Towards a Neural Foundation Model: A Probabilistic Spike Representation Model to Mitigate Neural Variability

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Spike train data are crucial for modeling neural dynamics but exhibit variability, hindering model generalization.
• Existing attention-based models struggle with neural stochasticity and electrode misalignment.
• A need exists for robust spike representation models that capture unified neural manifolds.

Purpose of the Study:

• To develop a novel framework, the Probabilistic Neural Representation Transformer (PNRT), for modeling variable spike train data.
• To address limitations of deterministic models in capturing neural stochasticity and electrode misalignment.
• To improve generalization of neural models across sessions and subjects.

Main Methods:

• Proposed the Probabilistic Neural Representation Transformer (PNRT) framework.
• Modeled variable spike activities into a consistent latent probabilistic distribution.
• Implemented an activity-based neuronal reordering method to decouple from physical electrode positions.

Main Results:

• PNRT demonstrated superior performance over deterministic baselines on three cross-subject datasets.
• The framework effectively models unified neural representations.
• PNRT showed robustness across significant spike variability.

Conclusions:

• PNRT successfully models unified neural representations from variable spike train data.
• The proposed methods enhance robustness and generalization in neural modeling.
• PNRT offers a promising approach for analyzing large-scale neural recordings.