FAPI-Net: A lightweight interpretable network based on feature augmentation and prototype interpretation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) offer high performance but suffer from large parameter counts, high computation (FLOPs), and a black-box nature.
• These characteristics impede deployment on resource-constrained devices and raise concerns in high-stakes decision-making fields.
• Existing research on lightweight DNNs often overlooks model interpretability.

Purpose of the Study:

• To propose FAPI-Net, a novel lightweight and interpretable deep neural network architecture.
• To reduce network parameters and computation while maintaining or improving accuracy.
• To provide visualization of the model's classification reasoning process.

Main Methods:

• Developed FAPI-Net by integrating feature augmentation convolution blocks (composed of lightweight feature-map augmentation (FA) modules and residual connections) with a prototype dictionary interpretability (PDI) module.
• Designed FAPI-Net based on the MobileNetV3 architecture.
• Conducted experiments on ILSVRC2012 and CIFAR-10 datasets, including ablation studies.

Main Results:

• FAPI-Net demonstrated a 2% reduction in parameters and a 20% reduction in FLOPs compared to MobileNetV3 on the ILSVRC2012 dataset.
• The inclusion of a trainable PDI module resulted in negligible accuracy loss compared to baseline models.
• Ablation experiments on CIFAR-10 confirmed the efficacy of the FA module.
• Visualization experiments confirmed FAPI-Net's ability to clarify classification decision processes for specific images.

Conclusions:

• FAPI-Net offers a promising solution for developing efficient and interpretable deep learning models.
• The proposed FA modules effectively reduce computational overhead without compromising accuracy.
• The PDI module enhances model transparency by visualizing classification reasoning, making DNNs more trustworthy in critical applications.