Theory and Data

by Neil Rickert

This is intended as a reply to John Wilkins recent series, and in particular on Dynamics and classification redux.

When I look at the history of the scientific investigation of electricity and magnetism, what I mostly see is a struggle for ways of getting data.  I see a number of different ways used, including the deflection of gold leaf in a Leyden jar and, later, the twitching of frogs legs when connected to an electric circuit.  Eventually, the scientists came up with the methods that we use today, and they developed measuring standards that they could follow to reliably provide data.  And when I look at the physical laws associated with electromagnetism, I see that, for all practical purposes, they are those measuring standards, as abstracted to make them independent of particular measuring units.

When I look at Newton’s laws, I see something similar.  The dyne, a basic unit of force, is normally defined as the force required to accelerate a mass of one gram at a rate of one centimetre per second squared.  That is just Newton’s law f=ma being used as a measuring standard.  Newton’s law of gravity was important particularly because of its use with respect to the motion of the planets.  And there, I see it as a measuring standard that would allow one to measure the mass of the planets and of the sun.  We cannot measure those masses with a traditional lab beam balance.  Cavendish’s famous experiment is widely seen as calibrating that standard.

The examples I have given are from physics, so they might seem far from the questions of biological classification that are of particular interest to John.  And measurement might seem different from classification, though I shall try to relate them later in this post.

Linnaeus gave us a classification system that was far more refined than what had been in use previously.  Instead of saying “That bird is a large water fowl”, we could now say “that bird is a member of Linnaean designator.  The second statement is far more precise.  Because it is more precise, it carries a higher information content.  The effect of the Linneaus classification scheme, is that it greatly increased the amount of information that we could convey in biological descriptions.  That increase in the amount of expressible information is, or should be, of great epistemic significance.  And it was probably an important part of what made evolution so apparent to Darwin.

Here’s little on my relevant background.  As a child, I developed an interest in mathematics and  in science (particularly physics).  I tended to look down on biology as mere classification.  More recently, I have been studying human cognition, and have come to recognize the theoretical importance of the kind of classification done in biology, and my comments above on the epistemic significance of Linnean classification reflect that change in view.

Classification sorts items into a collection of categories.  When we measure something, say electrical current, the measurement in effect places what was measured into one of a continuum of categories.  With measurement, we have chosen to use real numbers as the labels for the categories.  But it is still a very similar activity to the classification that we see in biology.  And the epistemic significance of measurement is the same – it greatly increases the amount of expressible information.

John says that we are learning machines, and he connects that term “learning machine” with the research into machine learning that is done in AI (artificial intelligence).  I prefer to say that we are learning systems, and avoid any commitment as to whether we are machines.  There is nothing coming out of machine learning research that comes even close to the learning that we see in humans.

The underlying assumption from machine learning is that we already have data, and learning is a search for finding patterns in that data, such as might be useful in classification.  That seems to be similar to John’s view.  My alternative (and heretical) view is that the problem for a newborn child is in getting data in the first place.  I cannot see why the output from sensory receptor cells could be more than the kind of bloomin’ buzzin’ confusion mentioned by William James.  My conclusion is that the child has to invent forms of classification (though “classification” is probably the wrong word; I prefer “partitioning”) in order to have any useful information about the world.  And that would make partitioning (classification) fundamental to epistemology.

Advertisements

2 Comments to “Theory and Data”

  1. It is not clear what your “partitioning” means. An example might help. Another blog discusses the problem of sorting objects into bleggs (blue, egg-shaped, smooth things) vs rubes (red, cubical, fuzzy things). And, of course, prior to sorting is the classification problem – noticing that rubes and bleggs seem to be natural kinds (given this collection of objects). One way, in machine learning, to do classification would be to use a Naive Bayes Classifier. Ok, so much for “classification”. Let us explore what prior tasks exist, tasks which might merit the name “partitioning”.
    One way to think about the prior tasks is to imagine the invention of the labels “red” and “blue” for two clusters along the color spectrum. Similarly for oval/cubic shapes and smooth/fuzzy textures. Is this the task that you mean by “partitioning”? Or, is it the task prior to that – the partitioning of the sense data into the data bearing on color, that bearing on shape, and that bearing on texture? Or is it the cognitive step even prior to that – the one where the infant notices that the world can be partitioned into objects – each with its own more-or-less temporally stable collection of attributes?
    Of course, the infant may well invert the order, coming up with concepts of red, blue, oval, and cubical even before the concepts of color and shape. Simply notice that blue items are always not-red and vise-versa, and you can name this empirical discovery “color”.
    So, where does your “partitioning” sit in this example, or does it reside outside it?

    Like

    • Partitioning is dividing up. I am using that in terms of dividing up the world, prior to getting data. The bleggs/rubes example you mention has to do with classification based on existing data.

      In effect, I am saying that learning and scientific advance is mostly to do with getting new data, rather than with finding patterns in existing data. Thus in science, we see the design of new instrumentation (microscopes, telescopes, etc) to enable getting new data. Linnaean classification led to new data by more finely dividing up the biosphere. Newton’s laws of motion were due to the newly available data that came from distinguishing between weight and mass (a distinction that we can credit to Galileo).

      For the learning human, I’m a sense data skeptic. I am reminded of an experience of my own. I was visiting relatives, and somebody decided to go on a kangaroo hunt. From time to time, a member of the party would point to a kangaroo. I could not see anything. It turned out that I would have to learn new ways of looking in order to be able to see them.

      Like

%d bloggers like this: