EI balance: necessary or inevitable?

TRAPP, PHILIP; ECHEVESTE, RODRIGO; GROS, CLAUDIUS

Conferencia; Bernstein Conference 2017; 2017

Spontaneous brain activity is characterized by an  asynchronous chaotic state for which otherwise large  excitatory (E) and inhibitory (I) inputs balance each  other [V96]. Such states may be obtained when the synaptic  weights scale, in the absence of strong correlations in the neural  activity, as 1/sqrt(K), where K is the number of afferent  synapses. The variance remains in this case constant [B16],  with the mean excitatory and inhibitory input contributions  individually diverging as sqrt(K). The question is then if and how  (a) the chaotic state and (b) EI balance are obtained in fully  autonomous networks, i.e. with ongoing dynamics coupled to   both intrinsic plastiticity (IP),  and synaptic plasticity, considering  here both Hebbian (Hebb.) and short term plasticity (STP) .  We examine here networks of continuous-time rate encoding neurons which adapt their afferent synaptic weight according toa self-limiting Hebbian learning rule deduced from the stationarity principle for statistical learning [E15]. The pruning of synaptic weights crossing zero allows then to study networks obeyingDale's law. The synaptic learning rules are considered in our study to be identical for excitatory and for inhibitory neurons [S17], for which have considered both a 1:1 or a 4:1 ratio. For the stabilization of the average activity level we use anintrinsic adaption rule for the threshold b=b(t) enteringthe sigmoidal [M12]. Under these conditions, we find that EI balance robustly emerges in a self-organized fashion, with inhibition closely tracking excitation (see Fig. A), without the need for an explicit 1/\sqrt(K} synaptic scaling. This effect is mainly produced by the intrinsic regulation (IP) of the mean neural activity, and is present even in the absence of synaptic plasticity, but is further increased by the effect of synaptic renormalization induced bythe ongoing Hebbian plasticity (cf. different bars in Fig. B). Further introduction of short term plasticity, which can produce fast changes in the effective connectivity between neurons, is not able to disrupt this state, making our results remarkably robust, and reinforcing the idea that EI balance is not just desirable in these networks, but almost inevitable under these conditions. Our results show furthermore that networks of rate-encoding neurons evolve, under the influence of self-limiting Hebbian plasticity, to a chaotic state fluctuating strongly on timescales of (30-50)ms. references---------- [B16]Barral, Jérémie, and Alex D. Reyes. "Synaptic scaling rule preserves excitatory-inhibitory balance and salient neuronal network dynamics." Nature neuroscience 19.12 (2016): 1690-1696. [E15]Echeveste, Rodrigo, Samuel Eckmann, and Claudius Gros. "The fisher information as a neural guiding principle for independent component analysis." Entropy 17.6 (2015): 3838-3856. [M12]Markovic, Dimitrije, and Claudius Gros. "Intrinsic adaptation in autonomous recurrent neural networks." Neural Computation 24.2 (2012): 523-540. [S17]Sprekeler, Henning. "Functional consequences of inhibitory plasticity: homeostasis, the excitation-inhibition balance and beyond." Current Opinion in Neurobiology 43 (2017): 198-203. [V96] Van Vreeswijk, Carl, and Haim Sompolinsky. "Chaos in neuronal networks with balanced excitatory and inhibitory activity." Science 274.5293 (1996): 1724-1726.