Emergent Virtual Analytics: Modeling Contextual Control of Derived Stimulus Relations

In order to provide a behavior-analytic account of artificial intelligence (AI) operations and its predictive potential, we analyzed the extent to which a current version of a deep neural network (DNN) is able to model and forecast human learning. Human participants received individual automated training focusing on the relations among four 3-member stimulus classes where 2 of the 4 classes were composed of positive, algebraic, exponential expressions; 2 other classes were composed of negative exponential expressions. During the generalization test of novel stimulus relations, we assessed our 3 human participants in a series of 4 alternating contexts with 8 tests per context for a total of 32 tests of novel relations. When the DNN algorithm analyzed human training and generalization outcomes in terms of contextual control, clear resemblances between human and simulated participants became apparent. These findings are provocative in the sense that the simulated participants' performances were predictive of the contextual control exhibited by humans during tests of novel relations. The degree to which these procedures might be adapted to enhance human potential is discussed. The outcomes from this study are related to several of the theoretical issues detailed within our separate conceptual AI study within this issue (Ninness & Ninness,2020).