Toward a Brain-Inspired Developmental Neural Network Based on Dendritic Spine Dynamics.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Neural networks with many parameters often overfit small datasets.
• Regularization techniques like weight penalties help mitigate overfitting.
• Dendritic spine plasticity, crucial for memory, involves dynamic changes in structure and synaptic strength.

Purpose of the Study:

• To propose a novel brain-inspired developmental neural network (BDNN-dsd) that utilizes dendritic spine dynamics.
• To address overfitting in neural networks by introducing adaptive weight constraints inspired by biological synaptic plasticity.
• To enhance classification performance and network convergence.

Main Methods:

• Developed a brain-inspired developmental neural network (BDNN-dsd) modeling dendritic spine dynamics (appearance, enlargement, shrinkage, disappearance).
• Implemented adaptive weight bounds modulated by synaptic activity, mimicking long-term potentiation/depression (LTP/LTD) and synapse formation/elimination.
• Constrained network weights to tunable bounds to limit redundant connections and promote effective ones.

Main Results:

• Demonstrated the effectiveness of BDNN-dsd on classification tasks using MNIST, Fashion MNIST, and CIFAR-10 datasets.
• Showcased improved network convergence rates compared to traditional methods.
• Achieved enhanced classification performance, particularly for compact networks.

Conclusions:

• The proposed BDNN-dsd effectively combats overfitting by dynamically adjusting weight bounds based on synaptic activity.
• This brain-inspired approach offers a promising alternative to standard regularization techniques like dropout and L2 regularization.
• The method improves both the speed of network training and the accuracy of predictions.