Taking a Look (Literally!) at the Raven’s Intelligence Test: Two Visual Solution Strategies

Taking a Look (Literally!) at the Raven’s Intelligence Test: Two Visual Solution Strategies Maithilee Kunda (mkunda@gatech.edu) Keith McGreggor (keith.mcgreggor@gatech.edu) Ashok Goel (goel@cc.gatech.edu) Design & Intelligence Laboratory, School of Interactive Computing, Georgia Institute of Technology 85 Fifth Street NW, Atlanta, GA 30332 USA over which various kinds of reasoning then take place. In this paper, we provide evidence from two different methods that Raven’s problems can be solved visually, without first converting problem inputs into propositional descriptions. Abstract The Raven’s Progressive Matrices intelligence test is widely used as a measure of Spearman’s general intelligence factor g. Although Raven’s problems resemble geometric analogies, prior computational accounts of solving the test have been propositional. Studies of both typical and atypical human behavior suggest the possible existence of visual strategies; for example, neuroimaging data indicates that individuals with autism may preferentially recruit visual processing brain regions when solving the test. We present two different algorithms that use visual representations to solve Raven’s problems. These algorithms yield performances on the Standard Progressive Matrices test at levels equivalent to typically developing 9.5- and 10.5- year-olds. We find that these algorithms perform most strongly on problems identified from factor-analytic human studies as requiring gestalt or visuospatial operations, and less so on problems requiring verbal reasoning. We discuss implications of this work for understanding the computational nature of Raven’s and visual analogy in problem solving. Keywords: Analogy; intelligence tests; representations; mental imagery; Raven’s Matrices; visual reasoning. Existing Computational Accounts knowledge Progressive Introduction The Raven’s Progressive Matrices tests (Raven, Raven, & Court, 1998) are a collection of standardized intelligence tests that consist of geometric analogy problems in which a matrix of geometric figures is presented with one entry missing, and the correct missing entry must be selected from a set of answer choices. Figure 1 shows an example of a 2x2 matrix problem that is similar to one in the Standard Progressive Matrices (SPM); other problems contain 3x3 matrices. The entire SPM consists of 60 problems divided into five sets of 12 problems each (sets A, B, C, D & E), roughly increasing in difficulty both within and across sets. Although the Raven’s tests are supposed to measure only eductive ability, or the ability to extract and understand information from a complex situation (Raven, Raven, & Court 1998), their high level of correlation with other multi- domain intelligence tests have given them a position of centrality in the space of psychometric measures (e.g. Snow, Kyllonen, & Marshalek 1984), and as a result, they are often used as tests of general intelligence in clinical, educational, vocational, and scientific settings. Computational accounts of problem solving on the Raven’s tests have, with the exception of Hunt (1974), assumed that visual inputs are translated into propositions, Carpenter, Just, and Shell (1990) used a production system that took hand-coded symbolic descriptions of problems from the Advanced Progressive Matrices (APM) test and then selected an appropriate rule to solve each problem. The rules were generated by the authors from a priori inspection of the APM and were validated in experimental studies of subjects taking the test with verbal reporting protocols. Bringsjord and Schimanski (2003) used a theorem-prover to solve selected Raven's problems stated in first-order logic. Lovett, Forbus, and Usher (2007) combined automated sketch understanding with the structure-mapping analogy technique to solve problems from the Standard Progressive Matrices (SPM) test. Their system took as inputs problem entries sketched in Powerpoint as segmented shape objects and then automatically translated these shapes into propositional descriptions. A two-stage structure-mapping process was then used to select the answer that most closely fulfilled inferred analogical relations from the matrix. In contrast to these propositional approaches, Hunt (1974) proposed the existence of two qualitatively different strategies: “Gestalt,” which used visual representations and perceptual operations like continuation and superposition, Figure 1: Example problem similar to one from the Standard Progressive Matrices (SPM) test.