[Paper Review] Efficient Saliency Maps for Explainable AI

T. Nathan Mundhenk, Barry Y. Chen+1 more|arXiv (Cornell University)|Nov 26, 2019

Explainable Artificial Intelligence (XAI)31 references47 citations

TL;DR

Introduces SMOE Scale saliency maps computed from five network scales during a single forward pass, yielding efficient and accurate XAI maps that can rival gradient-based methods, with optional Layer Ordered Visualization (LOVI) for scale-aware interpretation.

ABSTRACT

We describe an explainable AI saliency map method for use with deep convolutional neural networks (CNN) that is much more efficient than popular fine-resolution gradient methods. It is also quantitatively similar or better in accuracy. Our technique works by measuring information at the end of each network scale which is then combined into a single saliency map. We describe how saliency measures can be made more efficient by exploiting Saliency Map Order Equivalence. We visualize individual scale/layer contributions by using a Layer Ordered Visualization of Information. This provides an interesting comparison of scale information contributions within the network not provided by other saliency map methods. Using our method instead of Guided Backprop, coarse-resolution class activation methods such as Grad-CAM and Grad-CAM++ seem to yield demonstrably superior results without sacrificing speed. This will make fine-resolution saliency methods feasible on resource limited platforms such as robots, cell phones, low-cost industrial devices, astronomy and satellite imagery.

Motivation & Objective

• Motivate the need for efficient, fine-grained saliency maps suitable for real-time or embedded use.
• Develop a saliency estimation framework that operates during the standard forward pass by analyzing end-of-scale activations.
• Propose a statistically grounded, computationally light saliency statistic (SMOE Scale) and show its robustness.
• Introduce a method to combine multi-scale saliency maps into a single visualization (LOVI) with HSV encoding of layer contributions.
• Evaluate the method against state-of-the-art saliency techniques using ROAR/KAR metrics and compare efficiency.

Proposed method

• Compute saliency maps at the end of each spatial scale in a CNN (e.g., five scales in ResNet-50).
• Define a fast saliency statistic (SMOE Scale) derived from a simplified Gamma-scale parameter estimation expressed as a mean-activation times a log-information term across channel activations (Eq. 2).
• Assume Saliency Map Order Equivalence (SMOE) to avoid full iterative estimation and preserve ranking of salient pixels.
• Normalize each scale saliency map with a per-map 0-1 transformation using the normal CDF, then upsample to input size and combine with a weighted average (Eq. 3).
• Provide a Layer Ordered Visualization of Information (LOVI) by encoding per-location scale contributions in HSV color space (Hue = layer center of mass, Saturation = sparsity/uniqueness, Value = max activation).
• Optionally fuse SMOE Scale with Grad-CAM++ to create a stronger combined saliency signal.

Experimental results

Research questions

• RQ1 Can a single forward-pass-derived, multi-scale saliency method match or exceed the accuracy of gradient-based saliency methods (e.g., Guided Backprop, Grad-CAM) with dramatically reduced computation?
• RQ2 Does combining scale-wise saliency maps via SMOE Scale and simple normalization yield robust, interpretable visualizations across diverse architectures and datasets?
• RQ3 Does multi-scale visualization (LOVI) provide actionable insights into how information is distributed across network stages?
• RQ4 How do ROAR and KAR metrics evaluate the quality of the proposed saliency maps compared to existing methods on standard benchmarks?
• RQ5 What is the quantitative and qualitative impact of fusing SMOE Scale with Grad-CAM++ on saliency accuracy and speed?

Key findings

• SMOE Scale achieves strong KAR/ROAR performance, outperforming several fast gradient-based methods on multiple datasets.
• The method yields significant speed advantages, processing only five layers during a forward pass with negligible overhead.
• LOVI enables interpretability of scale-wise contributions, highlighting how early vs. late network stages contribute to saliency.
• Combining SMOE Scale with Grad-CAM++ improves overall saliency accuracy while maintaining fast inference relative to full-gradient methods.
• Compared to full backpropagation-based saliency methods, SMOE Scale demonstrates comparable or superior efficiency and competitive accuracy across tested tasks.