BEAN: Interpretable and Efficient Learning With Biologically-Enhanced Artificial Neuronal Assembly Regularization.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Deep neural networks (DNNs) excel at data analysis but lack interpretability, especially in dense layers.
• Classical DNNs assume conditional independence between neurons, hindering co-training and modularity.
• Biological neural networks utilize neuronal assemblies for robust representation encoding.

Purpose of the Study:

• To develop a novel regularization technique inspired by neuroscience to model neuronal correlations and dependencies.
• To enhance the interpretability and efficiency of deep neural networks.

Main Methods:

• Proposed Biologically Enhanced Artificial Neuronal assembly (BEAN) regularization.
• Modeled neuronal correlations and dependencies based on cell assembly theory.
• Evaluated BEAN on DNNs for interpretability, efficiency, and few-shot learning performance.

Main Results:

• BEAN regularization facilitated the formation of interpretable neuronal functional clusters.
• Achieved a sparse, memory, and computation-efficient network without compromising model performance.
• Demonstrated enhanced model generalizability in few-shot learning scenarios with limited training data.

Conclusions:

• BEAN regularization effectively models neuronal dependencies, leading to more interpretable and efficient DNNs.
• The approach shows promise for improving deep learning models, particularly in data-scarce environments.
• BEAN offers a biologically inspired method to advance artificial neural network design.