Fine-Pruning: A biologically inspired algorithm for personalization of machine learning models.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Deep neural networks (DNNs) mimic brain neuron design but use biologically implausible training methods like backpropagation.
• Backpropagation necessitates extensive computational resources and fully labeled datasets, creating significant development hurdles.
• Current DNN training methods are computationally intensive and data-hungry, limiting their application in resource-constrained settings.

Purpose of the Study:

• To investigate a biomimetic approach to machine learning model training inspired by biological brain pruning.
• To develop an efficient training methodology that bypasses the limitations of traditional backpropagation.
• To demonstrate the efficacy of biologically inspired learning for personalized and resource-efficient AI.

Main Methods:

• Implemented a novel training strategy based on neural network pruning, mimicking brain-based learning mechanisms.
• Applied the pruning-based method to personalize speech recognition and image classification models.
• Utilized ResNet50 on ImageNet for experimental validation of the proposed approach.

Main Results:

• Achieved significant model sparsity, approximately 70%, through biologically inspired pruning.
• Improved model accuracy to around 90% in personalized speech and image classification tasks.
• Demonstrated orders of magnitude reduction in computational resource requirements compared to backpropagation.

Conclusions:

• Biomimetic pruning offers an efficient alternative to backpropagation for training deep neural networks.
• This approach enables the creation of personalized machine learning models with reduced computational and data requirements.
• The findings present a promising direction for developing AI in resource-limited environments and advancing artificial general intelligence.