Finalist EPFL doctorate Award 2007 - Jean-Pascal Pfister
Jean-Pascal Pfister's scientific work is (according to the opinion of the jury) ranking among the best 3 % PhD theses awarded by EPFL in 2006. Thesis EPFL, no 3577 (2006). Dir.: Wulfram Gerstner.
A fascinating property of the brain is its ability to continuously evolve and adapt to a constantly changing environment. This ability to change over time, called plasticity, is mainly implemented at the level of the connections between neurons (i.e. the synapses). So if we want to understand the ability of the brain to evolve and to store new memories, it is necessary to study the rules that govern synaptic plasticity. Among the large variety of factors which influence synaptic plasticity, we focus our study on the dependence upon the precise timing of the pre- and postsynaptic spikes. This form of plasticity, called Spike-Timing-Dependent Plasticity (STDP), works as follows: if a presynaptic spike is elicited before a postsynaptic one, the synapse is up-regulated (or potentiated) whereas if the opposite occurs, the synapse is down-regulated (or depressed). In this thesis, we propose several models of STDP which address the two following questions: (1) what is the functional role of a synapse which elicits STDP and (2) what is the most compact and accurate description of STDP? In the first two papers contained in this thesis, we show that in a supervised scenario, the best learning rule which enhances the precision of the postsynaptic spikes is consistent with STDP. In the three following papers, we show that the information transmission between the input and output spike trains is maximized if synaptic plasticity is governed by a rule similar to STDP. Moreover, we show that this infomax principle added to an homeostatic constraint leads to the well-known Bienenstock-Cooper-Munro (BCM) learning rule. Finally, in the last two papers, we propose a phenomenological model of STDP which considers not only pairs of pre- and postsynaptic spikes, but also triplets of spikes (e.g. 1 pre and 2 post or 1 post and 2 pre). This model can reproduce of lot of experimental results and can be mapped to the BCM learning rule.