The emergence of integrated information, complexity, and consciousness at criticality

Sina Khajehabdollahi, Pubuditha M. Abeyasinghe, Adrian M. Owen, Andrea SoddubioRxiv 2019 · DOI

TL;DR

This paper explores the connection between consciousness and criticality by modeling integrated information (a measure of consciousness) within a simplified brain model (the Ising model). The study finds that integrated information exhibits a phase transition at the critical point, suggesting that consciousness may emerge from a system's tendency to self-organize to a critical state, similar to other complex phenomena.

What Problem Does This Paper Address?

The paper tackles the fundamental question of how consciousness emerges in complex systems, particularly the brain. While the Integrated Information Theory (IIT) provides a framework for quantifying consciousness as "integrated information" (Φ), calculating Φ for complex systems like the brain is computationally intractable. This research aims to bridge this gap by using a simplified, yet powerful, model from physics – the Ising model – to explore how and when integrated information arises in a controlled environment. The "critical brain hypothesis" suggests that the brain operates at or near a critical point, a state of balance between order and chaos, which is known to facilitate efficient information processing. This study investigates whether consciousness, as measured by IIT, also exhibits special properties at criticality.

Methods

The researchers employed a generalized Ising model to simulate simplified neural networks. Unlike the standard 2D Ising model, this generalized version allows for arbitrary network connectivity, enabling the modeling of small, fully-connected neural network motifs (5 nodes). They generated 159 random, positive-weighted networks to explore a diverse range of network structures.

Monte Carlo simulations were used to run the Ising model across a range of temperatures, mimicking different states of the system. For each simulation, they calculated several key metrics:

  • Magnetization (M): A standard order parameter in the Ising model, representing the overall alignment of spins.
  • Integrated Information (Φ): The core measure of consciousness from IIT.
  • Magnetic Susceptibility (χ): A measure of how sensitive magnetization is to changes in temperature, typically peaking at the critical point.
  • Generalized Susceptibility of Integrated Information (χΦ): Analogous to magnetic susceptibility, this measures the sensitivity of Φ to temperature changes.

By analyzing the peaks of these susceptibility measures, the researchers identified the critical temperature (Tc) for each network and observed how Φ behaved around this point.

Datasets

Not discussed in the paper. The study uses simulated data generated from the Ising model with randomly generated network connectivities.

Key Insights and Ideas

  • Consciousness as an Order Parameter: The study proposes that integrated information (Φ) can function as a macroscopic order parameter, similar to magnetization in the Ising model, exhibiting a phase transition at criticality.
  • Criticality Maximizes Sensitivity: The research strongly suggests that integrated information is maximized in its susceptibility to perturbations at the critical point of the system. This means that at criticality, the system's "consciousness" is most responsive to changes, a property that could be evolutionarily advantageous.
  • Bridging Physics and Consciousness: The paper successfully uses the well-understood physics of the Ising model to provide a computational framework for exploring a complex concept like consciousness, offering a concrete way to test theoretical ideas about IIT.
  • Universality of Critical Phenomena: The findings align with the broader understanding that many complex phenomena, including evolution, complexity, and potentially consciousness, emerge from systems self-organizing towards criticality.

Key Results

  • The generalized susceptibility of integrated information (χΦ) peaks at the critical temperature (Tc), mirroring the behavior of magnetic susceptibility (χ) for magnetization. This indicates that integrated information undergoes a phase transition.
  • The variance of integrated information (σ²(Φ)) shows distinct transition points, with a plateau region and a second transition near the classical critical point, suggesting complex behavior of Φ around criticality.
  • The study demonstrates that changing the network's connectivity can influence the system's susceptibilities, with local minima at critical temperatures and local maxima near criticality, implying that systems near criticality might be more controllable.

Limitations and Future Work

Limitations: * The study uses small, 5-node networks, which are highly simplified representations of actual neural systems. * The networks are fully connected, which may not accurately reflect the sparse and modular nature of real brain networks. * The dynamics are limited to the Ising model's rules and a specific temperature sweep. * The peculiar behavior of integrated information splitting into two branches near the critical point in some simulations is not yet understood.

Future Work: * Explore larger and more complex network structures (sparse, modular, hierarchical, small-world) to identify a richer taxonomy of integrated information phases. * Investigate different dynamical rules beyond the Ising model. * Incorporate evolutionary algorithms to study how networks themselves evolve and how this impacts integrated information. * Integrate insights from neuroscience, complexity science, material science, and statistical mechanics to better understand complex systems like the brain. * Further explore the idea that sub-critical systems might be easier to control, potentially offering evolutionary advantages.

Summary updated

Abstract

Abstract Using the critical Ising model of the brain, integrated information as a measure of consciousness is measured in toy models of generic neural networks. Monte Carlo simulations are run on 159 random weighted networks analogous to small 5-node neural network motifs. The integrated information generated by this sample of small Ising models is measured across the model parameter space. It is observed that integrated information, as a type of order parameter not unlike a concept like magnetism, undergoes a phase transition at the critical point in the model. This critical point is demarcated by the peaks of the generalized susceptibility of integrated information, a point where the ‘consciousness’ of the system is maximally susceptible to perturbations and on the boundary between an ordered and disordered form. This study adds further evidence to support that the emergence of consciousness coincides with the more universal patterns of self-organized criticality, evolution, the emergence of complexity, and the integration of complex systems. Author summary Understanding consciousness through a scientific and mathematical language is slowly coming into reach and so testing and grounding these emerging ideas onto empirical observations and known systems is a first step to properly framing this ancient problem. This paper in particular explores the Integrated Information Theory of Consciousness framed within the physics of the Ising model to understand how and when consciousness, or integrated information, can arise in simple dynamical systems. The emergence of consciousness is treated like the emergence of other classical macroscopic observables in physics such as magnetism and understood as a dynamical phase of matter. Our findings show that the sensitivity of consciousness in a complex system is maximized when the system is undergoing a phase transition, also known as a critical point. This result, combined with a body of evidence highlighting the privelaged state of critical systems suggests that, like many other complex phenomenon, consciousness may simply follow from/emerge out of the tendency of a system to self-organize to criticality.
BibTeX
@misc{khajehabdollahi2019emergence,
  author = {Sina Khajehabdollahi and Pubuditha M. Abeyasinghe and Adrian M. Owen and Andrea Soddu},
  title = {{The emergence of integrated information, complexity, and consciousness at criticality}},
  year = {2019},
  doi = {10.1101/521567},
  howpublished = {bioRxiv},
}
Figure