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

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Understanding neural systems' control over behavior is key in neuroscience.
• Neural population activity often exhibits low-dimensional dynamics.
• Preservation of these dynamics across tasks is crucial for learning but remains understudied.

Purpose of the Study:

• To develop a method for extracting preserved neural latent dynamics across different motor tasks.
• To investigate if neural dynamics learned in one task can facilitate learning in a new, related task.
• To enhance brain-machine interface (BMI) adaptability.

Main Methods:

• Proposed a Convolutional Transformer-based Variational Autoencoder (Conformer-VAE).
• Leveraged spatiotemporal patterns in neural activity for dynamic extraction.
• Validated using neural recordings from rats performing sequential one-lever and two-lever tasks.

Main Results:

• Conformer-VAE successfully captured preserved neural dynamics across tasks, outperforming baseline methods.
• Projecting inferred dynamics onto a 2D PCA plane visualized shared patterns.
• Preserved dynamics enabled faster decoder training for the new task via transfer learning.

Conclusions:

• The Conformer-VAE effectively extracts shared neural dynamics across tasks.
• Preserved dynamics facilitate rapid adaptation and task switching in BMIs.
• This approach has significant implications for neuroprosthetics and motor rehabilitation.