Fig. 6: Black-box explanations for state-of-the-art approaches.
From: Thermodynamics-inspired explanations of artificial intelligence
a LIME explanation for ‘Eyeglasses’ with top j = 10 features, b Sparse model size vs. hyperparameter α that regulates the strength of L1 regularization. SHAP values for c ‘Eyeglasses’, d, and ‘Male’ prediction respectively. Consistency of these results with explanations shown in Fig. 5 validates TERP.
