Networks, noise and models: Reconceptualizing the brain as a complex, distributed system

The rapidly amassing body of brain connectivity research heralds a paradigm change in the neuroscience of large-scale brain systems. From a theoretical perspective, it signals a shift away from the focus on functional specialization towards connectivity-derived measures of functional integration (Tononi et al., 1994; McIntosh, 2000; Friston, 1994, 2002). Cortical “activations” — the signature of the former approach — are being supplemented with research into the human connectome (Sporns et al., 2005). Interest in the dynamics of largescale neuronal populations has emerged in parallel, adding differential equations and attractor phase flows to compliment the general linear model in the neuroimager's methodological armory (for review see Deco et al., 2008; Braun and Mattia, 2010). In essence, the brain is being reconceptualized as a complex, distributed system (McIntosh, 2004). Although this approach is arguably more intuitive, its rapid emergence raises important methodological challenges. The study of functional specialization using neuroimaging techniques was provided with a robust statistical underpinning by the general linear model in concert with Gaussian field theory (Worlsey and Friston, 1995). Whilst the theoretical underpinning of dynamic networks gains leverage from well established results in the physical and mathematical sciences (Strogatz, 2001; Albert and Barabasi, 2002), developing the statistical and analytic tools for use in neuroimaging research is a very active endeavor. NeuroImage is one of the principal conduits of this development: A quick perusal of the journal website reveals that the top 7 current most downloaded articles in NeuroImage (and 15 of the top 25) belong to the fields of networks and dynamics. The range of these articles is impressive, covering theoretical underpinnings (Honey et al., 2010), networkbased measures (Rubinov and Sporns, 2010), statistics (Zalesky et al., 2010) and post-processing techniques (Calamante et al., 2010), multivariate analyses of the functional aspects of the default network (Spreng et al., 2010), disturbances of functional connectivity in autism (Assaf et al., 2010) and the mathematical foundations of neuronal spatiotemporal dynamics (Coombes, 2010). Effective connectivity— the influence that one brain region exerts on another— is at the confluence of connectivity anddynamics andhence a core concept in this developing field. The current issue of NeuroImage contains a fascinating dialog concerning effective connectivity in brain networks and, in particular, how to make appropriate inferences from empirical data. These articles are organized around a target “Comments and Controversies” article by Roebroeck et al. (2010a) with commentaries by Friston (2001) and David (2010), followed by a reply by Roebroeck et al. (2010b). These papers are principally concerned with two approaches to effective connectivity already well known to the imaging field, namely Dynamic Causal Modelling (DCM, Friston et al., 2003) and Granger Causality analysis (Goebel et al., 2003; Valdes-Sosa,