• ai
• artificial intelligence
• classification
• explainability
• machine learning
• trustworthiness

Explainability for black box decisions on image data is a research thesis on explainability in image classification decisions. For the methodology, I build on a latent-space exemplars and counter exemplars based model-agnostic explainability algorithm, ABELE (Guidotti, Monreale, Matwin, Pedreschi, 2019). My contribution consists in the pipeline “oab” that consists in three steps: first, the training data is sampled for a subset that is offline explained with the original method, that forms the “explanation base”. Then, in the model's latent space, incoming test records are matched with the existing explanation base; finally the explanation base's rules and counterrules are used to explain the test record.