Is Learning Just About Connecting (Meaningless) Dots?

This essay responds primarily to the question, Can Artificial Intelligence (AI) learn? In the next paragraphs, I do not try to give a concrete answer but analyze both postures. If we say learning is more than a mechanical process and depends on intelligence, then we will dig up the question: Can machines Think? from Turing, and compare it with the posture of Searle where he argues that a Turing machine will lack semantics (meaning) and intentionality. Then the discussion is not if the current AI programs can learn, but if it's possible to construct Artificial Intelligence in the future (maybe not in a Turing Machine), and by that also to discuss if it's possible for artificial agents to learn, if we think they couldn't, what could be lacking?

1. Turing and the origins of Machine Learning.

"May not machines carry out something which ought to be described as thinking but which is very different from what a man does?" (Turing, 1950, p.435)

In 1936, Turing showed for the first time what at the end of the 20th century would mark a new technological revolution, "The Computing Machine" or what we currently know as computers. In this same article, he also asks "What are the possible processes that can be carried out in a computing machine?" (Turing, 1936, p. 249). With this, Turing not only proposes a universal abstract machine capable of solving any (computable) problem but also sets a limit to what such a machine can do by proving that there is no such thing as a Turing machine capable of determining whether a program may or may not return a result in some finite number of steps, this problem is more commonly known as The halting problem.

His questioning led him to think about the limitations of machines, beyond The halting problem. What can be computable? What aspects of what we know can be replicated by a Turing machine? In short, any mechanical process. The importance of a universal machine, as he writes it:

We do not need to have an infinity of different machines doing different jobs. A single one will suffice. The engineering problem of producing various machines for various jobs is replaced by the office work of 'programming' the universal machine to these jobs. It is found in practice that LCMs can do anything that could be described as a 'rule of thumb' or 'purely mechanical'. (Turing, 1948, p.111)

Beyond producing a universal machine that we could program to perform any task, Could we train (educate) a machine so that it can find the solution to any problem by itself? Self-programming (modify itself)? Can intelligence be programmed? 75 years ago it made no sense to ask ourselves that, but today it is a question that surely many of us have asked ourselves. As Turing (1948) proposes, in order to investigate intelligence, it is necessary to abstract the way in which we develop this capacity. If a human develops his intelligence through learning then, What are the possible ways that we could mimic, abstract, or program our learning capacities? And to what extent is simulation = duplication?

Surely if we think learning is purely mechanical then we could say it can be programmed and simulated in a Turing machine. And perhaps if the current AI program doesn't model exactly learning we will find a way to program it. But, What are the current AI programs missing to agree that machine learning is the same as human learning?

2. Human Machines and Syntactic Information

One of the things that I found most interesting in Turing writings is this constant analogy between humans and actual digital computers as we know them today. What we missed is that in principle, humans were the first to be computers. They were the ones who made possible the discovery of Halley's comet, and we had computer factories where the computers were not machines but people. The job of a human computer was to carry out lots of mathematical operations in the most mechanical way. (Grier, 2001).

So during Turing's articles (1948; 1950), we found the definitions of what would constitute a paper Machine. "A man provided with paper, pencil, and rubber, and subject to strict discipline, is often a universal machine" (Turing, 1948, p.113). This definition implies that all sorts of algorithms can be done by a paper machine with sufficient time and life. And with the historical background, it may make us think that Turing Machines were an attempt to abstract in the most efficient and productive way, the capacities of the current human computers at the time.

If we can make the analogy between a human and a machine, it would be easy to extend the analogy between intelligence and computers. Maybe we do not need to replicate the body of a human-machine with all of its functions to replicate its intelligence. Duplicating with exact precision all our nervous systems would be too hard and too costly. This is something that Turing (1948) writes;

Certainly the nerve has many advantages. It is extremely compact, does not wear out (probably for hundreds of years if kept in a suitable medium!) and has very low energy consumption. Against these advantages the electronic circuits have only one counter-attraction, that of speed. This advantage is, however, on such a scale that it may possibly outweigh the advantages of the nerve. (Turing, 1948, p.117)

After this, Turing started thinking about the possible ways in which we could start creating a learning program. The problem is that the learning program that we want to create, as Searle (1980) pointed out, is just a set of symbols and syntactic data, with no connection to meaning. As an example, if you copy a sheet of paper with Chinese Symbols in a Word File, then by just the sheet of paper you wouldn't have any idea of what you're writing. You just would have copied symbols.

Then one of the major differences between Artificial Intelligence (AI) and Natural Intelligence (NI) is that AI only processes data (just symbols, patterns, binary, no meaning) and at some point, in our Natural Intelligence we start to process information (data with meaning). The problem can be then described as, How is it that we start to attach meaning to data? How do we construct meaning from our environment? (Floridi, 2004)

We can make the program extremely complex as we want, even imagine that we could formulate a program that simulates all of the synapses in the human brain. The problem is that this program is just pure syntactical information with no semantics and the connections can be just seen as an input and output (binary).

