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

All behaviors, even simple actions, result from the interplay of multiple neural computations, and understanding how a given behavior is generated requires delineating these computations from one another. Traditionally, identifying these critical ‘computational building blocks’ requires a researcher to impose structure on the data: specific task epochs are analyzed separately or specific population-level structure is sought using supervised dimensionality reduction methods. While these approaches can be successful, they have obvious drawbacks – many behaviors cannot be easily parsed into distinct epochs, and it is often not clear, particularly in early, exploratory phases of data analyses, what type of structure one should look for in neural data. Unsupervised methods, like principal component analysis (PCA) identify large neural signals, yet often do not provide the most interpretable low-dimensional view of the data.
Here, we present sparse component analysis (SCA), an unsupervised dimensionality reduction method that produces interpretable, low-dimensional representations across a broad range of neural datasets. SCA incorporates three empirically validated assumptions of neural data. First, population-level activity is low dimensional. Second, distinct neural computations tend to occur in orthogonal or nearly-orthogonal subspaces. Finally, and most critical for the present results, distinct computations are dissociable in time from one another. We applied SCA to four diverse neural datasets: motor cortex activity recorded from a reaching monkey, motor cortex activity from monkeys performing a unimanual and bimanual cycling task, and C. elegans interacting in an open field. In all cases, SCA found not only the structure previously reported using more supervised methods, but also found previously unreported structure, such as posture-related signals distinct from movement-related signals and discrete stopping signals related to the end of a cycling movement. SCA provides an unsupervised method for identifying interpretable latent factors and, in conjunction with supervised methods, provides a powerful tool for characterizing neural computations.