Tuesday, May 12, 2009

TB Interview with Matt Davis & Gareth Gaskell: Learning and Consolidation of Novel Spoken Words

New Talking Brains feature: the TB Interview! Here's a bit about Matt Davis and Gareth Gaskell's recent paper in JoCN...

Greg Hickok (Talking Brains): Tell me about your recent paper in JoCN. How did this project come about?

Matt Davis (MRC-CBU, Cambridge): Basically we put together two things that we’d worked on separately in the years since Gareth co-supervised my PhD. I really liked some of the behavioural studies of word learning that Gareth had been doing (e.g. Gaskell & Dumay, 2003).

Gareth Gaskell (University of York): And I was interested in your fMRI priming studies with Eleni Orfanidou and William Marslen-Wilson (Orfanidou, Marslen-Wilson & Davis, 2006). The idea was to combine these two projects.

Matt: In hindsight, they fit together like pieces in a jigsaw. The fMRI study showed clear differences in the fMRI response to familiar spoken words and novel pseudowords, and that these differences didn’t change with repetition priming.

Gareth: My behavioural studies with Nicolas Dumay showed that though you could learn a made-up word (like cathedruke) really quickly, these new words don’t compete for recognition with similar existing words (like cathedral) until some time after initial learning.

Matt: To me, though, the sleep study (Dumay & Gaskell, 2007) was the really jaw-dropping result. The result is that people who learn new words at 8am (AM group) don’t show a competition effect until 8am the following day. However, people who learned new words at 8pm (PM group) show lexical competition 12 hours later.

Gareth: It’s sleep that makes the difference. Newly learned words don’t behave the same way as existing words until you’ve had a chance to sleep on them.

Matt: Which in turn explains why our repetition priming study didn’t show pseudowords turning into real words – subjects didn’t fall asleep in the scanner.

Gareth: At least not on purpose…!

Greg: So what did you do in the present study?

Matt: Since we couldn’t scan people at 8pm and again at 8am we had to teach people two different sets of words on successive days instead. Participants learned one set of new words on Day 1, another set of words on Day 2, and were tested on these two sets and a set of untrained words after training on the second day. That way we can assess effects of training with and without overnight consolidation in a single test session.

Gareth: We’d not done any behavioural experiments using this design before, but Anna Maria Di Betta showed that it worked well and produced the same lexical competition effect that we’d seen before. This is Experiment 1 in the JoCN paper. There's no lexical competition from items learned and tested on Day 2, but there is a competition effect for items learned the previous day.

Matt: Mark Macdonald and I used the same design for the fMRI study (Experiment 2). To make sure that we could separate out effects of training and lexicality we taught people real words at the same time as the pseudowords. Then on the second day of the study, we used fMRI to look at how word/pseudoword differences change due to training and overnight consolidation.

Gareth: Apart from that, though, everything else was pretty similar to the behavioural study. We kept the same training task (phoneme monitoring), and test task (pause detection) from some of the behavioural studies and combined these with the fast, event-related sparse imaging design from your fMRI work.

Greg: So how did the fMRI data come out?

Matt: We were a bit confused at first. I’d expected to see an increased response to real words compared to pseudowords, but that was non-significant. I think that’s because we didn’t give participants any meaning for the pseudowords and because of the pause-detection task we used in the scanner emphasized phonological processing.

Gareth: But, we did see lots of activation for the reverse contrast.

Matt: That’s right – the superior temporal gyrus responds more to pseudowords than to real words. And that response stays the same for items trained just before people go into the scanner. In a way, this is similar to the Orfanidou result – you can’t turn a pseudoword into a real word with short term training.

Gareth: However, just like in the behavioural study, training plus overnight consolidation makes pseudowords respond more like real words. In the STG, the pseudoword response is significantly smaller for items that have been learned and consolidated. This novelty by consolidation interaction is even more significant in other areas that respond to pseudowords such as the precentral gyrus, SMA and right cerebellum.

Greg: So you need overnight consolidation to learn a new word? That seems wrong – people can learn new words much quicker than that.

Matt: I agree completely – people can learn new words quickly. But it seems that the cortex can’t learn as fast as people. In our fMRI experiment, it’s only the day after learning that you see changes to pseudoword responses in the cortex.

On the other hand, we see lots of evidence for rapid learning in the medial temporal lobe. There are three results in our study that suggest that the hippocampus is involved in initial learning of novel spoken words: (1) it responds more to untrained items that are truly novel at the time of scanning, (2) it habituates rapidly when untrained items are repeated, and (3) the strength of both these effects is correlated with how well individual participants learn new words.

Gareth: In conjunction with other results (e.g. Breitenstein et al., 2005), we suggest that there’s two systems involved in learning new words.

The hippocampus learns quickly, but doesn’t represent new words in the same way as existing familiar words. The cortex learns more slowly and uses overnight consolidation to ensure that new words and existing words can be stored in a single set of distributed representations.

Greg: Two complementary learning systems, one fast and one slow.

Matt: Exactly! This is the same idea that Jay McClelland, Bruce McNaughton and Randy O’Reilly proposed for neural network models of memory (McClelland, McNaughton & O’Reilly, 1995). To ensure that the cortex can learn new words without forgetting old words you have to interleave old and new items during training. We think the brain achieves this by storing new words initially in the hippocampus, and then transferring knowledge into overlapping cortical representations overnight whilst people sleep. Lexical competition is one hallmark of overlapping cortical representations which explains why you need to sleep after learning in order to show lexical competition (Dumay & Gaskell, 2007).

Gareth: We’re currently revising a review paper that summarises this “complementary learning systems account” of word learning. In this paper, we attempt to explain precisely what people can and can’t learn quickly about new spoken words.

Matt: We’ll probably get shot down in flames on this last point. We make the very strong prediction that as long as there isn't task-specific repetition priming, then training can not cause the cortical response to pseudowords to resemble the response to real words. In our experiment we showed that it was only after overnight consolidation that the pseudoword response was reduced in regions like the STG that show an elevated response for untrained pseudowords. We didn’t show the reverse pattern for regions that respond more to real words, but we’d predict the same thing –changes to these responses require learning and overnight consolidation.

Greg: Sounds like there’s plenty of opportunity for you to be proved wrong. Perhaps Talking Brains readers know of some counter-evidence already.

Gareth: We’ll look forward to hearing about it!


Breitenstein, C., Jansen, A., Deppe, M., Foerster, A. F., Sommer, J., Wolbers, T., et al. (2005). Hippocampus activity differentiates good from poor learners of a novel lexicon. Neuroimage, 25, 958–968.

Dumay, N., & Gaskell, M. G. (2007). Sleep-associated changes in the mental representation of spoken words. Psychological Science, 18, 35–39.

Gaskell, M. G., & Dumay, N. (2003). Lexical competition and the acquisition of novel words. Cognition, 89, 105–132.

McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successesand failures of connectionist models of learning and memory. Psychological Review, 102, 419–457.

Orfanidou, E., Marslen-Wilson, W. D., & Davis, M. H. (2006). Neural response suppression predicts repetition priming of spoken words and pseudowords. Journal of Cognitive Neuroscience, 18, 1237–1252.

No comments: