2021/12/05 by Dan Ley, Ley, Dan, Umang Bhatt +3 · 1 citation
Computer Science · #Explainable Artificial Intelligence (XAI) #Machine Learning in Healthcare #Adversarial Robustness in Machine Learning
paper · pdf · doi:10.48550/arxiv.2112.02646
To interpret uncertainty estimates from differentiable probabilistic models,\nrecent work has proposed generating a single Counterfactual Latent Uncertainty\nExplanation (CLUE) for a given data point where the model is uncertain,\nidentifying a single, on-manifold change to the input such that the model\nbecomes more certain in its prediction. We broaden the exploration to examine\n\δ-CLUE, the set of potential CLUEs within a \δ ball of the\noriginal input in latent space. We study the diversity of such sets and find\nthat many CLUEs are redundant; as such, we propose DIVerse CLUE\n(\∇-CLUE), a set of CLUEs which each propose a distinct explanation as to\nhow one can decrease the uncertainty associated with an input. We then further\npropose GLobal AMortised CLUE (GLAM-CLUE), a distinct and novel method which\nlearns amortised mappings on specific groups of uncertain inputs, taking them\nand efficiently transforming them in a single function call into inputs for\nwhich a model will be certain. Our experiments show that \δ-CLUE,\n\∇-CLUE, and GLAM-CLUE all address shortcomings of CLUE and provide\nbeneficial explanations of uncertainty estimates to practitioners.\n