Multistrategy Learning in Vision: Integrating Symbolic and Neural Net Learning for Vision Tasks

(Michalski, Zhang)

The project concerns the development of a novel multistrategy learning methodology that is specifically oriented toward vision learning. The methodology combines symbolic rule learning and neural-based learning strategies in order to achieve high efficiency and accuracy in learning visual object descriptions, and in applying these descriptions to rapid object recognition.

The initially developed vision system has several advantages: it can be easily modified and applied to new problems (due to learning), its learning speed can be at least an order of magnitude faster than neural net learning (due to symbolic pre-structuring of the net), it has short recognition times (due to its parallel architecture), and its underlying recognition rules are easy to understand by a human operator (due to the symbolic knowledge representation of the basic decision rules). The developed system was experimentally applied to natural scene recognition.

The method works in two stages: 1) rule learning using the AQ algorithm. This phase generates rules that generally and approximately describe the training examples, 2) neural net learning to determine the final visual concept description. The network is structured according to the rules obtained in stage. Each node in the hidden-layer of the network corresponds to a single rule. The degree of match of an example to the rule represents node activation. This activation value is input to the sigmoid transfer function associated with each node. Weight values for the connections between nodes and outputs are obtained using backpropagation method.

Selected References

Michalski, R.S., Zhang, Q., Maloof, M.A. and Bloedorn. E., The MIST Methodology and its Application to Natural Scene Interpretation, Proceedings of the Image Understanding Workshop, Palm Springs, CA, Feburary, 1996.

For more references, see publications section.