I have just finished reading The Cybernetic Brain an interesting history and philosophy of a small group of British cybernetics researchers.
The history is very interesting but in many ways the key point is the philosophy. Pickering claims that cybernetics is distinctive because it does not accept what he calls (follow Bruno Latour) a modern ontology, which is characterised by:
- Dualism between people and things. This is subtly different from the Cartesian dualism of mind and matter, though closely related. The dualism does not necessarily imply an acceptance of Cartesian dualism at a metaphysical level, but does imply a practical division of what and how we study: things in physics, chemistry, etc and people in social science and humanities. This seems to be an internalisation of cartesian dualism, even by those who do not accept it in theory.
- Representation and deliberation The purpose of science and study in general is to create representations of the world and to think about them
- Knowability In principle the world can be known in all its detail, we just need to put in more work to get more knowledge.
I’m a bit unconvinced by the use of the term “modern” in this case. As Pickering acknowledges it has too many resonances. In many cases I found myself naturally reading the term in ways that was not what was implied in its given use. Firstly the “modern” ontology has been challenged in many areas of modernism, primarily the arts (Pollock, Boulez and Joyce spring to mind) but also to some degree in science (some philosophies of quantum mechanics). So a reading equating “modern” ontology with modernism feels a bit wrong. Another natural reading when comparing work done 50 years ago with “modern” ways of thinking is to read “modern” as meaning “now”. I did that a lot without thinking, but as I will note below, that also seems very wrong. To save coining another term (and incompatibility with Pickering’s text) I will put it in quotes.
Pickering contrasts this “modern” ontology with a cybernetic one characterised by:
- Monism. people and things are not distinct. Its less clear from the book what type of monism this is. To me the implication is of a materialist ontology in which mind is an emergent property of matter. I think this covers a lot of the work, but there are also some spiritual dimensions to this (e.g. in Stafford Beer’s ideas) that I’m not sure how to place
- Non-representational and Performative. We do not represent and think about the world but figure out ways of acting in the world.
- Unknowability The world is fundamentally unknowable, primarily due to its create complexity.
I think that Pickering has identified one of the most interesting and enduring innovations of this early work in cybernetics. He himself sees this ontology as highly marginal, but I in fact I see it as a very common one in the areas of current AI, neuroscience and psychology that I interact with. In fact, my colleague Mark Bishop teaches a whole masters programme around this very philosophy (though he was trained as a cybernetician so it does all make sense).
The revival of this “cybernetic ontology” stems in large part from Rodney Brooks’ critique of the AI of the time (though it has many other streams, in psychology and neuroscience for instance). Good old fashioned AI was in many ways dualist (representation and thing itself were different); certainly was representation and deliberative and did assume that the world was knowable. Brooks contrasted this with an AI in which the world was unknowable but it was possible to construct material artefacts that could perform in the world without representation.
However, while Brooks’ technique was highly influential, his own suggested techniques never quite scaled up or became ubiquitous. So can we say that modern AI still uses a cybernetic ontology? I will take statistical and probabilistic and statistical methods as an example. One of the great recent successes in AI has been the development of statistical and probabilistic machine learning methods, what ontology do they represent. Probability is very clearly a principled mathematical method for dealing with an ultimately unknowable world, we can know it up to a certain degree of probability but no further. So these methods seem, at least, compatible with an onotology of unknowability. What about representation and performance. Many techniques are aimed at directly performing a task (e.g. classification) without a clearly understandable internal representation (e.g. Support Vector Machines). Many other methods are more hybrid, for example, Bayesian Networks do include clear representations and their purpose is to infer probability representations rather than act per se. Also feature extraction can be viewed as a way of creating representations from data (though automated feature extraction is often not humanly understandable. So the case isn’t quite clear, representation is mixed with non-representation and some deliberation is mixed with a lot of performance. What about dualism vs materialism? A lot of work in machine learning is closely linked with contemporary neuroscience (e.g. the Gatsby Unit), which does have a fundamentally materialist outlook in which mind emerges as a property of the interaction of matter. So we can say that the cybernetic ontology is alive an well, and in some ways dominant in certain domains, though still interacting with other philosophies and Pickering suggests it should in his last chapter.