Marked point process variational autoencoder with applications to unsorted spiking activities

Area of Science:

• Computational Neuroscience
• Machine Learning
• Data Science

Background:

• Spike train modeling in large neural populations is crucial for understanding neural coding.
• Marked point processes are used for neural population modeling, generating events with waveform features as marks.
• Existing models struggle with high-dimensional and nonlinear observations in joint mark intensity estimation.

Purpose of the Study:

• To develop a new joint mark intensity model for unsorted spike trains.
• To address limitations of existing models in capturing high-dimensional and nonlinear data.
• To improve the understanding of how covariates influence unsorted spikes.

Main Methods:

• Proposed a novel joint mark intensity model utilizing a variational autoencoder.
• Defined the joint mark intensity as a latent variable model with a neural network decoder.
• Derived a new log-likelihood lower bound for parameter estimation using variational bounds.
• Integrated the model into a nonlinear state space model for hidden state dynamics capture.

Main Results:

• The model effectively represents the dependency structure of unsorted spikes on covariates in a data-driven manner.
• Achieved superior performance in prediction and neural decoding tasks.
• Demonstrated effectiveness on synthetic data and real place cell spiking activity.

Conclusions:

• The proposed variational autoencoder-based model advances spike train analysis for large neural populations.
• The method offers a powerful tool for neural decoding and understanding neural coding.
• This approach enhances the ability to model complex neural data without prior spike sorting.