CONSTRUCTING BIOLOGICALLY CONSTRAINED RNNS VIA DALE'S BACKPROP AND TOPOLOGICALLY-INFORMED PRUNING.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Systems Neuroscience

Background:

• Conventional recurrent neural networks (RNNs) lack physiological and anatomical fidelity for modeling cortical function.
• Key biological constraints, Dale's law and structured connectivity, are often omitted in RNN models.
• Omitting these constraints raises questions about the validity of insights from RNNs studying neuronal interactions.

Purpose of the Study:

• To develop methods for training RNNs that incorporate Dale's law and sparse connectivity.
• To provide mathematical grounding and empirical validation for these biologically-constrained RNNs.
• To demonstrate the utility of these methods for inferring multi-regional neural interactions.

Main Methods:

• Developed novel methods to train RNNs respecting Dale's law and specific sparse connectivity patterns.
• Provided mathematical guarantees for the proposed constrained RNN approaches.
• Trained RNN models on 2-photon calcium imaging data from mice performing visual tasks.

Main Results:

• Empirically demonstrated that constrained RNNs match the performance of unconstrained RNNs.
• Successfully inferred multi-regional interactions in the mouse cortical network.
• Enforced data-driven, cell-type specific connectivity constraints across cortical layers and brain areas.

Conclusions:

• The developed methods enable biologically plausible RNNs that maintain high performance.
• Inferred neural interactions align with experimental findings and the theory of predictive coding.
• These biologically-informed RNNs offer a valid approach for studying complex neural systems.