Decimatable Boltzmann Machines for Diagnosis: Efficient Learning and Inference
Stefan Rüger · 1997
By using hidden nodes, Boltzmann machines can be employed to approximate probability distributions, store stochastic knowledge extracted from examples (a. k. a. learning) and retrieve the stored statistical information (a. k. a. inference or diagnosis). However, in general, the algorithms for such operations are NP-hard. Efficient algorithms can therefore only be expected to be found in special cases. Such algorithms have recently been specified for a certain set of so-called decimatable Boltzmann machines [9]. This approach is demonstrated for a small diagnosis system that was introduced in [5] as a didactic example: it deals with the joint probability of several medical variables. We show that a batch of fully diagnosed examples of medical data is sufficient not only to store the underlying stochastic correlations in a decimatable Boltzmann machine, but also to use this for inference or diagnosis. We argue that models like those obtained by decimatable Boltzmann machines — which learn from given examples only — may also help in building medically relevant models. 1