A latent variable approach to decoding neural population activity

Decoding is a powerful approach for measuring the information contained in the activity of neural populations. As a result, decoding analyses are now used across a wide range of model organisms and experimental paradigms. However, typical analyses employ general purpose decoding algorithms that do not explicitly take advantage of the structure of neural variability, which is often low-dimensional and can thus be effectively characterized using latent variables. Here we propose a new decoding framework that exploits the low-dimensional structure of neural population variability by removing correlated variability that is unrelated to the decoded variable, then decoding the resulting denoised activity. We demonstrate the efficacy of this framework using simulated data, where the true upper bounds for decoding performance are known. A linear version of our decoder provides an estimator for the decoded variable that can be more efficient than other commonly used linear estimators such as linear discriminant analysis. In addition, our proposed decoding framework admits a simple extension to nonlinear decoding that compares favorably to standard feed-forward neural networks. By explicitly modeling shared population variability, the success of the resulting linear and nonlinear decoders also offers a new perspective on the relationship between shared variability and information contained in large neural populations.