Biophysically interpretable recurrent neural network for functional magnetic resonance imaging analysis and sparsity...

Area of Science:

• Computational Neuroscience
• Machine Learning in Neuroscience
• Systems Neuroscience

Background:

• Recurrent Neural Networks (RNNs) are increasingly used in neuroscience for data analysis.
• Generic RNNs lack biophysical interpretability, hindering neuroscience-specific insights.
• Dynamic Causal Modelling (DCM) offers a biophysically grounded approach to inferring brain connectivity.

Purpose of the Study:

• To develop a novel Recurrent Neural Network (RNN) with inherent biophysical interpretability.
• To integrate Dynamic Causal Modelling (DCM) principles into a generalized RNN framework (DCM-RNN).
• To enhance causal architecture discovery in neuroscience using the proposed DCM-RNN.

Main Methods:

• Formulating Dynamic Causal Modelling (DCM) as a specialized generalized RNN.
• Defining DCM-RNN hidden states as neural activity and hemodynamic responses (blood flow, volume, deoxyhemoglobin).
• Utilizing $l_{1}$ connectivity regularization for sparse causal architecture discovery within the DCM-RNN framework.

Main Results:

• Demonstrated that DCM can be represented as a specific type of generalized RNN.
• Showcased DCM-RNN's ability to model neural and hemodynamic states and biologically relevant parameters.
• Empirically validated that DCM-RNN with $l_{1}$ regularization outperforms classic DCM with $l_{2}$ regularization in discovering sparse architectures, especially under noisy conditions.

Conclusions:

• DCM-RNN provides a biophysically meaningful and interpretable framework for analyzing neuroscience data.
• The proposed model offers a versatile tool for computational neuroscience, particularly when combined with deep learning.
• DCM-RNN with $l_{1}$ regularization advances causal architecture discovery, offering improved robustness and power over conventional DCM.