Constructing biologically constrained RNNs via Dale's backpropagation and topologically informed pruning.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Recurrent neural networks (RNNs) are used to model cortical function but lack physiological and anatomical fidelity.
• Conventional RNNs raise questions about the validity of insights into brain mechanisms.
• There is a need for biologically constrained computational models in neuroscience.

Purpose of the Study:

• To develop mathematically grounded methods for incorporating Dale's law and sparse connectivity into RNN training.
• To ensure that biologically constrained RNN models maintain performance comparable to unconstrained models.
• To apply these constrained RNNs for inferring multi-regional brain interactions from neural data.

Main Methods:

• Incorporated Dale's law (neuronal inhibition/excitation) and sparse connectivity into the RNN training pipeline.
• Trained RNN models with data-driven, cell type-specific connectivity constraints.
• Reconstructed two-photon calcium imaging data from mice during visual behavior across cortical layers and brain areas.

Main Results:

• Constrained RNN models achieved performance matching unconstrained RNNs.
• Successfully inferred multi-regional interactions using biologically plausible RNNs.
• The inferred interactions align with experimental findings and predictive coding theory.

Conclusions:

• Biologically constrained RNNs offer a valid and powerful approach for modeling cortical function.
• These methods enhance the physiological and anatomical fidelity of computational models.
• The approach provides insights into neural interactions consistent with established theories like predictive coding.