Reinforcement accounts for transitive inference performance

Abstract

Transitive inference is the ability, given that A >B andB > C, to infer that A > C. Pigeons, rats, chimpanzees, squirrel monkeys, and humans as young as 4 years have all been shown capable of this. In this paper, simple associative learning models are explored as accounts of nonverbal transitive inference performance. A Bush-Mosteller-based model can account for transitive inference under limited conditions. A Rescorla-Wagner-based model can account for transitive inference under all conditions in the literature, but cannot account for some additional nontransitive tests. Afinal configural model can also account for these nontransitive data. The ability of this model to account for transitive inference formation in humans is also considered.The transitive inference task is a form of syllogism familiar to the ancient Greeks and introduced into psychology by Cyril Burt (1911; 1919a, 1919b. Of Burt's tests for assessing the reasoning powers of 7-year-old children, two of the three he considered best for a short test were of the following form:Tom runs faster than Jim: Jack runs slower than Jim. Who is the slowest-Jim, Jack, or Tom? (Burt, 1919a, p. 73) Burt maintained an interest in this task because it correlated well with other measures of intelligence. Piaget was interested in Burt's syllogisms for the light that they could shed on the development of basic intellectual faculties (Piaget, 1928). For Piaget, success on transitive inference syllogisms is part of a general ability to seriate stimuli, as well as evidence for the concrete operational stage of intellectual development (Flavell, 1963).The ability to seriate sets of stimuli, given only partial information about the relationships between them, has far wider generality than does the simple completion of syllogisms. For social animals to be able to estimate their rank in relation to many conspecifics without entering potentially dangerous interactions with every other member of the group must often be adaptive (Cheney & Seyfarth, 1986). Similarly, the ranking offood-item prefThis research was supported by a grant to the author from the Department of Psychology, University of Western Australia, and by grants to John Staddon from the NSF and NIMH. I am grateful to Janice Steirn and Tom Zentall for making data available prior to publication and to Howard Eichenbaum, Tom Zentall, Yin LoLordo, and two anonymous reviewers for helpful comments on earlier drafts. Address correspondence to C. D. L. Wynne, Department of Psychology, University of Westem Australia, Nedlands, Perth, WA 6009, Australia (e-mail: c1ive@psy.uwa.edu.au). Wynne, 1994 [pigeons D. Given that verbal reasoning still seems outside the reach of all except the human species, we may therefore predict that the mechanisms underlying these kinds of performances can be much simpler than hitherto assumed. This paper demonstrates how a complex cognitive ability can be the result of familiar associative mechanisms. This argument will be developed with the use of data from pigeon subjects, but the logic a...

The discovery and comparison of symbolic magnitudes

“…Though we do not claim that the algorithm is psychologically realistic, it provides a functional model that can deal with partial orderings of elements coded by high-dimensional feature vectors. The function performed by this model is consistent with empirical evidence that both animals and humans can learn simple orderings from a partial set of ordered pairs (Merritt & Terrace, 2011;Trabasso & Riley, 1975;Woocher, Glass, & Holyoak, 1978;Wynne, 1995). Moreover, its output (feature weights) can readily be translated into empirical priors for learning one-place predicates.…”

“…However, there is ample evidence that many types of animals can learn simple orderings from a partial set of pairs. For both animals (Merritt & Terrace, 2011;Wynne, 1995) and humans (Trabasso & Riley, 1975;Woocher et al, 1978), orderings are typically learned ''from the ends in'', with the extreme or ''landmark'' objects being acquired prior to those that lie closer to the middle of a continuum. In the present simulations, we trained RankSVM with ordered pairs that mainly involved the half dozen animals with the highest or lowest values on the relevant continuum.…”

Direct and relational representation during transitive list linking in pinyon jays (Gymnorhinus cyanocephalus).

“…Direct encoding can be based on a physical attribute such as size but is most commonly interpreted in terms of the associative strength that each stimulus accumulates as a consequence of its history of reward and nonreward (Couvillon & Bitterman, 1992). Strength is assumed to build up both directly through rewarded responses to the stimuli themselves (Wynne, 1995(Wynne, , 1997 and in some models, indirectly, through transfer of value from other simultaneously presented and rewarded stimuli (von Fersen, Wynne, Delius, & Staddon, 1991;Zentall & Clement, 2001;Zentall & Sherburne, 1998). Direct representation is not purely a function of associative strength, however.…”

Transforming valences through transitive inference: How are faces emotionally dissonant?

“…Critically, the A and B stimuli elicited higher happiness ratings and greater right prefrontal alpha activation than the D and E stimuli. Given that neither the A nor the E stimuli constituted "end anchors" of the series, the pre versus post differences are unlikely to have been driven by associative value transfer processes (e.g., Wynne, 1995). The high happiness ratings and FAAs reported for Stimulus C during baseline and post measurements suggest a relation between FAA and valence, motivation and/or familiarity (cf.…”

“…Second, when comparing the pre and post valences measured for the face stimuli, the valences for some stimuli (particularly Stimulus A) were shown to be enhanced in the post condition, countering common knowledge regarding generalization gradients (e.g., Cofer & Foley, 1942;Hermans, Baeyens, & Vervliet, 2013). On balance, the "enhancement" may have been driven at least partially by the "end-anchor" effect (Wynne, 1995), which simply means that the stimulus whose selection was always reinforced (Stimulus A) and the stimulus whose selection was never reinforced (Stimulus E) evoked differentially ratings/FAAs due to their differential reinforcement histories.…”