Towards sparse coding of natural movements for neuroprosthetics and brain-machine interfaces

The correlation structure of natural hand & finger movements suggests that their motion is controlled in a lower-dimensional space than would be possible given their mechanical nature. Yet, it is unclear whether this low dimensional embedding is relevant to how the brain represents motor actions and how we can decode it for Brain-Machine Interface applications. We collected large data set of natural hand movement kinematics and analysed it using a novel sparse coding and dictionary learning approach - Sparse Movement Decomposition (SMD), which captures the embedding of the data in terms of spatial and temporal structure. We show that our sparse codes over natural movement statistics give a more parsimonious representation than the simple correlation structure. This suggest that, like V1 neuron receptive fields can be predicted from sparse code over natural image statistics, motor control may be encoded in such a manner. We further show how our sparse coding can help understand the temporal structure of behaviour, and thus our technique may be used for behavioural fingerprinting in diagnostics and for more naturalistic neuroprosthetic control.