Extracting Preserved Neural Latent Dynamics Across Tasks using Convolutional Transformer-based Variational Autoendecoder.

Abstract

Understanding how neural systems drive behavior is a fundamental goal in neuroscience. Numerous studies have demonstrated that the activity of large neural populations is often governed by low-dimensional neural dynamics. While much of the current research has focused on extracting informative and interpretable latent dynamics from individual motor tasks, it remains unclear whether these dynamics are preserved across different motor tasks. This question is particularly critical, as prior experience with a related task can facilitate faster learning in a new task. In this paper, we propose a Convolutional Transformer-based Variational Autoencoder (Conformer-VAE) to extract preserved neural latent dynamics across tasks by leveraging the rich spatiotemporal patterns in neural activity. We validate our approach using neural recordings from a rat, which first performed a one-lever pressing task (old task) and subsequently a two-lever discrimination task (new task). By projecting the inferred latent dynamics from both tasks onto a common 2D PCA plane, our results demonstrate that Conformer-VAE effectively captures preserved neural dynamics across tasks, outperforming baseline methods. Moreover, these preserved dynamics enable faster decoder training for the new task by transferring the neural-to-movement mapping learned from the old task. This capability facilitates seamless real-time task switching, offering promising applications for brain-machine interface systems.Clinical Relevance-This work facilitates faster adaptation in brain-machine interfaces by preserving neural dynamics across tasks, offering potential benefits for neuroprosthetics and motor rehabilitation in patients with motor impairments.