I’ve seen this quote excerpted in parts before, but not the whole thing:
I repeat, feedback is a method for controlling a system by reinserting into it the results of its past performance. If these results are merely used as numerical data for the criticism of the system and its regulation, we have the simple feedback of the control engineers. If, however, the information which proceeds backward from the performance is able to change the general method and pattern of performance, we have a process which may well be called learning.
– Norbert Wiener, The Human Use of Human Beings
It is a strange distinction Wiener is trying to make here. First, Wiener tries to make “numerical data” a simple special case, and equates control as the manipulation of numerical data. However, he doesn’t contrast numbers with something else (presumably non-numerical) which can “change the general method and pattern.” Taking it from the other direction, he implies that mere control engineering cannot accomplish “learning.” That is, from numerical data and “criticism of the system” we cannot change how the system works. By Wiener’s lights, pretty much all of the work in mathematical control and machine learning would be classified as control.
I am, of course, missing the context in which Wiener was writing. But I’m not sure what I’m missing. For example, at the time a “control engineer” may have been more of a regulator, so in the first case Wiener may be referring to putting a human in the loop. In the book he makes a distinction between data and algorithms (the “taping”) which has been fuzzed up by computer science. If this distinction leads to drawing a line between control and learning, then is there a distinction between control and learning?
5 thoughts on “Wiener on control versus learning”
Think of the difference between a fixed reflex (or what the psychologists call an “over-learned” habit), and acquiring a new habit. In the former case, the feedback alters what you do in the immediate future, but it doesn’t have any impact on how you’ll perform on the next trial; the system resets, the same transducer is ready to run again. In the second case, of learning, part of the response to feedback is not just alter immediate behavior, but to change how all future inputs (including future feedback signals…) will be processed. If you are using, say, a neural network for your controller, in the first case you just have a feedback channel as input to the network, but the weights (etc.) are fixed; in the second, you are running something like back-propagation to tune the weights. You could of course always embed the learning process in a larger, invariant system, but it seems like a tolerably clear distinction.
I see the distinction (changing the weights is a succinct way to describe it), but the way “control” is thought of now doesn’t seem to respect it. Sure, there’s the “what is the best linear controller to stabilize my plant,” but that’s just the first class in control systems.
Maybe in modern lingo we would say that “adaptivity” (changing the weights) is what distinguishes “classical control” from “modern control/learning.” This does make Wiener look more forward-thinking!
Maybe in modern lingo we would say that “adaptivity” (changing the weights) is what distinguishes “classical control” from “modern control/learning.”
I think that’s exactly what he’s getting at.
This does make Wiener look more forward-thinking!
Perhaps because many control engineers in the 40s, 50s and 60s and (proto-)computer scientists read Wiener and said “Hey, that would be really cool/useful”?
Maybe he refers to responding to feedback with changing the structure of control or of the system itself. Although in the context of some modern applications it is unclear what does structure mean. For example, if we consider non-parametric models used in visual feedback control, they are in a way always changing the structure to respond to the latest insight from the environment…
Of course, he could be referring even to a meta control problem, where you would change the performance goals according to the observations. Maybe you observe that surrogate goals could achieve better performance in competing metrics…
Or Wiener was just musing at replacing the basic LQR structure of control at the time :).
I think my initial response was because classical control feels “old” and modern control has become “learning” in the sense that Wiener makes. But it’s still called “control,” which makes reading distinctions like this confusing.