Bayesian CAIPI: A Probabilistic Approach to Explanatory and Interactive Machine Learning

Explanatory Interactive Machine Learning queries user feedback regarding the prediction and the explanation of novel instances. CAIPI, a state-of-the-art algorithm, captures the user feedback and iteratively biases a data set toward a correct decision-making mechanism using counterexamples. The counterexample generation procedure relies on hand-crafted data augmentation and might produce implausible instances. We propose Bayesian CAIPI that embeds a Variational Autoencoder into CAIPI’s classification cycle and samples counterexamples from the likelihood distribution. Using the MNIST data set, where we distinguish ones from sevens, we show that Bayesian CAIPI matches the predictive accuracy of both, traditional CAIPI and default deep learning. Moreover, it outperforms both in terms of explanation quality.