Structure mapping engine

Abstract: Learning concepts from sketches via analogical generalization and near-misses Matthew D. McLure ([email protected]) Scott E. Friedman ([email protected]) Kenneth D. Forbus ([email protected]) Qualitative Reasoning Group, Northwestern University, 2133 Sheridan Rd Evanston, IL 60208 USA concept in only one way. A near miss exemplar should be highly alignable with some instances of a concept 1 . This paper describes a model of concept learning that combines analogical generalization and near-miss analysis to capture both similarity-based and analytic aspects of concepts. Its inputs are labeled positive or negative examples of concepts. It uses SAGE to construct generalizations for each concept, thus capturing similarity- based aspects of concepts (and typicality, via probability). When a positive example is provided, the corresponding concept is updated. When a negative example is provided, analogical retrieval is used to find the closest prior positive example or generalization, and analogical matching is used to construct and update hypotheses about inclusion and Abstract Modeling how concepts are learned from experience is an important challenge for cognitive science. In cognitive psychology, progressive alignment, i.e., comparing highly similar examples, has been shown to lead to rapid learning. In AI, providing very similar negative examples (near-misses) has been proposed as another way to accelerate learning. This paper describes a model of concept learning that combines these two ideas, using sketched input to automatically encode data and reduce tailorability. SAGE, which models analogical generalization, is used to implement progressive alignment. Near-miss analysis is modeled by using the Structure Mapping Engine to hypothesize classification criteria based on differences. This is performed both on labeled negative examples provided as input, and by using analogical retrieval to find near-miss examples when positive examples are provided. We use a corpus of sketches to show that the model can learn concepts based on sketches and that incorporating near-miss analysis improves learning. Keywords: Concept learning; analogy; generalization. Introduction How concepts are learned from experience is a central question in cognitive science. It is well-known that some concepts can be viewed as analytic, having compact necessary and sufficient defining criteria (e.g., grandparent or triangle), whereas others are based on similarity or typicality (e.g., chair, bachelor). Prior work has explored analogical generalization as an explanation for learning similarity-based categories. The SAGE model of analogical generalization, an evolutionary improvement over SEQL (Kuehne et al 2000a) has been used to model learning of perceptual stimuli (Kuehne et al 2000b), stories (Kuehne et al 2000a), spatial prepositions (Lockwood et al 2008) and causal models (Friedman & Forbus, 2008; Friedman & Forbus, 2009). SAGE’s ability to construct probabilistic generalizations provides a model of typicality, i.e., high- probability relationships and attributes are more typical. SAGE has been used to model progressive alignment (Gentner et al 2007), where sequences of highly similar exemplars lead to more rapid learning (Kuehne et al 2000a). Progressive alignment alone may suffice to generate rule- like concepts (e.g., Gentner & Medina, 1998), but another possibility is to use negative examples to sharpen criteria for concepts. Winston (1970) proposed the idea of a near-miss, a labeled negative example that differs from the intended Figure 1: An example of the skeletal arm concept drawn in CogSketch. exclusion criteria for that concept. Near-miss analysis is also attempted when a positive example is provided, using analogical retrieval over negative examples to look for a candidate near-miss. (Using analogical retrieval to find positive concepts and near-misses is a significant advance over Winston’s model, which used hand-coded representations, a single abstract description for concepts and required a teacher to supply all negative examples.) To test the model, we use sketches to describe concepts, which are automatically encoded by a sketch understanding system. We show that the model can indeed learn concepts from sketches, and that including near-miss analysis improves learning. Our simulation is implemented using the Companions cognitive architecture (Forbus et al, 2009), which integrates analogical processing and sketching. For disjunctive concepts, some exemplars will not be similar.

Abstract: This paper compares the performance of the Structure-Mapping Engine (SME), a cognitive simulation of analogy, with two aspects of human performance. Gentner's Structure-Mapping theory predicts that soundness is highest for relational matches, while accessibility is highest for surface matches. These 'predictions have been borne out in psychological studies, and here we demonstrate that SME replicates these results. In particular, we ran SME on the same stories used in the psychological studies with two different kinds of match rules. In analogy mode, SME closely captures the human soundness ordering. In mereappearance mode, SME captures the accessibility ordering. We briefly review the psychological studies, describe our computational experiments, and discuss the utility of SME as a cognitive modeling tool.

Abstract: Leading accounts of analogy based on structure mapping theory (Gentner 1983, 1989) give an important explanatory role to re-representation. Structural alignment is insufficiently flexible to account for human analogical processing if semantically compatible, but non-identically coded, representational elements are not permitted to match. A process of re-representation can selectively allow non-identical representational elements to be considered matches and placed in structural correspondence during comparison. However, re-representation is only a posited theoretical construct with minimal supporting evidence. An experimental paradigm called inference probing is introduced which offers a new level of empirical support for the psychological reality of re-representation. Behavioural results are presented that bear on accounts of analogy, similarity, knowledge representation and reasoning.