How knowledge affects behaviour in a model system

Lead Research Organisation: Cardiff University
Department Name: Sch of Psychology


Learning can be separated into two components: The knowledge that has been laid down in the brain as a result of experience, and the translation of that knowledge into observable behaviour. Both components - knowledge and behaviour - are fundamentally important. You might know the title of a song or the name of the person who performed it, but be incapable of providing the answer in a quiz. I might know the score of a Mozart piano concerto, but be incapable of playing the piano. Both of us have knowledge that is potentially useful, but in neither case is it immediately revealed in our behaviour. In the laboratory, detailed analysis of how behaviour changes as a result of experience paved the way for precise theories of learning, which enabled the brain's role in learning to be better understood. This is a major success story, with people from different scientific fields collaborating with one another over an extended period. However, this success was built on a simple view about how knowledge is translated into observable behaviour: Increases in knowledge result in increases in behaviour. Recent evidence shows that this view cannot be sustained. There is dramatic variation in how acquired knowledge is expressed in the behaviour of different animals from the same species given identical training. This variation cannot be reconciled with extant theories, and exposes a significant gap in our understanding: How does knowledge affect behaviour? A programme of novel behavioural experiments will test a new theory of learning. This theory provides a more sophisticated analysis of the relationship between knowledge and behaviour, including how the marked individual differences mentioned above arise. The theory is expressed in specific terms that allow the accuracy of the unique predictions that it makes to be tested. This model has broad-ranging implications for our understanding of important features of learning, which have proven to be resistant to coherent explanation.

Technical Summary

Learning can be separated into two fundamental components: The knowledge that has been laid down in the brain as a result of experience, and the translation or expression of that knowledge in behaviour. Detailed analysis of how behaviour changes as a result of experience has provided important constraints on an increasingly accurate computational model of learning, which enabled the brain's role in learning to be better understood. This has been a major interdisciplinary success story. However, this story rests on a simple yet pivotal assumption. Namely, that there is a monotonic mapping between knowledge and behaviour; with increases in knowledge being mapped onto increases in the vigour or probability of behaviour. Recent evidence shows that this simple assumption cannot be maintained: Rodents given identical training express what they have learnt in qualitatively different ways; with two behavioural measures providing the basis for drawing contradictory conclusions about which rodents have learnt most. These dramatic individual differences in how knowledge is expressed in behaviour cannot be reconciled with extant theories of learning, and exposes a significant gap in our understanding: How does knowledge affect behaviour? Our project addresses this question by developing a new computational model of learning, called HeIDI, which provides a principled explanation for both individual differences in learnt behaviour and group-level behavioural phenomena. The formal nature of the model allows clear predictions to be derived, and their accuracy to be assessed with a novel programme of behavioural experiments in a model system: Pavlovian conditioning in rodents. This model has important implications for our understanding of many behavioural phenomena, which have proven resistant to coherent explanation.

Planned Impact

Economic impacts. The project has potential economic impacts, through informing research in the pharmaceutical industry, while complementing the 3RS agenda. This industry relies on assays of learning in rodents to assess therapeutic interventions. Individual differences in learnt behaviour could interact to an unknown extent with therapeutic interventions, and contribute variability to results. There are scientific, commercial and welfare costs associated with an incomplete understanding of individual differences in learnt behaviour. Moreover, increasing fundamental knowledge and tools (e.g., evidence-based computational models) cements the UK's place as a home for the pharmaceutical industry. The project also supports the pipeline of researchers, who are directly or indirectly associated with the project, to contribute the in vivo and computational skills to industry as well as to academia. The economic impacts described above also complement the 3RS agenda: to reduce, refine and replace the use of animals in research. Most obviously, the proposed project paves the way to better understand and thereby reduce variability in learnt behaviour, and could support the refinement and reduction of animals employed in research worldwide. Ultimately, the development of more accurate computational models of whole animal behaviour will help to achieve the goal of replacing the use of animals in research.

