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

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Unsupervised neural decoding requires dissociating internal neural pattern reactivation from external variable tuning.
• The Poisson Gaussian-Process Latent Variable Model (P-GPLVM) discovers low-dimensional structures in high-dimensional spike trains but lacks latent trajectory inference for new data.
• Estimating neural reactivation is limited by the inability to infer latent trajectories in previously learned latent spaces.

Purpose of the Study:

• To extend the P-GPLVM for latent variable inference in novel neural data, enabling unsupervised neural decoding and reactivation analysis.
• To develop principled methods for constrained latent variable inference in temporally compressed neural activity, such as population burst events.
• To introduce metrics for assessing neural pattern reactivation validity and inferring encoded experiences.

Main Methods:

• Extended the P-GPLVM to infer latent trajectories of new neural data using learned smoothness and mapping information.
• Developed constrained latent variable inference for temporally compressed neural activity (e.g., hippocampal sharp-wave ripples).
• Applied the extended framework to hippocampal ensemble recordings during maze exploration and running, analyzing neural pattern reactivation.

Main Results:

• Replicated P-GPLVM's ability to learn a latent space encoding animal position from hippocampal recordings.
• Demonstrated that the learned latent space can differentiate between maze contexts.
• Observed reactivation of neural patterns during running, correlating with experience similarity in the training data manifold.
• Successfully estimated neural pattern reactivation during population burst events, identifying replay events.

Conclusions:

• The extended P-GPLVM framework provides a powerful tool for unsupervised analysis of neural activity and decoding.
• The method enables robust inference of neural reactivation, even for compressed activity patterns like sharp-wave ripples.
• This approach facilitates answering critical questions in neuroscience regarding neural representations and memory replay.