Theory protection in human associative learning and formally representing uncertainty about novel stimuli

Abstract

Three collections of data are presented in this thesis, with the broad aim of investigating a potential role for theory protection in human associative learning. According to theory protection, people should resist updating their existing knowledge (i.e. resist new learning), even when faced with evidence that contradicts what they already know. In other words, people should maintain established associations between environmental cues and outcomes wherever possible. Theory protection differs from typical prediction error accounts of learning (e.g. Bush & Mosteller, 1951; Rescorla & Wagner, 1972; Rescorla, 2001). According to prediction error accounts, people should update existing associations (i.e. learn) most readily when the outcomes they encounter are most discrepant with what they predict. Details about these accounts are introduced in Chapter 1, along with several other theories and phenomena that are central to the subsequent chapters. In the first set of experiments (reported in Chapter 2), human participants were initially trained with a set of cues, each of which was followed by the presence or absence of an outcome. In a subsequent training stage, two of these cues were trained together, and the amount learned about each of the cues was compared, using a procedure based on Rescorla (2001). In each experiment, the cues differed in both their prediction error (with respect to the outcome), and the confidence participants should have about their causal status. The cue with the larger prediction error was always the cue with lower confidence about its causal status. In an apparent violation of prediction error, participants always learned more about the cue with the smaller prediction error, supporting a theory protection account of learning. Participants appeared to protect their existing beliefs about cues with a known causal status, instead attributing unexpected outcomes to cues with a comparatively ambiguous causal status. The second set of experiments (reported in Chapter 3) provides further evidence of theory protection, except that the cues, outcomes and experimental scenario differed to those in the Chapter 2 experiments. Chapter 3 also includes direct testing of the theory protection account against the predictions of both Pearce and Hall’s (1980), and Mackintosh’s (1975) attentional accounts of learning. The results were not consistent with either attentional theory. The final set of data (reported in Chapter 4) includes the results of formal model fitting simulations. The findings illustrate a simple way of representing participants’ lack of confidence about the causal status of novel cues. This was achieved by allowing the initial strength of associations (between cues and outcomes) to be an intermediate value. Importantly, the best fitting initial associative strength was shown to change in line with the overall proportion of trials on which the outcome occurs. Free and open resources to support formal modelling in associative learning are briefly introduced. Chapter 5 provides a general discussion of all the findings, whilst also setting out a future programme of research, so that the theory protection account can be developed into a formal model of human associative learning.