Testing a Connectionist Model of Acquired Equivalence

Over the past decades, experimental research with animals has demonstrated that the generalisation between two stimuli is determined not only by their intrinsic properties, but by their associative history. This phenomenon is illustrated with the use of acquired equivalence tasks, which show that stimuli are treated as more similar when they come to elicit the same response as a result of conditioning. Different associative learning theories have been proposed to accommodate extant experimental findings. Mediated conditioning can explain simple forms of acquired equivalence (Honey & Hall, 1989), but is unable to accommodate findings from more complex configural acquired equivalence tasks, where stimuli are equally reinforced and nonreinforced. Pearce’s (1994) connectionist model and its extended version (Honey & Watt, 1998) can explain findings from revaluation configural acquired equivalence procedures, but cannot anticipate findings from other forms of configural acquired equivalence. Alternatively, Honey and colleagues (Honey, 2000; Honey et al., 2010) proposed a connectionist model that allows for similar inputs that share a common reinforcer to share hidden units. This model was able to accommodate a wide range of experimental findings that other models failed to explain. Honey and colleagues claimed that their connectionist model could also accommodate the results from Intra dimensional/Extra-dimensional shift tasks (IDS/EDS), which consistently find that IDS is easier than EDS, without explicitly invoking the need for attention. In Chapter 2, we tested this claim by assessing the correlation between performance in a configural acquired equivalence task and two attentional set tasks: IDS/EDS and optional-shift. Findings revealed an overall positive correlation between test performance in acquired equivalence and optional-shift, but no correlation between performance in our acquired equivalence task and IDS/EDS, in what could be seen as a challenge to Honey et al. (2010). Chapter 3 tested the effects of various outcome manipulations in configural and non-configural acquired equivalence. Experiments in this chapter revealed an enhanced revaluation and acquired equivalence effect in participants who had experienced different outcomes across training and revaluation, compared to participants who had received the same outcomes across stages. However, these group differences disappeared in a second experiment that intermixed configural and non-configural trials during the initial discrimination. A second set of experiments in this chapter failed to replicate findings from Delamater (1998), which reported a faster reversal acquisition in a group of rats that received different outcomes within stimulus modality compared to a group of rats that received the same outcomes within stimulus modality. Although Honey and colleagues carefully described the characteristics of their network, verbal descriptions could be prone to error. To the aim of qualifying the model, Chapter 4 describes a series of simulations of the experimental data presented in chapters 2 and 3 of this thesis using a formal computer instantiation of Honey’s model that was recently published (Robinson et al., 2019). This Hebbian learning network successfully simulated data from our 2-Stages configural acquired equivalence task and confirmed an enhanced acquired equivalence effect in a simulation with different outcomes across training and revaluation. This instantiation of the model was also able to accommodate findings from Delamater (1998), despite our unsuccessful attempts to replicate and extend the generality of the findings to human participants