Spatial Variability and Peak Shift: A Challenge for Elemental Associative Learning?

Recent evidence from animal research suggests that peak shift effects produced with complex pattern stimuli are influenced by the level of spatial variability over successive presentations of the stimuli. Initial attempts to explain this have relied on postulating a shift between elemental and configural representations supporting learning. Here we outline an elemental model based on contemporary associative theory that can accurately simulate this effect, and furthermore can do so without any built-in assumptions of dimensionality and without requiring some strategic shift from elemental to configural processing. Further evidence in support for the model is provided from a categorization experiment examining the effect of spatial variability with human subjects. Peak shift with an artificial dimension Peak shift is a well-documented consequence of stimulus discrimination and generalization and continues to stimulate interesting research nearly half a century after its discovery (for recent reviews see Ghirlanda & Enquist, 2003; Honig & Urcioli, 1981). Typically, subjects will be trained to discriminate between two very similar stimuli, usually by learning to respond to one (S+) and not the other (S-). Subjects are then presented with a range of stimuli that lie at successive points along the dimension to test for stimulus generalization. Peak shift takes place when the peak response rate (or accuracy of response) occurs not for S+ but for a similar stimulus that is further away from S-. By its very nature, peak shift can only be observed over a series of stimuli that have a systematic relationship with one another. While very simple stimuli with characteristics that lie along a physical dimension such as wavelength of light obviously fit this requirement, some experiments have used much more complex stimuli including morphed faces (Spetch, Cheng, and Clifford, 2004) and patterns of small abstract shapes commonly referred to as ‘icons’ (Wills and Mackintosh, 1998; Oakeshott, 2002). The latter generally employ a range of icons organized so that a logical sequence of patterns is produced, akin to a series of points along a dimension. The frequency of occurrence of each type of icon is systematically varied from one stimulus to the next so that shifting one ‘step’ along the artificial dimension means replacing some icons with new ones but still retaining a proportion from the original. Figure 1 demonstrates one hypothetical set of four stimuli of this nature (see Table 1 below for further explanation of the dimensional design). Figure 1. Examples of ‘icon’ stimuli. A possible sequence of generalization test stimuli used here. Peak shift along an artificial dimension has been of particular interest to proponents of an elemental associative learning account of generalization and discrimination. Associative theories such as that proposed by Blough (1975) have been used to model peak shift effects with great success using several assumptions about the pattern of graded activation across hypothetical sensory units which are stimulated by the presentation of training and test stimuli. While the details of Blough’s (1975) model and others like it are beyond the scope of this paper, it is worth noting that one of the key reasons for studying peak shift along an artificial dimension was to gain more control over the similarity between stimuli and the pattern of activation across units that they might stimulate (Wills and Mackintosh, 1998). Indeed, peak shift effects with icon stimuli have now been shown with both human and pigeon subjects under a variety of experimental conditions (Jones and McLaren, 1999; Livesey, 2004; Oakeshott, 2002; Wills and Mackintosh, 1998), and several elemental associative models are able to predict peak shift and related generalization effects with impressive accuracy. Generally speaking, these models rely on assumptions about the underlying dimensionality of the stimuli in order to calculate how the patterns of activation of two similar stimuli might overlap. This seems plausible when modelling generalization between stimuli that have measurable, physical similarities, even though the units are