Research

See here for a complete list of publications.

We’re trying to understand how memory works. We build computer simulations that are designed to behave like the brain areas we believe are involved in remembering. If our simulations perform and fail in the same way as real people’s brains, and generate predictions that get validated by experiment, then we can start to be confident that understanding how our simulations work could lead to a deep understanding of the workings of the mind.

In general, we tend to focus on episodic memory - the ability to recall previously experienced events, and to recognise events as having been previously experienced. But this is not to say that memory can be understood as a single, monolithic recording device. It’s a complicated system, comprising various subsidiary mechanisms, interwoven with the way we process information and influenced by what we’re trying to do. The research below is intended to look at individual pieces, while seeing them as components playing complementary roles.

Major current avenues of research include:

Hippocampal models and the complementary learning systems theory   1

Theta oscillations and retrieval-induced forgetting   2

Temporal context and distributed pattern analysis of fMRI  2

Sleep, catastrophic interference and future plans   3

See the main publications page for a complete list of published research.

Hippocampal models and the complementary learning systems theory

There are three broad brain areas that are agreed to be of especial importance in memory. Here’s how we believe them to work and inter-relate. The hippocampus (along with other subregions of the medial temporal lobes) captures high-fidelity snapshots that are maintained over the medium term. The posterior cortex learns more slowly, absorbing these hippocampal snapshots into our broad base of knowledge and long-term memory. The prefrontal cortex maintains our current context and holds items in short-term memory, perhaps by linking them to hippocampal representations, as well as having an executive role in directing and controlling other areas.

Each of these brain areas is distinct in terms of internal architecture, neuronal properties and connectivity. By remaining faithful to these biological characteristics, our models of the different areas behave differently and complement each other in various ways analogous to the real brain, as shown with studies of memory performance in college students, studies of brain-damaged patients with memory disorders and neuroimaging studies that record brain activity during recognition and recall tests.

The idea of complementary learning systems has a long history.
For additional background information, see McClelland, McNaughton and O’Reilly (1995) or O'Reilly & Munakata, 2000.

For the most comprehensive exposition of the hippocampal model, see:

Norman, K. A. & O'Reilly, R. C. (2003). Modeling hippocampal and neocortical contributions to recognition memory: A complementary learning systems approach. Psychological Review. Abstract – PDF – projects as compressed tar file – documentation for projects – blog

or, for a briefer, high-level picture of the aims and design:

O'Reilly, R. C., & Norman, K. A. (2002). Hippocampal and neocortical contributions to memory: Advances in the complementary learning systems framework. Trends in Cognitive Sciences, 12(6), 505-510. Abstract PDF

For further information on work with Andrew Mayes testing the model’s predictions in amnesic patients, see:

Holdstock, J. S., Mayes, A. R., Roberts, N., Cezayirli, E., Isaac, C. L., O'Reilly, R. C., & Norman, K. A. (2002). Under what conditions is recognition spared relative to recall after selective hippocampal damage in humans? Hippocampus, 12, 341-351. Abstract  PDF

Theta oscillations and retrieval-induced forgetting

Most recently, we’ve been working on a new learning algorithm that makes sense of the mysterious theta oscillations as a means of efficiently learning new memories while keeping them separate from old memories. This research was driven by a desire to explain a series of intriguing behavioral results - the gist is that when we reinforce a memory, competing memories are weakened (see Anderson, Bjork & Bjork, 1994). Few alternative algorithms or explanations exist that satisfactorily capture the counter-intuitive details of these findings, which have been demonstrated in a very wide range of circumstances. The hope is to apply our algorithm to explain related results in sleep, cognitive dissonance, perceptions of self and negative priming.

Norman, K. A., Newman, E. L., Detre, G. J., & Polyn, S.M. (2004). How inhibitory oscillations can train neural networks and punish competitors. Computational and Systems Neuroscience, Cold Spring Harbor. PDF

Newman, E. L. & Norman, K. A. (2003) Oscillations Drive Learning in Retrieval Induced Forgetting.  Abstracts for Society for Neuroscience annual meeting. PDF

Temporal context and distributed pattern analysis of fMRI

By bringing together both hippocampal and prefrontal models, we’ve been able to capture many aspects of free recall in humans. Our model is based on the idea that prefrontal cortex actively maintains item and task information during encoding. As the environment and task demands change, different pieces of information are gated in and out of prefrontal cortex. This continually-changing prefrontal representation can be viewed as a 'drifting context vector', which is then bound (via the hippocampus) to representations in posterior cortex. These links make it possible, at test, to recall specific item information by reinstating the pattern of prefrontal activity that was present at study.

Polyn, S.M., Norman, K.A., & Cohen, J.D. (2003, March). Modeling prefrontal and medial temporal contributions to episodic memory. Cognitive Neuroscience Convention, NYC. PDF

Having successfully modelled the behavioral findings, our aim is to go even further and push the boundaries of fMRI data analysis to test the predictions of our models. So far, we’ve had very encouraging preliminary successes analysing images from brain scans as distributed representations, rather than just focusing on the most saliently active areas. This builds on work by Haxby et al (2001), who were able to tell whether you’re looking at an image of a certain category (e.g. shoe vs house vs face), based on the widespread activity patterns in your brain. By passing the fMRI data through a neural network classifier, we’re now able to discriminate the brain states corresponding to viewing a monkey vs a human face, or even tell genders apart. This potentially opens a whole new window onto the workings of areas like the PFC, and eventually, how they might represent our current mental context.

Sleep, catastrophic interference and future plans

We’re also making promising progress thinking about the specific computational role that sleep might play in the consolidation of memory as a means of dealing with catastrophic interference, an issue that has always plagued neural networks. We’re hoping to make greater use of fMRI in the future, as well as maybe looking at EEG as a means of testing some of our predictions at a higher temporal resolution.