Dilated Convolution with Learnable Spacings makes visual models more aligned with humans: a Grad-CAM study

DCLS makes AI vision models more interpretable: When researchers replaced standard convolution layers with Dilated Convolution with Learnable Spacing (DCLS) in eight different computer vision models, seven of them showed improved alignment with human visual attention patterns, as measured by correlation with the ClickMe dataset of human eye-tracking data.

Better interpretability doesn't require more complexity: DCLS achieves this improved human alignment while maintaining the same efficiency benefits as standard dilated convolution—it enlarges the models' visual field without adding extra parameters, suggesting that responsible AI doesn't always mean sacrificing performance.

Existing interpretability tools have blind spots: The researchers discovered that Grad-CAM, a widely-used method for visualizing what AI models "see," generated essentially random heatmaps for two model types (CAFormer and ConvFormer), highlighting that current explainability techniques can fail silently without proper validation.

Simple fixes can improve AI transparency: By introducing Threshold-Grad-CAM, a straightforward modification to the standard Grad-CAM method, the researchers enhanced interpretability across nearly all tested models, demonstrating that incremental improvements to explanation tools can make AI systems more trustworthy and understandable.

Reproducibility enables verification: The authors made their code and model checkpoints publicly available, allowing other researchers to verify these findings and build upon them—a practice essential for establishing trust in responsible AI research.