Recognition Networks for Approximate Inference in BN20 Networks

A recognition network is a multilayer perception (MLP) trained to predict posterior maxginals given observed evidence in a particulax Bayesian network. The input to the MLP is a vector of the states of the evidential nodes. The activity of an output unit is interpreted as a prediction of the posterior marginal of the corresponding variable. The MLP is trained using samples generated from the corresponding Bayesian network. We evaluate a recognition network that was trained to do inference in a large Bayesian network, similax in structure and complexity to the Quick Medical Reference, Decision Theoretic (QMR-DT) network. Our network is a binary, two-layer, noisy-OR (BN20) network containing over 4000 potentially observable nodes and over 600 unobservable, hidden nodes. In real medical diagnosis, most observables are unavailable, and there is a complex and unknown process that selects which ones axe provided. We incorporate a very basic type of selection bias in our network: a known preference that available observables are positive rather than negative. Even this simple bias has a significant effect on the posterior. We compare the performance of our recognition network to state-of-the-art approximate inference algorithms on a large set of test cases. In order to evaluate the effect of our simplistic model of the selection bias, we evaluate algorithms using a variety of incorrectly modelled selection biases. Recognition networks perform well using both correct and incorrect selection biases.