An unsupervised life

By Graham HarrisGraham Harris v3

In this blog I want to pull some threads together that I have written about previously and I will to try to make some bigger-picture connections. I am going to link cybernetics to embodiment and unsupervised learning in living systems.

Once we make the philosophical step of moving from 1st order cybernetics (rationalist and naïve realist science) to 2nd order cybernetics (with the observer in the loop) then we start down a path with significant consequences. Not, I might add, a complete move away from science (to head off the immediate scientific criticisms of subjectivity and muddled thinking).

 

I am just making a case for a new kind of science that makes a move away from the search for efficiency, equilibrium, optima and universal laws towards by embracing contingency, effectiveness, persistence, robustness and commonality of process. We move from rationalist “physics envy” to an unsupervised life.

I had not seen the bigger picture until I read some recent papers by Tom Froese and others on the continuum of thought that runs from 1st order, to 2nd order cybernetics and thence to embodiment and phenomenology [1].

So what is all this about? When we make the move to include the observer in the loop (which is the first move away from received science) we can then ask, “well, what kind of observer?” This is how the biology and cognitive powers of the agent become important. It was Varela who seems to have first made this “experiential turn” from cybernetics to the biology of cognition and to cognitive science.

If we reject the traditional view of cognitive science that Cartesian dualism is the order of the day (i.e. that the rational mind is a separate from the body and from the world and it exists to make universal data rich representations of the world) and accept, instead, that organisms and their nervous systems are embodied, enactive, embedded in the world and have extended networks and connections, then we end up asking quite different questions and come to some different conclusions.

Edmund Husserl and Maurice Merleau-Ponty [2] asked these different questions to develop a new philosophy that they termed phenomenology: to begin the systematic study of “the structure, qualities and dynamics of first-person experience” [3]. They let the observer into the loop and began to ask how that observer observes and what this means. So the experiential turn took us from cybernetics to a form of phenomenology.

Now phenomenology is, itself, not totally divorced from science – although many would see it that way as being too subjective. There are two important connections to received science one is old; one is new.

R.A. Crowson wrote a note in the journal Nature in 1969 quoting Husserl at length and linking phenomenology to science. He linked it however, not to a natural philosophy, which seeks “certain universally applicable mathematical formulae or laws”, but instead to a more contingent form of science, which we would know as natural history [4].

Crowson decried the ascendancy of natural philosophy (the Anglo-centric view of science I might say) and deplored the prestige given to its technological applications. He called instead for a more inclusive view that would “provide valid intellectual bases for natural history and natural philosophy, aesthetics and morals, psychology and sociology, progress and conservation”. That, he wrote, was “still an unattainable dream for us”. [It is worth pointing out here in parentheses that this “more inclusive view” has long existed in the countries of the Romance languages and much of it has not been translated into English.]

Perhaps, only now, nearly 50 years later in the Anglo-scientific world, some progress is being made. This is where we must turn to developments in cognitive science and artificial intelligence,

I have just been reading Andy Clark’s book “Surfing uncertainty: prediction, action and the embodied mind” [5]. This elegantly brings together new ideas in cognitive science and artificial intelligence in the context of embodiment, phenomenology and unsupervised learning. The breakthrough for me is that the book explains the science of the process that underlies what we might call natural history. Yes, the world we inhabit is complex, contingent and uncertain – even absurd [6] – and no laws entail, [7] but underlying it is a regular process, that of unsupervised learning which applies to life in general [8].

Clark lays out a framework based on embodiment which, as I have written before does not involve organisms making and using data rich representation of the world (in a Cartesian sense) Instead, Clark’s framework involves organisms using simple heuristics and past experience – as Bayesian priors – merely to “keep in touch with the world” (in an embodied sense) [9].

From brains to ecosystems and society life self-organises by unsupervised correlative learning (merely reinforcing what works) and, in its network structures in brains, ecosystems and society, carries forward a set of expectations and predictions in the form of simple, hierarchical priors.

Organisms generalise by finding and building upon regularities in past experience. These priors then propagate downwards through the networks of interaction at all levels and are then updated on the basis of errors and exceptions propagated back up. Unsupervised learning merely amplifies what works and throws out what does not. Together these two processes are a form of deep learning.

Once we realise that this is a universal form of networked biological organization that runs across many scales – “brains evolve and adapt like large ecosystems” [10] – then we have an entirely new way of thinking about life, contingency, chance and uncertainty. This is Morin’s recursive world of genetic, biological, ecological and social re-organisation in its full complexity [11].

Science and engineering (the universalist, rationalist and naïve realist Anglo-centric view) seeks explanation, representation and universal laws. Scientific management seeks optimization, efficiency and the minimisation of risk. Action follows prediction.

Life is different. Life is embodied and embraces uncertainty to learn on the fly. Life is therefore not efficient nor is it optimal – it carries a lot of distributed redundancy in its networked structures – but it is effective, persistent and resilient. Life is non-stationary – always on an uncertain trajectory to somewhere new. No laws entail and there is no ground.

