Although machine learning techniques have been applied with remarkable success to several problems of computer perception and vision, most of these problems have been fairly simple in nature. The difficulty with scaling up to more complex tasks is that inductive learning methods require a very large number of training examples in order to generalize correctly from complex sensor data.
This chapter proposes an approach to overcoming this difficulty, by relying on previously learned information to augment the available training data. In particular, we consider the task faced by a mobile robot learning to recognize new objects within an already-familiar environment. Because the robot has previously operated within this environment (here the corridors of a particular building), it has already had the opportunity to learn certain regularities that can be useful in subsequent learning tasks. Given a new task, such as learning to recognize the distance to the next door in the corridor, knowledge of these regularities enables the system to learn the new task more accurately from a limited quantity of new training data. We describe the Explanation-Based Neural Network (EBNN) algorithm for utilizing previously learned knowledge, and examine its performance for the mobile robot perception task of door recognition in a familiar corridor based on color vision and sonar sensor data. Experimental results indicate that EBNN is able to generalize more accurately than purely inductive methods such as Backpropagation, even when its prior knowledge is only approximately correct.