Learning is a critical research field for autonomous computer vision systems. It can bring solutions to the knowledge acquisition bottleneck of image understanding systems. Recent developments of machine learning for computer vision are reported in this paper. We describe several different approaches for learning at different levels of the image understanding process, including learning 2-D shape models, learning strategic knowledge for optimizing model matching, learning for adaptive target recognition systems, knowledge acquisition of constraint rules for labelling and automatic parameter optimization for vision systems. Each approach will be commented on and its strong and weak points will be underlined. In conclusion we will suggest what could be the “ideal” learning system for vision.

Inference is a very general reasoning process that allows us to draw consequences from some body of knowledge. Machine learning (ML) uses the three kinds of possible inferences, deductive, inductive, and analogical. We describe here different methods, using these inferences, that have been created during the last decade to improve the way machines can learn. We have already presented the most classical approaches in a book (Kodratoff, 1988), and in several review papers (Kodratoff, 1989, 1990a, 1992). These results will be described here very briefly, in order to leave room for newer results. We include also genetic algorithms as an induction technique. We restrict our presentation to the symbolic aspects of connectionism. A learning system can also be viewed as a mechanism skimming interesting knowledge out of the flow of information that runs through it. We present several existing learning systems from this point of view.

An approach to concept learning from examples and concept learning by observation is presented that is based on a intuitive notion of conceptual distance between examples (concepts) and combines symbolical and numerical methods. The approach is based on the observation that very different examples generalize to an expression that is very far from each of them, while identical examples generalize to themselves. Following this idea the authors propose some domain-independent and intuitively justified estimates for the conceptual distance. A hierarchical conceptual clustering algorithm that groups objects so as to maximize the cohesiveness (a reciprocal of the conceptual distance) of the clusters is presented. It is shown that conceptual clustering can improve learning from complex examples describing objects and the relation between them.>

RemovalThree different techniques that can be used for recursion removal are described: generalization of the function definition, study of the computation traces of the function, and nonprocedural languages.Despite the existence of implemented versions of these techniques, they are easy to use "by hand" and should therefore be part of every programmer's knowledge.