"Nature vs. artefacts, result vs. path, language vs. linguistics" Tobias Scheer

Speaker: Tobias Scheer (CNRS, Université Côte d’Azur, Laboratory Bases Corpus Langage (BCL))

Abstract: Adapted to the NeuroMod environment, we first look at some basic epistemological questions related to modelling, machines and learning. When somebody comes up with a machine or an algorithm that is able to achieve the same result as humans, this does not mean that human workings are matched. The goal of scientific inquiry is to find out about the workings of nature, at least for those disciplines that are concerned with natural (as opposed to artefactual) objects (but maybe also for other disciplines, see Marx on history). The successful mimicking of the results of natural workings in machines and algorithms does not tell us anything about how nature works. Beyond results, we want to know how exactly they are achieved. Google Translate produces correct results, but based on a technology that has nothing to do with any human or cognitive workings. Originally (1986 and following), connectionism was trying to mimic not only results, but also the workings of the brain (activation threshold, connection weight), but then was progressively driven away from this goal in the search of result-efficiency under the heading of Deep Learning (there is no backpropagation in the brain, or hidden layers of neurons, or supervision).

In sum: not just the result, but also the path matters. There are two completely different goals (which are of course both respectable and equally valid): finding out how nature works on the one hand, and building machines that achieve a specific purpose on the other. The latter may, and often does have, nothing to do with the former (but see bio-inspiration). In case the latter has the ambition to describe natural workings, it is one candidate among others: the way the result is achieved needs to be evaluated and compared to other candidate descriptions (backpropagation for example disqualifies deep learning for describing the natural workings of the brain). Saying "here is my machine that does the job, thus there is no need to further inquire on how things work with more complicated mechanisms" makes no sense.

We also need to agree that finding out how nature works is a goal of scientific inquiry at all. Quite surprisingly, this does not appear to be self-evident for everybody.
In a second step, we look at language, and more specifically at things that are real but cannot be heard, seen, smelled, be evidenced under a microscope or with the help of any other instrument. For example, in "Mary buys this nice house", the sequence "this nice house" has bonds that the sequence "buys this" has not: this is what linguists call a (syntactic) constituent. It is shown what kind of evidence suggests that there are these constituents, which have no physical reality but are cognitively real (and ultimately of course have a neural reality, but which for the time being we are far from being able to characterize). More examples of this kind of invisible but cognitively real linguistic objects will be discussed. Surprisingly, linguists need to struggle and argue for convincing people to believe that these objects are real: in adult science (biology, chemistry, physics), everybody knows and accepts that the existence of objects is determined by conjecture based on a remote glimpse of a trace they leave somewhere (molecules, electrons, quarks, particles etc.), but when it comes to language many people go by the materialist idea that "what you get is what you see, if I can't see an item I won't believe it exists". This shows that the division into adult and non-adult sciences is an appropriate choice of wording: linguistics, and a number of other disciplines in the Humanities for that matter, are still in their infantile development.

In a third step, the difference between language and linguistics is introduced. Many scholars study language, from all kinds of perspectives (psychologists, neurologists, mathematicians etc.), but only linguists build theories involving the kind of invisible objects discussed. My personal interest is in these linguistic theories, and more specifically in finding experimental evidence (neurophysiological, behavioural) for cognitive activity that speaks to linguistic theory. Not to language as such: to linguistic theory. Relevant examples where experimental evidence impacts linguistic theory are discussed.