End-to-end neural system identification with neural information flow.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• System identification in neuroscience aims to understand how brain regions process information.
• Existing methods often require invasive techniques or struggle with complex, multi-region neural computations.
• Developing noninvasive methods to model neural information flow is crucial for advancing brain research.

Purpose of the Study:

• To introduce Neural Information Flow (NIF) as a novel, end-to-end trainable system identification framework for neuroscience.
• To model neural computations across multiple brain regions using coupled tensors.
• To validate the NIF model's ability to recover meaningful neural representations from noninvasive data.

Main Methods:

• Developed a NIF model representing neural processing as a network of coupled tensors.
• Trained the model end-to-end using stochastic gradient descent on noninvasive fMRI data.
• Utilized low-rank observation models to link tensor elements (cortical columns) to brain region activity, decomposing into spatial, temporal, and feature receptive fields.

Main Results:

• Successfully trained a NIF model on fMRI data from early visual areas.
• Recovered plausible visual representations consistent with known cortical organization.
• Identified population receptive fields that align with empirical findings in visual neuroscience.

Conclusions:

• NIF offers a powerful, noninvasive approach for system identification in neuroscience.
• The model effectively captures neural information processing and sensory representations.
• NIF demonstrates potential for advancing our understanding of brain function from large-scale neuroimaging datasets.