Leveraging Guided Backpropagation to Select Convolutional Neural Networks for Plant Classification.

Area of Science:

• Computer Science
• Machine Learning
• Plant Science

Background:

• Convolutional Neural Networks (CNNs) excel at plant classification, but complex models often overfit training data.
• Traditional evaluation metrics like accuracy and loss curves do not guide model improvement for overfitting.
• Understanding the relationship between model capacity, learned features, and performance is crucial.

Purpose of the Study:

• To analyze the impact of model capacity on feature learning in CNNs for plant classification.
• To investigate the diversity of features learned across different layers of a CNN.
• To propose a new method for selecting optimal CNN model depth to mitigate overfitting.

Main Methods:

• Analysis of the relationship between CNN representational capacity and learned features.
• Examination of feature diversity in shallow versus deep CNN layers.
• Development and application of the SSIM cut curve method using Guided Backpropagation visualizations.

Main Results:

• Models with higher representational capacity may learn excessive subtle features, negatively impacting performance.
• Shallow CNN layers learn more diverse features compared to deeper layers.
• The SSIM cut curve effectively visualizes feature similarity across depths.

Conclusions:

• The SSIM cut curve offers a novel approach to selecting appropriate CNN model depth for plant classification.
• This method provides insights into feature learning and model overfitting.
• The proposed technique has the potential to improve the selection of CNN models for enhanced performance.