The Importance of Long-term Memory in Infant Perceptual Categorization

Quinn and Eimas (1998) reported that young infants include non-human animals (i.e., cats, horses, and fish) in their category representation for humans. To account for this surprising result, it was proposed that the representation of humans by infants functions as an attractor for non-human animals and is based on infants’ previous experience with humans. We report three simulations that provide a computational basis for this proposal. These simulations show that a “dual-network” connectionist model that incorporates both bottom-up (i.e., short-term memory) and top-down (i.e., long-term memory) processing is sufficient to account for the empirical results obtained with the infants. Introduction During the last decade, an increasing amount of computational research, in particular, connectionist modeling, has been devoted to the basic mechanisms underlying human categorization (e.g., Anderson & Fincham, 1996; Kruschke, 1992). Our own research has focused on developing a computational model of early infant categorization and testing that model empirically (French, Mermillod, Quinn, & Mareschal, 2001; Mareschal, & French, 1997; Mareschal, French, & Quinn, 2000; Mareschal, Quinn, & French, 2002). Quinn, Eimas, and Rosenkrantz (1993) observed a surprising categorization asymmetry in young infants between 3 and 4 months of age. After being exposed to a series of photos of cats, the infants showed greater interest in an image of a novel dog compared to a novel cat. However, after exposure to a series of dogs, infants of the same age showed no significantly different interest in either a new dog or a new cat. We hypothesized that this categorization asymmetry was due to the greater perceptual variability of dogs and to the fact that the ranges of perceptual features of cats were largely included in those of dogs. In short, when familiarized on dogs, a new cat was perceived as something very much like what had already been seen. But, when familiarized on cats, a new dog was generally outside of what the infants had been familiarized on (i.e., cats). The explanation required that very young infant categorization of these animals be essentially a bottom-up process. For reasons that are given in detail elsewhere (see, Mareschal & French, 1997; Mareschal et al., 2000) we used a three-layer, non-linear autoencoder to model this categorization asymmetry. The model predicted a reversal of this categorization asymmetry when the original variances and inclusion relationship between the two sets of stimuli was reversed by selecting a highly varied set of cats and a set of dogs with low variability. This prediction was subsequently verified experimentally with young infants (French et al., 2001). The model also predicted a disappearance of this categorization asymmetry when the inclusion relationship was removed by careful selection of cat and dog breeds for the stimuli. Again, we were able to empirically verify that the asymmetry did, in fact, disappear (French, Mareschal, Mermillod, & Quinn, 2003). This work strongly supports the view that categorization by young infants of certain types of objects (cats, dogs, horses, cars, etc.) is almost exclusively a bottom-up, statistically driven process with no contribution from prior conceptual knowledge. “Perceptual attractors” Recently, however, Quinn and Eimas (1998) reported a very interesting effect that suggests that this picture has to be modified when human perceptual features are involved. The essence of their experiment is as follows. Using an experimental design identical to that used in Quinn et al. (1993), they showed 3and 4-month-old infants images of a series of pairs of horses, followed by a pair of test images consisting of a novel horse and a human (or a fish or a car). As expected, the infants looked longer at the novel category (humans, fish, and cars) than the new exemplar from the familiarization category (horse). However, when the infants were familiarized on twelve images of humans, and then were presented with an image of a novel human or a horse (or a fish or a car), there was no significant increase in looking time for the exemplar from the novel animal categories, although there was a significant increase in looking time for the car. In other words, infants do not seem to be able to recognize an animal exemplar from a novel category after being familiarized with humans. The result was initially attributed to a lack of power of the experiment. However, two replications were done with a large number of subjects and the effect remained. Further, control experiments show that there is no discrimination bias among the exemplars used in the experiments and no spontaneous preference for exemplars of humans over instances of non-human animals. It was also suggested that the possibility of broader variance of the human category, combined with overlapping distributions of various perceptual features of the images, might have produced categorization asymmetries similar to those in Quinn et al. (1993) and