Conference on Cognitive Computational Neuroscience

Variational autoencoders (VAE) have received significant attention from the deep learning and neuroscience communities as the VAE is good at finding meaningful low-dimensional latent representations from high-dimensional data. Although the VAE was originally developed as an unsupervised learning algorithm, recent advances in the VAE allow us to use it in a semi-supervised manner (e.g., conditional VAE). This new model allows neuroscientists to condition the VAE on their experimental variables and identify the condition-free low dimensional representations. Here, we trained conditional VAEs to model brain responses to different levels of noxious heat stimulation with a training fMRI pain dataset (n = 87). By conditioning the data on different levels of heat intensity, we extracted the condition-free low-dimensional latent variables, with which we were able to generate brain responses for any given conditions (i.e., heat intensity) from no-pain data in an independent test dataset (n = 37). Further analyses revealed that the condition-free latent variables can identify different individuals with high accuracy, suggesting that the latent variables contain each individual’s idiosyncratic features. Overall, we show that the conditional VAE can model the effects of heat intensity as well as each individual’s idiosyncratic features, providing a powerful analysis strategy both for population-level and personalized pain neuroimaging.