Societal impacts. 'Behaviour change' - whether involving increasing adaptive behaviours or reducing maladaptive ones - is a key societal challenge. It arises in the clinic, classroom, home and beyond. The development of our new computational model, HeiDI, has potential impact on our understanding of aspects of behaviour change. For example, it provides the first formal analysis of why 'extinction treatments' are often of limited efficacy in producing lasting changes in behaviour. Extinction refers to the reduction in a learnt behaviour to a stimulus that occurs when the reinforcer involved in the development or maintenance of that behaviour is removed. The fact that this treatment has limited long-term efficacy has constrained its use in a variety of practical domains. HeiDI provides a clear explanation for why this is the case, and predicts what can be done to increase the efficacy of extinction as a behaviour change intervention: While removing the reinforcer will undermine the forward relationship between the stimulus and that reinforcer, it will have no impact on the relationship between the reinforcer and the stimulus. HeiDI supposes that both relationships contribute to behaviour, and anticipates that extinction procedures will necessarily be of limited effectiveness unless they also involve exposure to the US, which will undermine the efficacy of the US-CS association. The development of more effective behaviour-change treatments is a goal of practitioners that we can potentially support.

The general public has a significant interest in the behaviour of animals in their natural habitats, and the pets in their homes. Pet owners widely report that their pets respond idiosyncratically to training, and that they have temperaments. Similarly, professional animal trainers (e.g., those who work for The Guide Dogs for the Blind Association) also report marked differences in trainability across animals. Our project will shed light on the origin of some of these differences. We have sought additional funding to develop an interactive website ("Idiosyncratic animals") that will focus on the idea that animals, like people, respond idiosyncratically to objectively equivalent treatment or training. This website will include a simplified interactive cartoon version of HeiDI juxtaposed with cartoon animals (e.g., dogs) illustrating how marked differences in learnt behaviour can emerge from the interaction of very simple processes (e.g., perceived reinforcer value).


10 25 50

publication icon
Honey R (2020) Elaboration of a model of Pavlovian learning and performance: HeiDI. in Journal of Experimental Psychology: Animal Learning and Cognition

publication icon
Honey RC (2021) Higher-Order Conditioning: What Is Learnt and How it Is Expressed. in Frontiers in behavioral neuroscience

publication icon
Honey RC (2022) Associative change in Pavlovian conditioning: A reappraisal. in Journal of experimental psychology. Animal learning and cognition

publication icon
Iliescu AF (2020) Individual differences in the nature of conditioned behavior across a conditioned stimulus: Adaptation and application of a model. in Journal of experimental psychology. Animal learning and cognition

publication icon
Sánchez J (2022) Perceptual learning after rapidly alternating exposure to taste compounds: Assessment with different indices of generalization. in Journal of experimental psychology. Animal learning and cognition

Description We have developed a computational model of learning and how learning is expressed in behaviour. The model has been published and then elaborated to accommodate the results form a novel analysis of archival data. This analysis, raw data and the elaborated model have also been published. The canonical model was published in a leading journal and implemented as an open-source app accompanied by the associated computer code. This app allows the predicted effects of manipulating key features of the model (i.e., its parameters) to be readily visualised and understood. We have tested the accuracy of some of these novel predictions in a series of large-scale empirical studies. The breadth of the new model has been extended significantly, with three key theoretical papers being accepted for publication across specialist and multidisciplinary journals.
Exploitation Route The computational model and apps that we develop can be used to compare the predictions of our model with others in the field. We envisage that the theoretical analyses that we have developed will feed into those who study the neurobiology of learning and memory, as well as those with a general interest in individual differences.
Sectors Pharmaceuticals and Medical Biotechnology

Title HeiDI: A model for Pavlovian learning and performance with reciprocal associations. 
Description This is an implementation of the model described in: Honey, R.C., Dwyer, D.M., & Iliescu, A.F. (2020). HeiDI: A model for Pavlovian learning and performance with reciprocal associations. Psychological Review, 127, 829-852. It allows exploration of the parameter space for a range of fundamental empirical phenomena. 
Type Of Material Computer model/algorithm 
Year Produced 2020 
Provided To Others? Yes  
Impact None.