In this thesis, a methodology for integrating knowledge-based techniques into connectionist approaches for visual pattern recognition is presented. We first propose a connectionist model called Adaptive Inference Network (AINET) which has expert system capability as well as instance-based learning capability. The proposed model has acyclic feedforward network structure and can be trained by Back-Propagation learning algorithm. In the model, there are two types of connections which represent the fuzzy relations between feature variables. We discuss the role of the dual connections in comparison with the conventional neural network models. We describe the network structure and behavior, the learning method, and other features of the proposed model, and then introduce a knowledge representation technique based on fuzzy relations and a knowledge utilization method in the learning process. To evaluate the usefulness of the model for the practical applications, we consider a pattern recognition system model which consists of two stages : feature extraction stage and classification stage. For the feature extraction stage, modular structure neural networks and conventional approaches are used together to extract the features more effectively. The proposed model exhibits four major characteristics: 1) logical inference ability, 2) knowledge acquisition by learning, 3) performance improvement by utilizing expert knowledge and 4) the capability of explaining about the final decisions. Networks for simple logical operations such as AND and XOR functions are illustrated to evaluate these effects. Some advantageous features of the proposed model for the general pattern recognition applications are also discussed in this thesis. First, the model upgrades the learning speed in comparison with the ordinary Multi-Layer Perceptron which is the most popular neural network model. Second, the model provides the useful facilities for rule validation and refinement, and learning data ...