Diffusion-Based Generation of Neural Activity from Disentangled Latent Codes.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Advances in neural recording technology allow simultaneous monitoring of thousands of neurons.
• Latent variable models are crucial for simplifying complex neural data into interpretable representations.

Purpose of the Study:

• To develop an unsupervised approach for inferring disentangled behavioral variables from neural activity.
• To leverage conditional generative modeling for enhanced neural data analysis.

Main Methods:

• Proposed a novel approach, Generating Neural Observations Conditioned on Codes with High Information (GNOCCHI), based on InfoDiffusion.
• Augmented diffusion models with latent variables to capture data variations.
• Applied GNOCCHI to synthetic and biological time-series neural data from reaching movements.

Main Results:

• GNOCCHI learned higher-quality, more structured, and disentangled latent spaces compared to VAE-based models.
• The model accurately generated novel samples representing unseen behavioral conditions.
• Demonstrated effective unsupervised discovery of interpretable latent spaces from neural data.

Conclusions:

• Unsupervised, information-based models like GNOCCHI hold significant potential for neural data analysis.
• GNOCCHI facilitates the generation of high-quality synthetic neural data for unseen conditions.
• This approach enhances the interpretability of neural recordings and aids in understanding brain function.