Extended Poisson Gaussian-Process Latent Variable Model for Unsupervised Neural Decoding

Area of Science:

• Computational Neuroscience
• Machine Learning for Neural Data Analysis

Background:

• Unsupervised neural decoding aims to understand internal neural state repetition without external variable tuning.
• The Poisson Gaussian-process latent variable model (P-GPLVM) discovers low-dimensional structures in high-dimensional spike trains but lacks inference for new data.
• Inferring latent trajectories in new neural data is crucial for estimating internal state repetition.

Approach:

• Extended the P-GPLVM to infer latent variables for new neural data using learned smoothness and mapping information.
• Developed a principled approach for constrained latent variable inference in temporally-compressed activity patterns like population burst events (PBEs).
• Introduced metrics to assess the congruence of inferred latent variables with the learned manifold.

Key Points:

• Applied the extended P-GPLVM to hippocampal recordings during maze exploration, confirming latent space encodes animal position.
• Demonstrated the latent space differentiates between maze contexts.
• Inferred latent variables revealed repeating internal neural states during running, correlating with similar experiences.

Conclusions:

• The extended P-GPLVM framework enables unsupervised analysis of neural activity, including PBEs.
• This approach facilitates the estimation of internal neural state repetition for identifying replay events.
• The framework aids in answering critical scientific discovery questions regarding neural dynamics and experience encoding.