Multivariate classification shows associative learning reduces working memory load
Talk presentation by William Ngiam at the Australasian Cognitive Neuroscience Society 2025 conference. This was part of a symposium titled “A model collaboration” encouraging connection between mathematical psychology and cognitive neuroscience.
Abstract
The amount of information that can be maintained in mind – or working memory capacity – is sharply limited. Longstanding debate about the nature of this capacity limit stems from disagreement on the representational ‘unit’ of working memory’; some argue for an item-based limit, whereas others argue for resource-based limits. A recent trend amongst cognitive neuroscientists is to use machine learning with neuroimaging data to decode the contents of working memory. But what is being decoded? We should not rely on simply using a “working memory task” to declare we are decoding working memory. Without a reliable theory or formal model constraining what a working memory representation is and is not, interpretation will be very difficult. I will argue that the complexity of modelling working memory is unappreciated – that there is a plurality of viable models. I will present one formal model of working memory recall in the whole-report task (Ngiam et al., 2024), which dovetails with recent successful decoding of itembased load in working memory (Thyer et al., 2022). I will then show recent empirical work exploring how this multivariate signature of item-based load changes with associative learning. The key result is that decoding differs between subjects who show explicit learning and those who do not. Learners show a reduction in item-based load (in line with ‘chunking’), but multidimensional scaling reveals this reduction is not straightforwardly explained with pure item-based models. Thus, I hope to demonstrate that progress can be made with careful iteration between formal modelling and neuroimaging.