DeepVID: Deep Visual Interpretation and Diagnosis for Image Classifiers via Knowledge Distillation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep Neural Networks (DNNs) demonstrate high performance across disciplines but often function as "black-boxes", hindering understanding.
• Model interpretability is crucial for trust, especially in safety-critical domains such as medical diagnosis and autonomous driving.

Purpose of the Study:

• To introduce DeepVID, a novel deep learning framework for visually interpreting and diagnosing DNN models, with a focus on image classifiers.
• To address the challenge of DNN interpretability by developing a method that enhances trust and understanding of model decision-making processes.

Main Methods:

• Training a small, locally-faithful model to approximate the behavior of a complex DNN around specific data instances.
• Employing knowledge distillation to transfer knowledge from the large DNN to the simpler, interpretable local model.
• Utilizing a deep generative model (Variational Auto-Encoder) to create semantically meaningful neighbors for probing DNN behavior.

Main Results:

• DeepVID successfully mimics DNN behavior locally, enabling visual interpretation.
• The generated neighbors effectively probe DNNs, aiding the local model's learning process.
• Comprehensive evaluations and expert case studies validate DeepVID's effectiveness in interpreting DNNs.

Conclusions:

• DeepVID provides a viable solution for the visual interpretation and diagnosis of Deep Neural Networks.
• The proposed method enhances the transparency and trustworthiness of DNNs, particularly for image classification tasks.
• DeepVID facilitates a deeper understanding of DNNs, crucial for their reliable deployment in critical applications.