Identifying Interpretable Latent Factors with Sparse Component Analysis.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Understanding neural computations requires identifying shared latent factors within neural populations.
• Current methods often rely on supervision, which can be limiting when the structure of these factors is unknown.
• Identifying distinct computational roles of neural signals is crucial for linking neural activity to behavior.

Approach:

• Introduced Sparse Component Analysis (SCA), an unsupervised machine learning technique.
• SCA identifies latent factors by enforcing sparsity in time and orthogonality.
• Applied SCA to diverse datasets including primate motor cortex, C. elegans neural activity, and artificial neural networks.

Key Points:

• SCA successfully identified interpretable latent factors without prior supervision.
• The method demonstrated clear parcellations of neural activity across different behaviors.
• Identified factors were effective in describing network computations in all tested systems.

Conclusions:

• Sparse Component Analysis offers a powerful unsupervised approach for dissecting neural computations.
• This method facilitates the discovery of meaningful latent structures in complex neural data.
• SCA advances our ability to understand the relationship between neural activity and behavior.