Model agnostic interpretability for multiple instance learning