Skip to the content.

Synergistic short-term plasticity mechanisms for working memory

Authors: Florian Fiebig, Nikolaos Chrysanthidis, Anders Lansner, Pawel Herman

Presentation type: Poster


Working memory (WM) is a key component of most cognitive models, and biologically detailed models of its underlying computational mechanisms become more important as neuroscientific constraints become clearer and AI models become more brain-like. Various WM models based on a diverse set of biologically plausible synaptic and neural plasticity mechanisms have been proposed. We show that these proposed short-term plasticity mechanisms may not necessarily be competing explanations, but instead yield interesting functional interactions on a wide set of WM tasks and enhance the biological plausibility of spiking neural network models of synaptic WM.

We evaluate the interactions between three commonly used types of plasticity, namely intrinsic excitability, synaptic facilitation/augmentation, and Hebbian plasticity. Combinations of these are systematically tested in a spiking neural network model on a broad suite of tasks deemed principally important for WM function, such as one-shot encoding, free and cued recall, active and silent multi-item maintenance, and rapid updating. We compare the performance and biological plausibility of a robust, integrated model against other combinations and parameterizations with fewer plasticity mechanisms.

While more reductionist models (WM explained by one particular mechanism) are theoretically appealing and have increased our understanding of specific mechanisms, we also know that mechanistic WM models need to become more capable, robust and flexible to account for new experimental evidence of bursty and activity-silent multi-item maintenance in more challenging WM tasks. We believe that spiking models should address known electrophysiological constraints from recordings and generally solve more than one task.

With this in mind, we identify principled problems of models that employ a reduced set of plasticity mechanisms to accomplish WM function. We show that more reductionist models fail to achieve comparable performance in some tasks. Parameter tuning can improve their performance, but only at the price of reduced biological plausibility or task generality. Our results indicate synergies between commonly proposed plasticity mechanisms for WM function and delineate significant tradeoffs between them.