Interpreting Image Classifiers by Generating Discrete Masks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models, particularly image classifiers, often function as "black-boxes", lacking transparency and interpretability.
• Understanding the decision-making process of these models is crucial for trust and debugging.

Purpose of the Study:

• To develop a novel, interpretable approach for deep image classifiers.
• To generate discrete masks that highlight discriminative image regions relevant to a model's prediction.

Main Methods:

• Utilized a generative adversarial network (GAN) framework where the deep model acts as the discriminator and a generator explains its decisions.
• Trained the generator to produce probability maps for discrete mask sampling, with the discriminator evaluating mask quality.
• Employed policy gradients for generator training due to sampling operations and incorporated gradients as auxiliary information.

Main Results:

• Demonstrated the method's ability to provide reasonable explanations for deep image classifier predictions.
• Achieved superior performance compared to existing interpretability approaches on the ILSVRC dataset.
• Validated the model's reasoning capabilities through the model randomization test.

Conclusions:

• The proposed GAN-based method offers a robust and effective way to interpret deep image classifiers.
• The generated masks provide meaningful insights into the model's attention and decision-making process.
• This approach enhances the transparency and trustworthiness of deep learning models in image classification tasks.