Guided Zoom: Zooming into Network Evidence to Refine Fine-Grained Model Decisions.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep single-label classification models often show higher top-k accuracy than top-1, particularly in fine-grained datasets with subtle class differences.
• Leveraging information from top-k predictions can significantly boost model performance.
• Explainability methods are increasingly explored to improve model reliability and performance.

Purpose of the Study:

• To introduce Guided Zoom, a novel method for enhancing deep learning model performance using explainability.
• To ensure deep neural networks make predictions based on "right reasons" or valid evidence.
• To investigate the coherence of evidence used for predictions against training data.

Main Methods:

• Guided Zoom examines the reasonableness of evidence (pixel-space grounding) for top-k predictions.
• It assesses if test-time evidence aligns with evidence from similar correct training decisions.
• The study explores various grounding techniques and their complementarity for evidence computation.

Main Results:

• Guided Zoom leads to improved classification accuracy in deep learning models.
• The method achieves state-of-the-art performance on four fine-grained classification datasets.
• Evidence coherence checking enhances the informativeness of model predictions.

Conclusions:

• Explainability, through Guided Zoom, can be effectively used to boost model performance.
• Verifying prediction evidence against training data improves classification accuracy, especially for challenging datasets.
• Guided Zoom offers a promising direction for developing more robust and accurate deep learning classifiers.