Hi there, My research focuses on the neuroscience of learning and memory, specifically, how we learn general patterns [1,2], new vocabulary [3], and the layout of unfamiliar places [4,5]. I have also investigated what kinds of information are forgotten with time and why this happens [6]. I use a variety of research methods including computational modelling of behavioural and neuroimaging data.

I am happy to supervise any research projects that relate to my interests. If you have your own research ideas in mind, please get in touch to see if we can work something out. Alternatively, please find a list of project ideas copied below.

[1] https://doi.org/10.1371/journal.pcbi.1010566. [2] https://doi.org/10.1002/hipo.22688. [3] https://doi.org/10.1016/j.cub.2018.02.042. [4] https://doi.org/10.1162/jocn_a_01654. [5] https://doi.org/10.1038/s41386-020-00811-8. [6] https://doi.org/10.1038/s41562-020-0888-8.

Modelling individual differences in learning and generalisation

How do we learn to recognise different objects and come to understand what they mean? Theoretical models suggest that this kind of learning relies on a ‘representational hierarchy’ of brain regions, with different levels of the hierarchy being engaged depending on how difficult an object is to recognise. Learning about objects that are harder to recognise should engage more anterior brain regions at higher levels of the hierarchy. Nonetheless, these higher-level processes depend on processes at lower levels of the hierarchy working efficiently (e.g., Saksida & Bussey, 2010).

A consequence of these hierarchical models is that they predict individual differences in learning ability to follow a very particular pattern. People who are very proficient at learning about complex objects should also be very proficient at learning about simpler objects. However, the reverse may not necessarily hold - learning proficiency for simpler objects should not necessarily predict learning proficiency for more complex objects.

As well as learning to recognise the significance of individual objects, we can often generalise what we have learnt to new situations. For instance, by observing that topaz is hard enough to scratch quartz, and that quartz is hard enough to scratch gypsum, one can infer that topaz must be harder than gypsum - even if these materials have never been seen together. Some theoretical models claim that these kinds of generalisations only depend on learning processes at specific levels of the representational hierarchy. Accordingly, the only thing that should predict individual differences in generalisation performance are individual differences in the learning process at a particular hierarchical level.

In this project, we will test these theoretical predictions by asking participants to learn discriminations between stimuli that vary in complexly (see Saksida & Bussey, 2010). We will then use cross-validated factor analysis to identify whether individual differences in the ability to learn certain types of discriminations/generalisation predict individual differences in others as predicted by the theories.

Saksida, L. M., & Bussey, T. J. (2010). The representational–hierarchical view of amnesia: Translation from animal to human. Neuropsychologia, 48(8), 2370-2384.

Grokking in humans

“Grokking” refers to learning that results in a complete, indurative understanding of something. For instance, most native English speakers “grok” English grammatical rules meaning that they can effortlessly construct complex but grammatically valid sentences on the fly.

Recent machine learning experiments have highlighted that artificial neural networks can also grok particular tasks, e.g., the division of two large numbers (Power et al., 2022). We know this because, even though the neural networks are only trained on a subset (e.g., 50%) of all possible division problems, they eventually generalise what they have learnt to division problems they have not encountered before (i.e., 100%).

Surprisingly, however, grokking in artificial neural networks takes much longer than we may expect. Neural networks typically learn to solve the initial training subset very quickly but, after this point, it looks like they stop learning anything well before they can generalise what they have learnt to unseen problems. Nonetheless, as long we repeatedly present the network with the same training problems, the network will eventually ‘grok’, being able to generalise what it has learnt to all possible problems of the same type.

Little is known about when and how grokking occurs in artificial neural networks but these networks are increasingly used as models of how humans learn. In this project, we will investigate whether grokking in humans takes just as long as it does in artificial neural networks - that is, we will test whether people only learn to generalise a small number of example problems by repeatedly rehearsing them well after learning seems to have stopped. The results may validate the use of certain artificial neural networks in modelling brain function, or highlight significant differences between how humans and machines learn.

Power, A., Burda, Y., Edwards, H., Babuschkin, I., & Misra, V. (2022). Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177.

