Working Memory modelling and fast Hebbian synaptic plasticity
by
Florian Fiebig(CB/CSC/KTH)
→
Europe/Stockholm
F0 (Lindstedtsvägen 24)
F0
Lindstedtsvägen 24
Description
Working memory (WM) is among the most important components functions of cognition. It maintains information over seconds and minutes in a form which is easy to access and update. WM is updated by selectively attended external information and activated long-term memory representations. It is rather easily distractible and refresh processes are crucial for its maintenance and reactivation. WM is a core function of general intelligence and dysfunctional WM is commonly associated with attention and learning problems, language and reading difficulties.
The most common theory about the neural mechanisms of WM is that of persistent elevated activity in a recurrently connected neural network, presumably located in the prefrontal cortex, yet recent re-examination of the supporting experimental data have shown that the activity in single units during delay periods is much more variable than the activity in the computational models. Alternative theories of WM have been suggested, including those based on synaptic plasticity rather than persistent activity. I am going to present a WM model based on a simulated piece of cortex endowed with fast spike-based Hebbian synaptic plasticity.
I will highlight the rich temporal dynamics and performance of this memory model in a common word-list learning task, aiming to show how this functional neural network model captures both cognitive performance characteristics of WM, such as primacy and recency, as well as underlying neural activity characteristics, such as attractors and nested oscillations. The new model further pushes the boundaries of plasticity-based memory modelling by dropping the top-down control of plasticity, unleashing a network that stays continually plastic and thus reshapes itself through its own self-generated activity rather than only outside stimuli.
Summarily, I hope to convince you that we can and should replace the current working memory paradigm in the light of new experimental and modelling evidence.
