Improved transfer learning using textural features conflation and dynamically fine-tuned layers.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Transfer learning leverages existing knowledge for new tasks, but effectiveness diminishes with task dissimilarity.
• Selecting relevant data and model layers is crucial for successful knowledge transfer.
• Previous methods focused on data or layer selection independently.

Purpose of the Study:

• To develop a unified model pipeline for transfer learning by combining data and layer selection strategies.
• To minimize knowledge loss during transfer learning by utilizing least divergent textural features and model layers.
• To enhance the accuracy and efficiency of fine-tuning pre-trained models.

Main Methods:

• Utilized five pre-trained models (ResNet50, DenseNet169, InceptionV3, VGG16, MobileNetV2) across nine diverse datasets.
• Implemented a novel approach combining data point selection based on minimal textural feature divergence and dynamic layer selection based on positive weight analysis.
• Evaluated the impact of individual and combined selection methods on model performance.

Main Results:

• Data points with lower textural feature divergence improved accuracy by 3.54% to 6.75% on CIFAR-100.
• Selecting layers with more positive weights boosted accuracy by 2.42% to 13.04% on CIFAR-100.
• The combined approach yielded an additional 1.56% accuracy improvement, demonstrating synergistic benefits.

Conclusions:

• Selecting data points with low divergence from source samples significantly improves target task adaptation in transfer learning.
• Choosing pre-trained model layers with predominantly positive weights streamlines the fine-tuning process, reducing experimental effort.
• The integrated pipeline offers a more robust and efficient strategy for optimizing transfer learning performance.