimotor error-based learning in a bottom-up manner, bringing insights to the functional structure and disorders of the cerebellum. Moving from sensorimotor control to representational learning, we emphasise more on the statistical structure of natural behaviour, i.e. the prior p(x), rather than individual executions. We assume that sensorimotor representations are shaped by the natural behaviour statistics, and based on this assumption, we introduce machine learning models of topological neural coding and manifold learning. Our topological neural coding models, with neurons arranged in a 2D topological map, are optimised towards the optimal representational power under p(x). Applying topological neural coding to natural arm movement data reproduces features of proprioceptive representations observed in macaque S1. Our manifold learning models aims to explore the hidden structure of p(x) while learning a non-linear low-dimensional manifold embedded in the high-dimensional space. This allows us to achieve analytic density description and auto-clustering of human behaviour, which directly leads to symbolic representations of behaviour and metrics to measure behaviour similarity.