The abstraction of intervening concepts from experience with multiple input-multiple output causal environments

The purpose of this article is threefold: (a) introduce a new paradigm for investigating how intervening concepts are learned, (b) report four new experiments that provide converging evidence for the acquisition of intervening concepts, and (c) propose a simple associative learning mechanism to account for the results. The new paradigm utilizes a stimulus-response-feedback task in which subjects learn trial by trial how a multivariate set of inputs maps into a multivariate set of outputs. The first two experiments use evidence based on a principal component analysis to replicate the finding that intervening-concept learning occurs spontaneously, but only in environments that contain an intervening factor. The next experiment provides a second converging line of evidence for this conclusion by showing that subjects can use an intervening concept to make accurate inferences to a new fourth output during a transfer test. The last experiment provides a third line of evidence by showing that subjects can use an intervening concept to make accurate inferences from a new fourth input. The results are explained by a hidden-unit connectionist learning mechanism that includes both accuracy and parsimony as learning objectives.