Laura Bella Naumann, BCCN Berlin / TU Berlin

Computational Models of Presynaptic Inhibition

Synaptic plasticity in recurrent neural networks is believed to underlie learning and memory in the brain. One practical problem of this hypothesis is that recurrent excitation forms a positive feedback loop that can be easily destabilised by synaptic plasticity. Numerous homoeostatic mechanisms have been suggested to stabilise plastic recurrent networks, but recent computational work indicates that all these mechanisms share a major caveat: An effective rate stabilisation requires a homoeostatic process that operates on the order of seconds, while experimentally observed mechanisms such as synaptic scaling occur over much longer timescales. Here, presynaptic inhibition is suggested as an alternative homoeostatic process, which does not suffer from this discrepancy in timescales. Experimental studies have revealed that GABAB receptor mediated presynaptic inhibition of synaptic transmission is triggered by excess network activity and acts on timescales of 100s of milliseconds, thus constituting a candidate mechanism for the rapid compensation of synaptic changes. To highlight the beneficial properties of presynaptic inhibition in excitatory recurrent circuits, simple rate-based recurrent network models are analysed. Presynaptic inhibition is mimicked by multiplicatively scaling down recurrent excitatory weights in response to excess population activity. Using analytical and numerical methods, it is shown that presynaptic inhibition ensures a gradual increase of firing rates with growing excitation, even for very strong recurrence. In contrast, classical subtractive postsynaptic inhibition is unable to control recurrent excitation once it has surpassed a critical value. Furthermore, presynaptic inhibition is found to boost the f-I curve for weak external inputs, while leaving it unchanged for high inputs. In a recurrent network model containing Hebbian assemblies this property enables to gradually shift from pattern completion to expression of an externally applied signal as stimulus strength is increased.

Hence, the multiplicative character of presynaptic inhibition does not only provide a powerful homoeostatic mechanism to rapidly reduce effective recurrent interactions while retaining synaptic weights, but also allows to regulate the interpretation of sensory stimuli. On the one hand, it might therefore set the stage for stable learn- ing without interfering with plasticity, and on the other hand it generates interesting properties for sensory processing.

Additional Information

Master thesis defence in the international master program Computational Neuroscience.

Organized by

Henning Sprekeler / Robert Martin