Toward Automated Sketching Collaborators: An Analogical Route


Automated sketch collaborators might help us create more dynamic intelligent tutoring systems, work out designs, reduce bias in solving spatial social problems, and organize our ideas. Here, we examine some properties of sketch recognition methods designed to help serve that goal. Structure Mapping techniques are applied to symbolic structural descriptions of sketched objects in order to classify them. First, we evaluate processes for encoding ink into a structural description in ways that are controllably tractable. We contribute a perceptual organization scheme involving edge-cycles, which is more abstract, concise, and effective than using edges. We also provide filtering strategies for keeping input tractable for analogical matching, which can be applied to edge-cycles, edges or both, and we demonstrate the value of a hybrid encoding with a filter based on ink-coverage. Second, we evaluate two variants of an approach for extending similarity-based classifiers with classification criteria gleaned from near-miss comparisons – highly similar positive-negative pairs – and we look at its performance over a range training data compression rates. We find that in moderation, criteria from near-misses can improve similarity-based classification. Third, we demonstrate that an alternative approach to learning concept boundaries, which uses linear support vector machines built on analogical generalization, achieves higher, more stable accuracy than near-miss and similarity-based classifiers with compression of training data that is equal or better. Finally, we discuss how the analogical approach compares to related work and close by suggesting future work that that could build on these contributions toward the goal of sketching collaborators.

Date created
Resource type
Rights statement