Accepted by:

We are interested in enabling a generic sketch recognition system that would allow more natural interaction with design tools in various domains. Instead of writing recognizer code for each new domain, new shapes should be added by describing them in a shape description language. While writing such descriptions is easier than writing code, it is still not a particularly easy or natural mode of interaction. The most natural way to teach new symbols to the system would be simply drawing them. This thesis presents a learning system that takes in a drawn symbol and produces a textual description of it. The main challenge is to decide which properties of the example are relevant. People cope with this task in part, we believe, through the use of perceptual biases. We use studies of human perception of geometry to understand these biases and use them to help select the relevant properties from a single example. The main generalization power is derived from two sources: 1) a qualitative description vocabulary that reflects properties that people pay attention to and 2) mechanisms, derived from the observations about perception, that adjust the relative importance of different properties based on the particular configuratio