Drop, Swap, and Generate: A Self-Supervised Approach for Generating Neural Activity.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Understanding neural processing requires simplified representations of complex neural activity.
• Unsupervised learning is challenging for discovering brain-behavior links without explicit labels.

Purpose of the Study:

• Introduce Swap-VAE, a novel unsupervised approach for learning disentangled representations of neural activity.
• Develop a method to uncover the relationship between neural representations and behavior.

Main Methods:

• Utilized a generative modeling framework (Variational Autoencoder).
• Implemented an instance-specific alignment loss to maximize representational similarity between augmented views of neural data.
• Augmented data by dropping neurons and jittering samples to promote temporal consistency and invariance.

Main Results:

• Successfully learned disentangled representations from neural activity.
• Demonstrated effectiveness on both synthetic datasets and real neural recordings from primate brains.
• Identified latent dimensions within neural data that correlate with behavior.

Conclusions:

• Swap-VAE enables unsupervised learning of meaningful neural representations.
• The method effectively disentangles neural datasets, revealing behavior-relevant dimensions.
• This approach advances the study of brain function and information processing.