Conceptual inference during language comprehension

How do we effortlessly decode continuous streams of speech or text in order to understand the events in a story or unfamiliar concepts presented during a lecture? The ability to make conceptual inferences is thought to be central to language comprehension (Cain et al., 2001). However, little is known about the relationship between these two abilities as there are many different types of inferences. Furthermore, it is unclear what aspects of the inference-making process are important for language comprehension.

One type of conceptual inference is known as ‘associative inference’. To illustrate this, imagine that you go for a walk in your local park and see a woman walking large, fluffy Golden Retriever. The next day, you are walking in the same park and see that dog again, but this time they are being walked by a man. In this case, you may make an inference that the woman and the man belong to the same household and own the dog together.

A different, but related type of conceptual inference is known as ‘transitive inference’. To illustrate this, imagine that you witness a running race between two of your friends, Alice and Bob. Alice wins the race comfortably and so, as a result of a bet they made, Bob is forced to buy everyone a round of drinks. A couple of months later, Bob wants to try his luck again and makes a bet with Charlotte that he can beat her in another running race. This time Bob wins the race and it is Charlotte who has to buy a round of drinks. Given these events, you may infer that Alice is faster than Charlotte, even though the two of them have never raced each other before.

Both ‘associative’ and ‘transitive’ inferences are hypotheses to contribute to language comprehension in different ways. This project will test this hypothesis by attempting to show that ‘associative’ and ‘transitive’ inference abilities predict language comprehension.

Cain, K., Oakhill, J. V., Barnes, M. A., & Bryant, P. E. (2001). Comprehension skill, inference-making ability, and their relation to knowledge. Memory & cognition, 29(6), 850-859.

Spatial constraints on non-spatial learning

How do humans learn to solve complex problems for which they have little prior experience? While machine learning techniques allow artificial agents to acquire specific skills, artificial general intelligence remains elusive as we do not understand the basis of flexible thinking.

Recently, theoretical models have opened up exciting possibilities for new progress in this domain by suggesting that flexible thinking is grounded in our experience with space (Whittington et al., 2020). Just as we use spatial maps to compare distances between different spatial locations, we may use 'conceptual maps' to compare different concepts when making non-spatial decisions. For example, conceptual maps that relate the size and colour of different apples to their sugar content and bitterness may be invaluable when deciding which fruit to pick, and which fruit to avoid.

The idea that we use our spatial memory systems to learn about non-spatial problems raises an interesting possibility… our experience with 3-dimensions of space limits the kinds of non-spatial associations that we can learn about. For instance, humans are generally well-practised at quickly learning the locations of specific objects in reference to distances from 2 adjacent walls and a ceiling. However, we don’t commonly learn spatial locations in reference to more than 3 spatial dimensions. Indeed, it is extremely hard for us to even imagine 4 dimensions of space.

Does this limit also apply to conceptual maps that could, in principle, represent concepts embedded in more than 3 non-spatial dimensions? This project will test this hypothesis.

Whittington, J. C., Muller, T. H., Mark, S., Chen, G., Barry, C., Burgess, N., & Behrens, T. E. (2020). The Tolman-Eichenbaum machine: unifying space and relational memory through generalization in the hippocampal formation. Cell, 183(5), 1249-1263.

Transfer effects in sequence learning

We are able to rapidly learn complex sequences of actions and events enabling us to plan for the future and make informed decisions (e.g., Wittkuhn et al., 2022). For instance, when getting used to riding on the roads, cyclists must quickly learn action sequences that involve looking, signalling and positioning their bikes in order to safely turn into a side road.

It has been long hypothesised that after a sequence of actions and/or events has been learnt, it may be generalised to different contexts thereby avoiding the need to completely relearn sequences that are similar (e.g., Kumaran, 2012). For example, having gained experience in riding bicycles on the road, cyclists may generalise what they know when learning to drive motor cars. This may make them more efficient at acquiring safe diving behaviours (Beanland & Hansen, 2017).

Despite this hypothesis, it is not known whether humans are readily able to generalise knowledge of multi-step sequences from one context to another, and if so, what kind of signals we use to infer that a generalisation is possible. For generalisation to take place, is it enough for two sequences to be similar in structure, or do the sequences need to be composed of similar actions/events? Furthermore, it is unclear how sequence generalisation may be managed effectively when there are subtle but important sequence differences.

