High-level description Evolution has shaped neural networks through millenia, to arrive at the most complex “computing machines” known to humanity-cortical networks. The dynamics and the function of cortical networks is notoriously difficult to capture with mathematical models. A promising and insightful approach is to study how neural networks should be designed if they were to perform their function optimally. In this paper, we build on the previous work from Deneve’s lab and address spiking neural networks derived from principles of optimal efficiency.
In particular, our focus was to understand how the structure in the recurrent connectivity drives the coding and dynamics of the spiking network with efficient coding. In such a network, we observe spontaneous bursts of synchronized spiking that are internally generated by recurrent connections. We provided a mechanistic explanation for these network-wide spontaneous bursts from recurrent excitatory synaptic interactions. We described a continuum of network states from unresponsive networks to “overactive” networks that remain in a persistent state of bursting. We found that the optimal network state is characterized by occasional bursts of activity when the sensitivity of the network to the inputs is such as to lead to the most precise neural representation of the stimulus. As bursts occur with moderate frequency, they do not disturb network’s representation, since burst-related representation is orthogonal to the representation driven by external stimuli.
Authors: Veronika Koren and Sophie Denève
Published in: PLOS Computational Biology, 2017 | to the paper