GRP-021 SHAP-Explainable Image-to-Topology Regression

We evaluated whether deep regression models predicting vectorized topological features (in the form of persistence landscapes) actually learn the underlying persistent homology of the image. A DenseNet-121 is trained to regress 300-dimensional persistence landscapes from grayscale scene images. Using SHAP, we evaluate the contribution of pixels in the original images to the persistence landscapes. Across all six classes, SHAP-feature overlap is consistently lower than the baseline, implying that DenseNet may not be truly learning the underlying persistent homology.