In this project, we will directly address these questions by as teaching participants to reproduce distinct, but similar, sequences of abstract ‘actions’ in a computer-based task.

Wittkuhn, L., Krippner, L. M., & Schuck, N. W. (2022). Statistical learning of successor representations is related to on-task replay. bioRxiv. Kumaran, D. (2012). What representations and computations underpin the contribution of the hippocampus to generalization and inference?. Frontiers in Human Neuroscience, 6, 157. Beanland, V., & Hansen, L. J. (2017). Do cyclists make better drivers? Associations between cycling experience and change detection in road scenes. Accident Analysis & Prevention, 106, 420-427.

Determinants of perceptual and semantic forgetting

Recent research has suggested that certain aspects of a memory are more stable (i.e., less prone to forgetting) than others (Lifanov et al., 2021). For instance, if you come across a tree with leaves that are unusually bright red, you may be more likely to remember the fact that you saw a colourful tree than the precise colour of the leaves. In other words, you may be more likely to remember the general semantic information about the event, than specific perceptual details.

Despite this, it is unclear why perceptual information may be forgotten faster than semantic information. One possibility is that our brains prioritise the encoding of semantic information because this information is generally more relevant to the behaviours and decisions that we make. Within a predictive coding framework, this prioritisation may result from generating stronger predictions about the semantic aspects of an event, leading to larger semantic prediction errors.

Another possibility is that our brains prioritise the ‘consolidation’ of semantic information. Consolidation is an offline process that protects memories from forgetting and interference by making them reliant on a network of highly stable brain regions.

Yet another possibility is that semantic information is less prone to interference from newly encountered information. To illustrate this, consider that having seen a tree with bright red leaves, you may subsequently come across very similar shades of red in various different contexts. These experiences may ‘interfere’ with your original colour memory in some way.

This project will attempt to shed light on which of these mechanisms contribute to the forgetting, or the lack thereof, of perceptual and semantic details. To do this, we will ask participants to remember a number of different objects, shown within colour photographs, over a 12-hour period. The objects with vary in how common they are (common vs uncommon objects), and the photographs will show these objects in a variety of different spatial orientations (typical vs atypical orientations).

Following the 12-hour retention period, we will examine how patterns of responses in an 8-alternative forced-choice recognition test change as a function of three factors: 1) how common the objects are, 2) whether they were shown in typical or atypical orientations, and 3) whether the retention interval includes a period of sleep or not. This analysis will help us to infer why perceptual and semantic features of an event may be forgotten at different rates.

Lifanov, J., Linde-Domingo, J., & Wimber, M. (2021). Feature-specific reaction times reveal a semanticisation of memories over time and with repeated remembering. Nature communications, 12(1), 1-10.

Generalising models of spatial navigation to visual memory

Models of the mammalian spatial memory system have been very successful at predicting both navigation behaviour, and the activity of spatially sensitive neurons (Bicanski & Burgess, 2018). These models have now been used to explain how we recognise visual stimuli, such as faces (Bicanski & Burgess, 2019). However, the evidence that our spatial memory systems are actually employed in visual processing is lacking.

In this project, we will test whether memory systems that help us navigate our 3-dimensional world are also involved in learning the locations of objects on a 2-dimensional visual display.

We will test this hypothesis by exploiting known properties of the mammalian spatial memory system. When animals perform a memory-guided search of a known 3-d area, experimentally manipulating environmental features such as the angles/lengths of the surrounding walls is known to affect search behaviours in a specific way (O'Keefe & Burgess, 1996; Krupic et al., 2015). We will test whether similar manipulations embedded within a 2-dimensional visual display produce the same effects.

Bicanski, A., & Burgess, N. (2018). A neural-level model of spatial memory and imagery. eLife, 7, e33752. Bicanski, A., & Burgess, N. (2019). A computational model of visual recognition memory via grid cells. Current Biology, 29(6), 979-990. O'Keefe, J., & Burgess, N. (1996). Geometric determinants of the place fields of hippocampal neurons. Nature, 381(6581), 425-428. Krupic, J., Bauza, M., Burton, S., Barry, C., & O’Keefe, J. (2015). Grid cell symmetry is shaped by environmental geometry. Nature, 518(7538), 232-235.