2019/08/13 by Wenbo Gong, Sebastian Tschiatschek, Gong, Wenbo +9
Computer Science · #Machine Learning in Healthcare #Explainable Artificial Intelligence (XAI) #Gaussian Processes and Bayesian Inference
paper · pdf · doi:10.48550/arxiv.1908.04537
In this paper we introduce the ice-start problem, i.e., the challenge of\ndeploying machine learning models when only little or no training data is\ninitially available, and acquiring each feature element of data is associated\nwith costs. This setting is representative for the real-world machine learning\napplications. For instance, in the health-care domain, when training an AI\nsystem for predicting patient metrics from lab tests, obtaining every single\nmeasurement comes with a high cost. Active learning, where only the label is\nassociated with a cost does not apply to such problem, because performing all\npossible lab tests to acquire a new training datum would be costly, as well as\nunnecessary due to redundancy. We propose Icebreaker, a principled framework to\napproach the ice-start problem. Icebreaker uses a full Bayesian Deep Latent\nGaussian Model (BELGAM) with a novel inference method. Our proposed method\ncombines recent advances in amortized inference and stochastic gradient MCMC to\nenable fast and accurate posterior inference. By utilizing BELGAM's ability to\nfully quantify model uncertainty, we also propose two information acquisition\nfunctions for imputation and active prediction problems. We demonstrate that\nBELGAM performs significantly better than the previous VAE (Variational\nautoencoder) based models, when the data set size is small, using both machine\nlearning benchmarks and real-world recommender systems and health-care\napplications. Moreover, based on BELGAM, Icebreaker further improves the\nperformance and demonstrate the ability to use minimum amount of the training\ndata to obtain the highest test time performance.\n