EvoPruneDeepTL: An evolutionary pruning model for transfer learning based deep neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep Learning models excel at complex optimization but require extensive datasets.
• Transfer learning adapts pre-trained models, making layer configuration critical for performance.
• Optimizing Deep Learning models is computationally intensive, often involving pruning strategies.

Purpose of the Study:

• To propose EvoPruneDeepTL, an evolutionary pruning model for Transfer Learning based Deep Neural Networks.
• To optimize the configuration of fully-connected layers in transfer learning models.
• To investigate pruning and feature selection for improved computational efficiency and accuracy.

Main Methods:

• Developed EvoPruneDeepTL, an evolutionary pruning model for Deep Neural Networks (DNNs) utilizing Transfer Learning (TL).
• Replaced last fully-connected layers with sparse layers optimized via a genetic algorithm.
• Implemented solution encoding strategies for optimized pruning and feature selection.

Main Results:

• EvoPruneDeepTL demonstrated significant contributions to computational efficiency.
• The model improved network accuracy while simultaneously reducing the number of active neurons.
• Experiments validated the benefits of EvoPruneDeepTL and feature selection on various datasets.

Conclusions:

• EvoPruneDeepTL effectively optimizes DNNs for transfer learning tasks.
• The proposed method enhances computational efficiency and model accuracy.
• Evolutionary pruning and feature selection offer a viable strategy for improving DNN performance.