Training a hierarchical image model without a high-level teacher

Summary form only given. The coupled Markov random field (MRF) model of images is specified by local energy parameters. In previous studies, line-process energy parameters were successfully estimated by an algorithm previously proposed by the authors (1989), when teaching information about the line process was given and the learning was conducted in MRF equilibrium. The same algorithm was implemented on a massively parallel supercomputer to investigate the following two problems. First, learning was conducted while the MRF was still in a transient state. Surprisingly and fortunately, the transient learning was even faster than equilibrium learning. Second, a new scheme was proposed which works even when a high-level teacher for a hierarchical model is absent. Using this scheme the line-process energies were efficiently estimated when only image intensity data were provided