Linking macroscale structure and function in brain-like recurrent neural networks

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Linking brain structure to function is crucial in neurobiology and AI.
• Artificial neural networks (ANNs) lack comprehensive macroscale structural constraints, hindering interpretation.
• It's unclear if brain structure-function principles apply to ANNs or if constraints yield cortical findings.

Purpose of the Study:

• To introduce BrainRNN, an ANN architecture inspired by human cortical structure.
• To investigate if brain-like structural constraints can induce functional organization in ANNs.
• To explore structure-function coupling within a structurally constrained ANN.

Main Methods:

• Developed BrainRNN, a recurrent neural network architecture.
• Incorporated macroscale human cortical structural constraints into the model design.
• Analyzed connectivity distribution, unit activation, and emergent functional organization.

Main Results:

• BrainRNN selectively regulates connectivity distribution under structural constraints.
• Increased activated association units in BrainRNN correlate with higher-order cognitive capacity.
• Emergent macroscale functional organization, including modules and gradients, mirrors human cortical findings.

Conclusions:

• Brain-like structural constraints shape functional organization in artificial neural networks.
• Function can be discovered from structure in structurally grounded AI.
• This approach holds potential for advancing neuroscientific research through AI models.