AutoTune

AutoTune optimizes fine-tuning of pre-trained convolutional neural networks (CNNs) using Bayesian Optimization to improve transfer learning performance on target image-classification datasets.

Key Features:

• Bayesian Optimization: Employs Bayesian Optimization to systematically adjust pre-trained CNN layers using insights derived from the target data.
• Softmax Layer Adjustment: Trains the final layer of a base CNN and modifies the number of neurons in the softmax layer to match the class count of the target task.
• Automated Fine-tuning: Fine-tunes the entire network automatically after final-layer training.
• Validation-guided Greedy Criteria: Guides fine-tuning using performance metrics observed on validation data via a greedy criterion.
• Transfer Learning Focus: Specifically targets adaptation of pre-trained models for tasks with limited target-data availability.
• Benchmark Evaluation: Evaluated on CalTech-101, CalTech-256, and Stanford Dogs with reported classification accuracies of 95.92%, 86.54%, and 84.67%, respectively.

Scientific Applications:

• Image Classification Benchmarks: Improves classification accuracy on image datasets such as CalTech-101, CalTech-256, and Stanford Dogs.
• Transfer Learning for Limited Data: Enhances adaptability and performance of pre-trained CNNs when target datasets are small or task-specific.

Methodology:

AutoTune first trains the final layer of a base CNN, modifying the softmax neuron count to match target classes; it then employs Bayesian Optimization to adjust pre-trained CNN layers and automatically fine-tunes the entire network guided by validation performance using a greedy criterion, and was evaluated on CalTech-101, CalTech-256, and Stanford Dogs with reported accuracies of 95.92%, 86.54%, and 84.67%.