REAL-TIME VARIATIONAL METHOD FOR LEARNING NEURAL TRAJECTORY AND ITS DYNAMICS.

Area of Science:

• Computational Neuroscience
• Machine Learning

Background:

• Latent variable models are crucial for understanding neural computation.
• Offline algorithms for neural trajectory extraction are well-developed.
• Real-time alternatives for immediate feedback and enhanced experimental design are less explored.

Approach:

• Introduced the exponential family variational Kalman filter (eVKF), an online recursive Bayesian method.
• eVKF infers latent trajectories and learns the underlying dynamical system simultaneously.
• Utilizes the constant base measure exponential family for latent state stochasticity and works with arbitrary likelihoods.

Key Points:

• Derived a closed-form variational analogue to the Kalman filter's predict step.
• Achieved a provably tighter bound on the Evidence Lower Bound (ELBO) compared to other online variational methods.
• Validated eVKF on both synthetic and real-world neural data.

Conclusions:

• eVKF demonstrates competitive performance in inferring latent neural dynamics.
• The method provides a powerful tool for real-time analysis in computational neuroscience.
• Facilitates immediate feedback and improved experimental design in neural recording studies.