Theory protection in associative learning: Humans maintain certain beliefs in a manner that violates prediction error.

TLDR: Three experiments were conducted to investigate a possible role for certainty in human causal learning, and participants appeared to attribute outcomes to cues with a comparatively uncertain causal status, in an apparent violation of prediction error.

Abstract: Three experiments were conducted to investigate a possible role for certainty in human causal learning. In these experiments, 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, 2 of these cues were trained in a causal compound, and the change in associative strength for each of the cues was compared, using a procedure based on Rescorla (2001). In each experiment, the cues differed in both their causal certainty (on the part of participants) and size of their prediction error (with respect to the outcome). The cue with the larger prediction error was always the cue with the more certain causal status. According to established prediction error models (Bush & Mosteller, 1951; Rescorla, 2001; Rescorla & Wagner, 1972), a larger prediction error should result in a greater updating of associative strength. However, the opposite was observed, as participants always learned more about the cue with the smaller prediction error. A plausible explanation is that participants engaged in a form of theory protection, in which they were resistant to updating their existing beliefs about cues with a certain causal status. Instead, participants appeared to attribute outcomes to cues with a comparatively uncertain causal status, in an apparent violation of prediction error. The potential role of attentional processes (Mackintosh, 1975; Pearce & Hall, 1980) in explaining these results is also discussed. (PsycInfo Database Record (c) 2020 APA, all rights reserved).

Figures

Table 1. The design of the experiments

Figure 4. Experiment 2 Stage 1 training data.

Figure 8. Experiment 3 Stage 2 data.

Figure 5. Experiment 2 Stage 2 training data.

Figure 2. Experiment 1 Stage 2 data.

Figure 6. Experiment 2 Test stage ratings for all single stimuli and the two compound cues.

Full text

University of Plymouth
PEARL https://pearl.plymouth.ac.uk
Faculty of Health: Medicine, Dentistry and Human Sciences School of Psychology
2019-09-26
Theory protection in associative
learning: humans maintain certain
beliefs in a manner that violates
prediction error
Spicer, S
http://hdl.handle.net/10026.1/14792
10.1037/xan0000225
Journal of Experimental Psychology: Animal Learning and Cognition
American Psychological Association
All content in PEARL is protected by copyright law. Author manuscripts are made available in accordance with
publisher policies. Please cite only the published version using the details provided on the item record or
document. In the absence of an open licence (e.g. Creative Commons), permissions for further reuse of content
should be sought from the publisher or author.

1
Theory protection in associative learning: humans maintain certain beliefs in a manner that violates
prediction error
Stuart G. Spicer, Chris J. Mitchell, Andy J. Wills and Peter M. Jones
Plymouth University
Running head: Theory protection in learning
Author note:
The experiments reported here were conducted as part of Stuart Spicer’s PhD. Correspondence
concerning this article should be addressed to Peter M. Jones, School of Psychology, Plymouth
University, Plymouth, PL4 8AA, United Kingdom. Email: peter.m.jones@plymouth.ac.uk

2
Author contributions
Stuart G. Spicer (lead author): Co-contributer to the rationale, theoretical basis and design of the
experiments. Programmed the experiments, collected and analysed the data as part of PhD project.
Wrote up the experiments.
Peter M. Jones: Co-contributer to the rationale, theoretical basis and design of the experiments.
Consulted on analysis and write up as PhD supervisor.
Chris J. Mitchell: Co-contributer to the rationale, theoretical basis and design of the experiments.
Consulted on analysis and write up as co-author.
Andy J. Wills: Consulted on analysis and write up as PhD supervisor.

3
Abstract
Three experiments were conducted to investigate a possible role for certainty in human causal
learning. In these experiments, 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 in a causal compound, and the change in associative strength for each of
the cues was compared, using a procedure based on Rescorla (2001). In each experiment, the cues
differed in both their causal certainty (on the part of participants) and size of their prediction error
(with respect to the outcome). The cue with the larger prediction error was always the cue with the
more certain causal status. According to established prediction error models (Bush & Mosteller,
1951; Rescorla & Wagner, 1972; Rescorla, 2001), a larger prediction error should result in a greater
updating of associative strength. However, the opposite was observed, as participants always
learned more about the cue with the smaller prediction error. A plausible explanation is that
participants engaged in a form of theory protection, in which they were resistant to updating their
existing beliefs about cues with a certain causal status. Instead, participants appeared to attribute
outcomes to cues with a comparatively uncertain causal status, in an apparent violation of
prediction error. The potential role of attentional processes (Mackintosh, 1975; Pearce & Hall,
1980) in explaining these results is also discussed.
Keywords: associative learning, prediction error, theory protection, redundancy effect, uncertainty

