Learning identifiable and interpretable latent models of high-dimensional neural activity using pi-VAE
This work addresses the challenge of analyzing complex neural data for neuroscientists, offering a more interpretable method, though it appears incremental as it builds on existing identifiable VAE techniques.
The authors tackled the problem of interpreting and identifying latent models of high-dimensional neural activity by proposing pi-VAE, which integrates latent models with task variables, resulting in improved interpretability and identifiability, and demonstrated better data fitting and novel insights on neurophysiological datasets.
The ability to record activities from hundreds of neurons simultaneously in the brain has placed an increasing demand for developing appropriate statistical techniques to analyze such data. Recently, deep generative models have been proposed to fit neural population responses. While these methods are flexible and expressive, the downside is that they can be difficult to interpret and identify. To address this problem, we propose a method that integrates key ingredients from latent models and traditional neural encoding models. Our method, pi-VAE, is inspired by recent progress on identifiable variational auto-encoder, which we adapt to make appropriate for neuroscience applications. Specifically, we propose to construct latent variable models of neural activity while simultaneously modeling the relation between the latent and task variables (non-neural variables, e.g. sensory, motor, and other externally observable states). The incorporation of task variables results in models that are not only more constrained, but also show qualitative improvements in interpretability and identifiability. We validate pi-VAE using synthetic data, and apply it to analyze neurophysiological datasets from rat hippocampus and macaque motor cortex. We demonstrate that pi-VAE not only fits the data better, but also provides unexpected novel insights into the structure of the neural codes.