Sunday, December 6, 2015

Raising Consciousness

New methods to detect and characterize consciousness in the brain. AKA, going beyond Granger causality to understand brain dynamics.

Neuroscience and the study of consciousness has to date peeked at the living human brain via MRI, and analyzed it by correlation, trying to join slight activity signatures to various mental tasks and subjective experiences. There has been a great deal of speculation about what the neural correlates of consciousness are, complete with fanciful mathematical theories (phi, critique of phi, gamma waves, etc.) But none has been convincing, though there has been a general coalescence around some ideas- that consciousness involves fleeting coalitions among many brain areas coordinated to some degree by anatomical connection and rhythmic oscillation in the gamma band.

Getting beyond that, to a more detailed theory of consciousness, will take not only better techniques of looking at the brain, with higher time and space resolution, but also better analytical methods & theories to make sense of the vast amount of activity we see and data we already gather.

As David Eagleman illustrated in his excellent PBS show on the brain, there is a storm of activity taking place all the time, associated, as usual, with the words "billions" and "trillions". Somehow, it gets the job done, but figuring all this out from the outside requires another order of analysis. A recent paper describes new mathematical methods that appear very promising, for determining causality within a complex network like the brain.
"A dilemma is that overly realistic and detailed simulations often require a number of unknown parameters and can obscure physiological principles. It is, therefore, not straightforward to instantiate an appropriate reductive model that capture dynamical complexity and diversity across multiple brain areas. "

The problem, for a complex dynamic system, is that correlation methods alone are extremely crude. Imagine probing a computer's internal circuits with a correlation meter, and expecting to figure out its logic and mechanisms- it would be impossible. The system is non-linear, which means that a signal here can lead to negative signals there, or to vastly amplified signals, or altered patterns of waves, etc., so that simple correlations among points of activity have very limited analytical power. Then there is feedback and other network behavior that can obscure the directionality of causation.

The next step of analysis has been Granger causality, which is an extension of correlation analysis to a time series. Compare two time series of events, and if changes in one series routinely predict changes in the other, (i.e. correlate with a time lag), that supports a hypothesis of causality from one series to the other. But this method relies on the two variables being independent, and also struggles with non-linear effects, as other correlation methods do.

The new method, called convergent cross-correlation mapping (CCM) or cross-embedding, arrives from a recent ecology paper that asked what causes the unusual cycles of anchovies and sardines, each of which go through boom and bust cycles which seem anti-correlated with each other. Do they compete with each other's food sources? Is there a predator cycle that determines their abundance? The authors conclude that each population is driven by sea surface temperature, independently of the other. The causality went precisely from temperature to each species' abundance, not from either species to the other. The new method uses correlation, but through some higher-level math that better accommodates non-linear systems with feedback characteristics.

This paper showed a nice example of the method, analyzing the predator-prey relationship of the classic protists, Didinium and Paramecium. The dynamic cyclicity is very clear (in A), but what causes what? The Granger method could call it either way, depending on how you set the lag time. The cross-embedding method (rho, on the Y-axis of B) finds a higher amount of information provided by the Paramecium graph against the Didinium graph, suggesting that the predator exerts stronger top-down control of Paramecium than the reverse. In other words, the subject variable (Paramecium in this case) is more dependent on, and thus directly informative about, the driving variable (Didinium) than the reverse. A feature of this method is that it provides stronger results (suggesting true causality) the longer the time series, whence the "convergent" in its name.

A Didinium eating a Paramecium.
Population dynamic between the predator Didinium and its prey, Paramecium. 

Getting back to humans, (or thereabouts), another recent paper used this method to look at neural correlates of consciousness in macaque monkeys. They used electrodes applied directly to the exposed brain surface, getting much higher signal and resolution than when they only applied to the scalp. The monkeys were either anesthetized or conscious and behaving in various ways to test visual and other forms of perception / consciousness. The question was whether, over large distances in the whole brain, correlates of consciousness can be detected. (They also have prior work, using Granger methods.)

Electrode map and anatomical codes, top. At bottom, an example of one electrode pair and its analysis shows directionality of signaling, from #118 (red, visual cortex) to #41, (green, motor and somatosensory cortex).

The answer is that they can get significant signals from awake and behaving brains that are not only different from anestheized brains, but also reflective of an expected hierarchy of directionality and complexity, for instance that the visual system is causal towards signals in the frontal cortex during visual perception, and that the frontal areas overall have significantly higher complexity than the primary sensory processing areas at the rear of the brain.

Complexity and directionality measures among major brain areas in macaque brains under different conditions, using cross-embedding analysis. Consciousness is easily detectable, among other significant characteristics.
"Based on this method, we simultaneously characterized the large-scale cortical interaction and the dynamical complexities embedded in individual area activities. It revealed that the awake brain has a hierarchical structure of the dynamical complexity, where the frontoparietal areas had more complex dynamics than visual areas. Intriguingly, this hierarchy was linked to the directed cross-area interaction from visual to frontoparietal areas. To our best knowledge, this is the first study reporting clear cortical hierarchy in terms of dynamical complexity, as well as its relationship to the global cortical interaction. Moreover, we found that this hierarchy was universal across different behavioral/sensory conditions and disappeared after the loss-of-consciousness induced by either of two different anesthetization methods. These results indicate that this hierarchical structure is correlated with the level of consciousness rather than its specific contents reflecting perception or action."

This work is not unique in finding distinctions between conscious and unconscious states, but the improved data analysis is a significant step forward in teasing out reliable correlates of consciousness as observed from (sort-of) outside. Of course it is still a long, long way from telling what the content of that conscious state is, let alone sharing it in any rich way. And, being based on naked brain EEG, it is not clinically useful either.

  • Who is subsidizing the coal industry? You are.
  • One party is at fault.
  • And is also wrong. I mean really, really wrong.
  • And is also corrupt.
  • Update from the god & gun nuts. Update two.
  • Other things that are wrong...
  • What is your anti-capitalist stance?
  • Using dispersants on oil spills does no good.
  • This week in the WSJ: corruption at the Afghan anti-corruption committee.

No comments:

Post a Comment