Network Self-organization Explains the Distribution of Synaptic Efficacies in Neocortex

Abstract

Event Abstract Back to Event Network Self-organization Explains the Distribution of Synaptic Efficacies in Neocortex Pengsheng Zheng1*, Christos Dimitrakakis2 and Jochen Triesch1 1 Frankfurt Institute for Advanced Studies (FIAS), Germany 2 École Polytechnique Fédérale de Lausanne, Switzerland Computations performed by cortical circuits depend on their detailed patterns of synaptic connection strengths. Interestingly, the distribution of synaptic efficacies in neocortex has an approximately lognormal shape. Many weak synaptic connections coexist with few very strong connections such that only 20% of synapses contribute 50% of total synaptic strength [1]. Furthermore, recent evidence shows that weak connections fluctuate strongly while the few strong connections are relatively stable, suggesting them as a physiological basis for long-lasting memories [2]. It remains unclear, however, through what mechanisms these properties of cortical networks arise. Here we show that lognormal-like synaptic weight distributions and the characteristic pattern of synapse stability can be parsimoniously explained as a consequence of network self-organization. We have developed a self-organizing recurrent neural network model (SORN) composed of binary threshold units [3]. The network receives no external input but self-organizes its connectivity structure through different forms of plasticity: additive spike-timing-dependent plasticity (STDP), the formation of new synaptic connections via structural plasticity, inhibitory spike-timing dependent plasticity, homeostatic synaptic scaling, and intrinsic plasticity of neuron excitability. Across a wide range of parameters, the network produces lognormal-like synaptic weight distributions and faithfully reproduces experimental data on synapse stability as a function of synaptic efficacy. We show that the lognormal-like weight distribution arises from a rich-get-richer mechanism induced by STDP: the probability that a synapse gets strengthened due to STDP grows approximately linearly with its current efficacy, while homeostatic synaptic scaling induces competition between synapses. Our model also predicts a power law scaling of the life times of newly established synaptic connections. Overall, our results suggest that the fundamental structural and dynamic properties of cortical networks arise from the self-organizing forces induced by different forms of plasticity.