Conflicting emergences. Weak vs. strong emergence for the modelling of brain function

Abstract

The concept of "emergence" has become commonplace in the modelling of complex systems, both natural and man-made; a functional property" emerges" from a system when it cannot be readily explained by the properties of the system's sub-units. A bewildering array of adaptive and sophisticated behaviours can be observed from large ensembles of elementary agents such as ant colonies, bird flocks or by the interactions of elementary material units such as molecules or weather elements. Ultimately, emergence has been adopted as the ontological support of a number of attempts to model brain function. This manuscript aims to clarify the ontology of emergence and delve into its many facets, particularly into its "strong" and "weak" versions that underpin two different approaches to the modelling of behaviour. The first group of models is here represented by the "free energy" principle of brain function and the "integrated information theory" of consciousness. The second group is instead represented by computational models such as oscillatory networks that use mathematical scalable representations to generate emergent behaviours and are then able to bridge neurobiology with higher mental functions. Drawing on the epistemological literature, we observe that due to their loose mechanistic links with the underlying biology, models based on strong forms of emergence are at risk of metaphysical implausibility. This, in practical terms, translates into the over determination that occurs when the proposed model becomes only one of a large set of possible explanations for the observable phenomena. On the other hand, computational models that start from biologically plausible elementary units, hence are weakly emergent, are not limited by ontological faults and, if scalable and able to realistically simulate the hierarchies of brain output, represent a powerful vehicle for future neuroscientific research programmes.

Keywords: Bayesian inference; Brain; Computational models; Emergence; Free energy principle; Integrated information theory; Multi-scale; Oscillators; Strong emergence; Weak emergence.

Figure 1

Our overview of the brain reveals structure and function at multiple spatial and temporal scales A->B. Structural connectivity can be explored, both at the macroscopic scale, i.e. regions to region, and at the microscopic level by measuring interactions between cells within and between cortical layers. A,B -> C. Each of these macroscopic and microscopic descriptions of function, forms a hierarchy, which reveals different, yet complementary information about the function of the underlying tissue. For example, functional MRI reveals a temporally ‘slow’ time-course of activity over a wide region of the brain, whereas electrophysiological measures reveal highly detailed spiking time-courses of spatially highly localised tissue. However, these two levels of description are strongly interlinked.

Figure 2

An illustrative approach to Strong Emergence (Authorized reproduction – S. Harris – Science Cartoon Plus)

Figure 3

The use of coupled oscillators to explore emergent properties of neural connectivity at the macroscopic scale. A) The generalised overview of experiments which aims to simulate from structural connectivity of the macroscopic brain, the overall functional activity of a putative neural network – such approaches often generate simulations of fMRI or MEG signals which are then correlated with empirical measures. B) The underlying dynamics at each node, can be simulated using a range of different underlying equations, here we show the simple Kuramoto oscillator system (left), which considers each node as a single reduced phase oscillator, or the more complex (right) Wilson-Cowan model, which exposes for each node, 4 separate interconnections which represent localised (microscopic) connectivity. C) The dynamics of the brain however, do not live in isolation of interactions with the external world, but are a weakly emergent property of this interaction. In our previous work ((Hellyer et al., 2017)), we demonstrated one approach for extending exploration of emergent dynamics into the behavioural space – inextricably linking the internal dynamics of the model to their emergent behavioural consequences (Portions of Figure 3, adapted with permission from (Hellyer et al., 2016)& (Hellyer et al., 2017))