Inhibitory control of correlated variability in cortical networks

dynamical systems [Curto et al., 2009] or probabilistic frameworks in which variability is modelled as stochastic and correlated variability arises through abstract latent variables whose origin is assumed to lie either in unspecified circuit processes [Ecker et al., 2014, Macke et al., 2011, Pachitariu et al., 2013] or elsewhere in the brain [Goris et al., 2014, de la Rocha et al., 2007]. While these models are able to accurately reproduce many features of cortical activity and provide valuable summaries of the phenomenological and computational properties of cortical networks, their parameters are difficult to interpret at a biophysical level. One alternative to these abstract stochastic models is a biophysical spiking network, which can generate variable neural activity through chaotic amplification of different initial conditions [van Vreeswijk and Sompolinsky, 1996, Amit and Brunel, 1997, Renart et al., 2010, Litwin-Kumar and Doiron, 2012, Wolf et al., 2014]. These networks can be designed to have interpretable parameters, but have not yet been fit directly to multi-neuron recordings and, thus, their use has been limited to attempts to explain qualitative features of cortical dynamics through manual tuning of network parameters. This approach has revealed a number of different network features that are capable of controlling dynamics, such as clustered connectivity [Litwin-Kumar and Doiron, 2012], synaptic coupling strength [Ostojic, 2014], or adaptation currents [Latham et al., 2000, Destexhe, 2009], but the inability to fit the networks directly to recordings has made it difficult to identify which network features play an important role in vivo. To overcome this limitation, we used a novel computational approach that allowed us to fit a spiking network directly to individual multi-neuron recordings. By taking advantage of the computational power of graphics processing units (GPUs), we were able to sample from the network with millions of different parameter values to find those that best reproduced the activity in a given recording. We developed a biophysical spiking network with intrinsic variability and a small number of parameters that was able to capture the apparently doubly chaotic structure of cortical activity [Churchland and Abbott, 2012]. Like classical excitatory-inhibitory networks, the model generates deterministic microscopic trial-totrial variability in the spike times of individual neurons [van Vreeswijk and Sompolinsky, 1996], as well as 3 peer-reviewed) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not . http://dx.doi.org/10.1101/041103 doi: bioRxiv preprint first posted online Feb. 24, 2016;