Single-example learning of novel classes using representation by similarity Evgeniy Bart * and Shimon Ullman Dept. of Computer Science and Applied Mathematics Weizmann Institute of Science Rehovot, Israel 76100 evgeniy@csail.mit.edu, shimon.ullman@weizmann.ac.il Abstract We describe an object classification method that can learn from a single train- ing example. In this method, a novel class is characterized by its similarity to a number of previously learned, familiar classes. We demonstrate that this similarity is well-preserved across different class instances. As a result, it generalizes well to new instances of the novel class. A simple comparison of the similarity patterns is therefore sufficient to obtain useful classifica- tion performance from a single training example. The similarity between the novel class and the familiar classes in the proposed method can be evaluated using a wide variety of existing classification schemes. It can therefore com- bine the merits of many different classification methods. Experiments on a database of 107 widely varying object classes demonstrate that the proposed method significantly improves the performance of the baseline algorithm. 1 Introduction Recent methods of visual object classification achieve high performance levels. However, they require hundreds of examples for training [12, 13, 5, 2]. The cost of collecting such large amounts of training data may be prohibitive. For example, when learning to avoid dangerous objects (e.g. predators), situations that permit acquisition of training examples are hazardous. Since the system’s behavior is incorrect until a sufficient number of examples has been gathered, minimizing this number is crucial to allow adaptation to new situations. Obtaining useful performance with very few training examples is also important when the learning is incremental (i.e. the examples are presented sequentially and the system is updated after each presentation). In addition, realistic classification schemes should be able to handle a large number of classes. For example, it is estimated that humans are familiar with tens of thousands of different classes [3]. As a result, the accumulated cost of learning all classes may become excessive. Reducing as much as possible the number of required training examples may help deal with this problem. In this paper, we propose an object classification method that can learn from a single example. With this method, a system that can already classify several classes can be ex- tended to classify an additional novel class using a single training example. We assume * http://people.csail.mit.edu/evgeniy