Sparse Component Analysis: An interpretable dimensionality reduction tool that identifies building blocks of neural computation

Behavior results from neural computations that occur at the level of populations of neurons, and understanding how a given computation produces a particular behavior requires a characterization of how distinct signals interact and evolve. For example, in motor cortex, discrete reaches are generated by a combination of ‘preparatory’ and ‘execution-related’ signals. While the number of computationally-relevant signals is much lower than the number of neurons in a population – and is therefore amenable to dimensionality-reduction methods – standard techniques, like principal components analysis, find entangled representations in which individual dimensions reflect a mixture of distinct processes. This ‘mixing’ often impedes our ability to interpret and characterize a given computation. Here, we present sparse component analysis (SCA), a dimensionality reduction method to find interpretable, temporally sparse, low-dimensional representations. On datasets recorded from monkey motor cortex and C. elegans, we demonstrate how SCA helps to reveal the building blocks of neural computation across behaviors.