The article introduces Prototype Generation, a robust form of feature visualisation for model-agnostic, data-independent interpretability of image classification models. It counters previous claims that feature visualisation algorithms are untrustworthy due to the unnatural internal activations by generating prototypes similar to natural images. It provides a global understanding of what a model has learned, necessary for comprehensive analysis and trust in automated systems. This method also helps in identifying systemic biases, uncovering spurious correlations, and potentially refining the model for better performance and fairness.
Publication date: 2 Oct 2023
Project Page: arXiv:2309.17144v1 [cs.CV] 29 Sep 2023
Paper: https://arxiv.org/pdf/2309.17144
