Machine learning is one of the terms claimed by tech culture that continues to resonate, even among the learned. It is used so casually these days one might think a toaster may soon be capable of waxing philosophic on the relative merits of butter vs jam in the great toast topping debate. All one need do is search the technical literature for the words and you are quickly overwhelmed. What exactly is Machine learning? More importantly what is a Machine? What does it mean to learn? And could a Machine, once sufficiently defined, actually learn anything?
As as starting point the dictionary typically makes a fairly good guide for definitions of terms, thus we shall consult it first.
Definition of machine from Merrimack-Webster
1a archaic : a constructed thing whether material or immaterial
b : conveyance, vehicle; especially : automobile
c archaic : a military engine
d : any of various apparatuses formerly used to produce stage effects
e (1) : an assemblage (see assemblage 1) of parts that transmit forces, motion, and energy one to another in a predetermined manner (2) : an instrument (as a lever) designed to transmit or modify the application of power, force, or motion
f : a mechanically, electrically, or electronically operated device for performing a task <a machine for cleaning carpets>
g : a coin-operated device <a cigarette machine>
h : machinery —used with the or in plural1a archaic : a constructed thing whether material or immaterial
A computer would seem to fit most appropriately in either the archaic (1a) or the (f :) description. It is certainly a material constructed thing an it is an electronically operated device for performing a task. In the case of the modern computer performing many tasks, often in massively parallel fashion, at incredible speeds and with unerring precision. Therefore for purposes of this discussion a computer is a Machine defined as in 1a and (f:) above.
Next we examine the definition of learn from the same source.
1: the act or experience of one that learns <a computer program that makes learning fun>
2: knowledge or skill acquired by instruction or study <people of good education and considerable learning>
3: modification of a behavioral tendency by experience (such as exposure to conditioning)
Let’s tackle each of these one at a time and ask if a Machine (defined as above for the case of a computer) could actually learn. Definition (1:) describes learning as the act or experience of one that learns. The irony of the example case has not escaped my notice and I hope it has not escaped the readers either. While a computer/computer program can certainly facilitate learning in others can we say that it itself could learn? (1:) requires the “experience of one that learns.” Can a computer have experiences? I would think most would agree it cannot. To have an experience requires feelings and sensations of things in the world. Some might say consciousness is even a requirement to have genuine experiences. I am not prepared to go that far as non conscious beings such as insects and animals seem to experience things much as humans do. A computer cannot feel or have sensations therefore by def (1:) it cannot learn. Def (1:) also requires “one that learns.” It could be argued that to be one that learns requires a “one” that is capable of learning. The “one” in this case is implied to be a living being since non living things are not capable of learning. A computer is a non living thing therefore it cannot be the “one” required by the definition. Taken together, the inability to experience things combined with lack of ability to be a “one” shows that def (1:) could not apply to the case of a Machine/computer.
We shall skip def (2:) for the moment as it is the most problematic for my argument and move on to def (3:). This can be dispensed of easily since we have already shown that a Machine cannot have experiences. If it cannot have experiences it cannot learn as per (3:) requirement that learning is “modification of a behavioral tendency by experience.” Even if we were to allow that a Machine could have experiences the idea of a Machine exhibiting a particular “behavior” strikes us as very strange indeed. That said it is certainly possible to program a Machine to alter its output when a given input or set of inputs changes. If this change were extremely rare one could say the computer mostly tended to output X however it modified its behavior (the output, say to Y) when the input changed. In this example the input substitutes for the experience. Therefore it “learned.” An input however is not an experience. As we discussed above to have an experience requires a “one” capable of experiencing things. A Machine/computer no matter how cleverly programmed simply cannot be this “one”. Therefore def (3:) also fails to define learning in the case of a computer.
Finally we come to def (2:), which, seems the most favorable as a possible way that a computer could be said to learn something. As with def (1:) the example usage section directly references human involvement via the use of the term “people.” In and of itself this is not fatal to this option for a defintion of a computer, for as I said above it is generally accepted that many non human animals are capable of learning things. Could a Machine acquire “knowledge” or “skill” by instruction or study. Certainly a computer can be given instruction. Its program is nothing more than a set of instructions. Moreover one could argue that a computer is capable of study in the sense that its program might instruct it to, for example, “scan these 500 pages and return the word that occurs most often as output.” Viewed broadly enough the scanning could be said to be a form of study. After all human students study books by scanning their pages and recording things they are asked to record by their teachers all the time. We have so far shown that a computer can be given instruction and may even be said to study things. Can it use its instructions and the things it studies to acquire knowledge or a skill? Again if viewed broadly enough a computer could be said to have acquired knowledge through study. In the example of study I described above the returned word (output) could be thought of as a form of knowledge. After all the computer did not know the most common word until it was instructed to find it. It is more difficult to make even a stretch case for a computer acquiring a skill but the definition is an either/or proposition. The acquisition of knowledge through study alone is enough to satisfy the definition of learn in (2:).
Ultimately the entire case of the proponents of machine learning rests upon very thin ice. First we must accept very broad definitions of the commonly used terms “study” and “knowledge.” These terms are generally not understood in that broad way but I do think at least a small majority of persons could be persuaded to accept them as such. More damningly however it requires us to accept that acquiring a form of knowledge is the same thing as acquiring knowledge itself. By virtue of the fact that a form of a thing is not identical with the thing itself we are forced to reject this proposed equivalence and admit that a Computer/machine cannot learn. Therefore Machine learning is an absurd proposition and the term should no longer be used or at least no longer used in the strong sense it is now.
One might be tempted to try and extend this line of reason to attack the very possibility of artificial intelligence as well. After all a machine/computer is artificial. We have just shown that a computer cannot learn and the ability to learn is a generally agreed upon requirement for intelligence. Therefore an artificial entity/computer could not be intelligent. I do not believe this line of attack to be particularly persuasive. Intelligence is a much broader, much more iIl defined, much more highly debatable concept than learning. It is simply too easy to modify any given definition of intelligence (i.e one that does not require the ability to learn or conceives of learning in a fundamentally different manner than is typical) to evade an attack of this sort. There are any number of other more fruitful and less easy to explain away arguments better suited to that task. I will explore a few of these in a later post. Or maybe I won’t. Honestly nobody reads this shit anyway.