Poly-CAM: High resolution class activation map for convolutional neural networks

Current explainability methods have significant limitations: Existing techniques for visualizing what neural networks "see" suffer from problems like low resolution (CAM), excessive smoothing (perturbation-based methods), or scattered noise patterns (gradient-based approaches), making it difficult to pinpoint exactly which image features drive predictions.

Poly-CAM improves localization precision: The proposed method combines information from both early and late layers of neural networks to create higher-resolution activation maps that more accurately identify which specific image features the network uses to make its decisions.

Better feature localization matters for trust: Precise identification of class-specific features is important for responsible AI because it allows humans to verify whether models are making decisions based on relevant, meaningful image features rather than spurious correlations or artifacts.

Performance validated on standard metrics: The approach maintains competitive performance on "insertion-deletion faithfulness metrics" (standard tests measuring whether highlighted features actually matter to the model's prediction) while achieving superior precision in locating features.

Explainability remains an active challenge: This work demonstrates that making AI systems interpretable requires ongoing technical innovation, as no single existing method adequately balances resolution, smoothness, and accuracy in explaining neural network decisions.