XAI-BO: an architecture using Grad-CAM technique to evaluate Bayesian optimization algorithms on deep learning models

Abstract

The Bayesian optimization algorithm, proposed in the 1980s, is known for its advantages in optimizing parameters and conserving system resources. However, the specific application of this algorithm in various deep-learning models has not been thoroughly evaluated. This study proposes an experimental model to assess the dependency on reducing hyperparameters in Dense layers and dropout layers, as well as the performance and accuracy across 13 deep learning models, implemented on two datasets with different characteristics and sample sizes (Cucumber disease recognition dataset and tomato physiological states dataset). Additionally, the research utilizes Grad-CAM to clarify the positive impact of parameter reduction when using the BO algorithm. Through evaluations based on hyperparameter data, performance by each iteration, accuracy, and model explainability, the study demonstrates that the BO algorithm optimizes hyperparameters through selective processes. However, it significantly impacts the model's accuracy. This is particularly evident in the InceptionResNetV2 and NasNetMobile models. The research results contribute to a clearer understanding of the impact of optimization algorithms on deep learning models, opening up new research directions in optimizing and elucidating models through the use of explainable artificial intelligence.

Keywords

• Optimization algorithm
• disease classification
• Bayesian optimization
• hyperparameter tuning
• Grad-CAM