By constantly updating its internal models (simple heuristics) in the face of change it is able to evolve and reflexively change in response to an environment which it is part of and which it, itself, partly creates. So it is a universal recursive learning process that creates the Crowson’s contingency and the uncertainty that we see in life. Embracing contingency and uncertainty by allowing the observer into the loop (rather than trying to eliminate it by taking a detached 3rd person view) leads us down a path that looks more like his natural history.

So we have a dilemma here. On the one hand we would like to have a predictable world; on the other hand we must acknowledge its uncertainty and absurdity. We have two complementary views.

Let me take you back to the early BBC Reith Lectures – to the ones given by Robert Oppenheimer in 1953 on “Science and the common understanding”. Oppenheimer was head of the Los Alamos laboratory during the war and was one of those credited with being the father of the atom bomb.

It seems people were expecting some “great truth” about the bomb from Oppenheimer in his Reith Lectures but, instead, he wisely chose to talk about the new science of atomic physics and its relationships to society. In his lectures he talked about the wider implications of the physical concept of complementarity (that in quantum theory particles can be both a particle and a wave). Oppenheimer pointed out that neither explanation is entirely correct and that we have to hold in our minds two mutually contradictory viewpoints [12].

There are many undecidable propositions in this life that are complementary – and they are lessons for the ways we run our lives. There is, as Oppenheimer observed, a complementarity between being actors and observers, between freedom and constraints. There is complementarity between the universal and the local, between society and the individual, between natural philosophy and natural history.

Life has found a way of balancing the complementarity between the learning individual and system and the constraints of being on this one planet. Oppenheimer called for an open society in which there are no God given rules – through diversity and debate, and without absolutism, we achieve a richer understanding. Complementarity is not a problem to be solved; through debate and learning, and through trial and error, success and failure we develop and evolve a richer world.

We must be accepting of limited knowledge; we have no perfect predictive powers. If we work on the basis of simple heuristics that “keep us in touch with the world” it requires us to also learn on the fly, to be adaptive, open and nimble. To quote Hannah Arendt “we are not Gods”; we are but limited beings in an uncertain world.

[1] See Froese T. (2010) From cybernetics to second-order cybernetics: A comparative analysis of their central ideas. Constructivist Foundations 5(2): 75-85. http://constructivist.info/5/2/075 ; Froese, T. (2011) From second-order cybernetics to enactive cognitive science: Varela’s turn from epistemology to phenomenology. Systems Research and Behavioral Science, 28, 631-645; Stapleton, M and Froese, T. (2016) The enactive philosophy of embodiment: from biological foundations of agency to the phenomenology of subjectivity. Chapter 8 in “Biology and subjectivity” Eds M. Garcia-Valdecasas et al. Historical-Analytical Studies on Nature, Mind and Action, Vol 2, p.113-129; Vörös S., Froese T. & Riegler A. (2016) Epistemological odyssey: Introduction to special issue on the diversity of enactivism and neurophenomenology. Constructivist Foundations 11(2): 189–204. http://constructivist.info/11/2/189

[2] Merleau-Ponty, M. (2013) Phenomenology of perception. Translated by Donald A. Landes. Routledge, London. (First published in French 1945)

[3] See Gallagher, S and Zahavi, D. (2008) The phenomenological mind: an introduction to the philosophy of mind and cognitive science. Routledge, London.

[4] Crowson, R.A. (1969) Science and phenomenology. Nature, 223, 1318-9

[5] Clark, A. (2016) Surfing uncertainty: prediction, action and the embodied mind. Oxford.

[6] Jean-Paul Sartre (1946) L’existentialisme est un Humanisme, Éditions Nagel, Paris

[7] Kauffman, S. (2013). Evolution Beyond Newton, Darwin and Entailing Law: The Evolution of Complexity in the Biosphere. In “Complexity and the Arrow of Time,”Eds. Lineweaver, C.H., Davies, P.C.W. and Ruse, M., Cambridge

[8] Power, D.A., Watson, R.A. et al. (2015) What can ecosystems learn? Expanding evolutionary ecology with learning theory. Biology Direct, 10:69 DOI 10.1186/s13062-015-0094-1

[9] Alva Noë (2004) Action in perception, MIT press

[10] Adams, P. (1998) Hebb and Darwin. Journal of Theoretical Biology, 195, 419-438

[11] Edgar Morin (1977) La Méthode, Vol 1, La nature de la nature. Published in English as “Method: towards a study of humankind, Vol 1, The nature of nature”, translated by J.L.R. Bélanger, Peter Lang, publishers New York. (1992) American University Studies: Ser 5, Philosophy, Vol 111.

[12] A useful commentary on Oppenheimer’s 1953 Reith Lecture by Brian Cox can be found on the BBC web site. http://www.bbc.co.uk/programmes/p05hctvq

One thought on “An unsupervised life

Leave a Reply to pascal perez Cancel reply

Your email address will not be published. Required fields are marked *