Diagnosing Shortcut Learning in Spatial Vision with XAI Heatmaps

You build a computer vision model for spatial gesture recognition. The training loss drops smoothly, validation accuracy hits 98%, and everything looks solid in notebook evaluation. Then you test it i