4
Prediction error is one of the most common components in theories of how associations between
cues and outcomes are learned (e.g. Bush & Mosteller, 1951; Mackintosh, 1975; Rescorla &
Wagner, 1972). According to such theories, learning occurs when humans and non-human animals
encounter surprising outcomes. Specifically, associations are formed when there is a discrepancy
(i.e. error) between an expected outcome and an experienced outcome. For example, if a person
expects that a type of food is safe to eat, but they experience an allergic reaction after eating that
food, then learning will take place and expectations about that food will update accordingly. If they
eat the same food on subsequent occasions and an allergic reaction occurs again, then further
learning will take place. This learning will continue until the asymptote of learning is reached. At
this point, no more learning can take place, because there is no longer any error between what is
expected and what is experienced. In other words, if a person correctly anticipates that a food will
cause an allergic reaction then there will be little or no learning, because there is little or no error.
Consequently, the amount of learning is greatest when there is a large error and smallest when there
is a small error. Much of the research supporting prediction error as a determinant of learning stems
from animal conditioning research (e.g. Pavlov, 1927; Kamin, 1969; Rescorla, 2001).
Not all kinds of prediction error are the same. According to Bush and Mosteller’s (1951) theory,
learning about each cue is determined by the discrepancy between the outcome that occurs and the
outcome that was predicted by that cue alone. This is referred to as individual prediction error.
However, following the observation that knowledge about one cue can interfere with learning about
another (e.g. the blocking effect, Kamin, 1969), Rescorla and Wagner (1972) suggested that
learning was instead governed by overall prediction error; that is, the discrepancy between the
outcome that occurs and an aggregate prediction derived from all the cues that are present.
Subsequently, Rescorla (2001) provided evidence that learning is governed by a mixture of
individual prediction error and overall prediction error. In other words, when more than one cue is

Frequently asked questions

Q1. What is the common component in theories of how associations between cues and outcomes are learned?

Keywords: associative learning, prediction error, theory protection, redundancy effect, uncertaintyPrediction error is one of the most common components in theories of how associations between cues and outcomes are learned (e.g. Bush & Mosteller, 1951; Mackintosh, 1975; Rescorla & Wagner, 1972). 

Q2. What is the reason why the cue with the more certain causal status needed to be judged?

the cue with the more certain causal status needed to be judged as a less likely cause of the outcome prior to the compound conditioning phase, so that a prediction error account would predict greater learning for this cue. 

Q3. What were the ten individual cue types represented on the screen?

The ten individual cue types were represented on screen as photographs of fruits: apple, banana, cherry, kiwi, mango, orange, peach, pear, plum and strawberry. 

Q4. What was the reason for the higher ratings given to overshadowed cues?

the higher causal ratings given to overshadowed cues, compared to blocked cues, meant that there was a theoretical basis for expecting less learning during the XY+ stage, since the prediction error for X would be slightly smaller. 

Q5. What is the plausible explanation for the behavior of participants?

A plausible explanation is that participants engaged in a form of theory protection, in which they were resistant to updating their existing beliefs about cues with a certain causal status. 

Q6. Why was there a theoretical basis for expecting the results of Experiment 2 to be smaller?

As outlined above, there was a theoretical basis for expecting the Experiment 2 mean differences to be smaller than the values observed in Experiment 1 (because of the higher ratings given to overshadowed cues than blocked cues). 

Q7. Why were people who had previously taken part in similar experiments excluded from this study?

People who had previously taken part in similar experiments were excluded from this study, to ensure participants were naive to the purpose of the experiment. 

Q8. What is the theory protection account for cues that are trained in compound?

In the absence of a formal theory of certainty and learning, the authors propose a simple theory protection account for cases in which two cues are trained in compound.