Towards a "universal translator" for neural dynamics at single-cell, single-spike resolution

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Current neuroscience understanding of the brain is fragmented.
• Reading neural activity from arbitrary brain regions remains a challenge.

Purpose of the Study:

• To develop a foundation model for neural spiking data.
• To enable automated readout of encoded information in neural activity.

Main Methods:

• Introduced a novel self-supervised modeling approach: multi-task-masking (MtM).
• The MtM model alternates between masking and reconstructing neural activity across time, neurons, and brain regions.
• Evaluated using the International Brain Laboratory dataset with Neuropixels recordings from 48 animals.

Main Results:

• MtM significantly improved performance over state-of-the-art population models.
• Enabled effective multitask learning for diverse prediction tasks (single-neuron, region-level, forward prediction, behavior decoding).
• Training on multiple animals enhanced model generalization to unseen subjects.

Conclusions:

• The MtM approach provides a foundation for a comprehensive brain model.
• This work advances the goal of understanding brain function at single-cell, single-spike resolution.
• The developed model paves the way for a generalized foundation model of the brain.