Learning from data with errors

An inference criterion representing a trade-off between “fit” and simplicity leads to a general method for constructing concept descriptions from a class of descriptors, given examples with errors. In one application, this method is combined with a specific descriptor generator into a program learning structural descriptions of shape.