As an example of this, let me try to explain what I call the Chinese Room Experiment 2.0 (the one with water pipes). Imagine that you are in a room with water pipes, and someone gives you instructions to move and manipulate the water pipes (the program). On the other side of the room, there is someone who throws Chinese symbols into the water pipes (input). You start the process by moving and connecting the water pipes (make this the relation of the neural connections). By this system at the end, the water will flow and someone else on the other side of the room will receive the Chinese symbols and will think you know Chinese, but you know nothing! You just did a mechanical process with zero information about the meaning of the Chinese Symbols. (Searle, 1980)

The problem is when we think of our neural connections, as just purely mechanical, How do we construct meaning ourselves? Maybe computers just have binary data, but how different are a graph and our neural connections? Can it be the difference between a discrete state and a continuous state?

What we can conclude about this experiment is that semantics is not syntaxis. And we may think that we could construct a robot with the most robust and sensible sensors with a processor which is extremely fast. But if the program that sustains this robot is equivalent to a Turing Machine then, it falls from the same nature, it's pure syntactical with no meaning. (Searle, 1980)

As long as the program is defined in terms of computational operations on purely formally defined elements, what the example suggests is that these by themselves have no interesting connection with understanding. (Searle, 1980, p.418)

The thing is that maybe we cannot make a program that runs in a Turing machine because that implies it lacks semantics. But if our brains are just biochemical processes, something that can be just purely mechanical, then how do we acquire the sense of semantics? Is it something intrinsic to the natural nature of the brain? Do we need carbon to create an intelligent agent? In order to create artificial intelligence, do we need to create artificial life? And if we create artificial life, how could it be proven that these agents think?

Floridi (2004) selects this question, as the "Turing Problem: Can (forms of) natural intelligence be fully and satisfactorily implemented nonbiological?" (p.568) The problem with these sorts of questions is that anything can be an information process with a certain level of abstraction. And by this if our thinking is just a form of information processing, where is the line, when could we say if something has or does not have consciousness?

3. Imitation = learning?

What a Turing machine can do so well is the imitation game. A mimicking process. I myself have fallen to distinguish if I am messaging with a human or a chatbot (a contemporary form of the Turing Test). And although this is an anecdote, we are approximating that level of uncertainty. In an informational environment, in an infosphere (Floridi, 2010), what are going to be our tools to distinguish between an AI output and a human creation?

Trying to prove to someone 75 years ago, that at some point the question; Can Machines Think? will be something of popular doubt, would have constituted to give a clear explanation about the possible ways in which we would see something like that happening in the future. And in my opinion, that's what Turing tries to explain with the Imitation Game. The argument for which I think Turing tries to pursue in Computing Machinery and Intelligence (1950) is to try to explain how in the future, we would have to pass the Turing Test for the ones who did not believe it. At the beginning of the text he mentions;

The short answer is that we are not asking whether all digital computers would do well in the game nor whether the computers at present would do well, but whether there are imaginable computers that would do well. (Turing, 1950, p.436)

If we had (or will) pass the Turing Test? What is next? What can we say AI lacks in intelligence or learning?

For Searle (1980) what an AI lacks is semantics and intentionality, something that seems to him a product of the brain. For him, if AI produces intentionality it must replicate the causal powers of the brain. But what are these causal powers? What Dennett replies in the same paper, is that by this Searle could be incorporating a form of Élan vital, comparing the intentionality (or consciousness) to something intrinsic to nature, and with something that only nature can replicate. But I don't think Searle would attach life to consciousness, the truth is that we don't know exactly how our consciousness works and with that, How could we know what is needed for consciousness? In order to know if something is a duplicate of intentionality we must first know what is intentionality and how it works in the human body. We have artificial hearts that can function as well as the heart, but could we create an artificial brain?

As Turing (1950) stated, there is a kind of paradox when trying to find consciousness. As Dennett's reply to Searle (1980), the Chinese Room may fall from being sophistry;

This argument appears to be a denial of the validity of our test [The Turing Test]. According to the most extreme form of this view, the only way by which one could be sure that a machine thinks is to be the machine and to feel oneself thinking. One could then describe these feelings to the world, but of course no one would justify taking any notice. Likewise according to this view the only way to know what a man thinks is to be that particular man. It is in fact the solipsist point of view. (Turing, 1950, p.446)

But then at the end of this argument, he also points out the remaining mysteries of consciousness.

"I do not wish to give the impression that I think there is no mystery about consciousness. There is, for instance, something of a paradox connected with any attempt to try to localise it. But I do not think these mysteries necessarily need to be solved before we can answer the question which is concerned in this paper". (Turing, p.446).

With this paragraph, I will argue that Turing did not mean that passing the imitation game was proof of consciousness, instead passing the Turing Test is again another proof that a computer can do any mechanical process and that mimicking even a human is mechanical, it's just repetition. And if we think learning is just pure mechanical, just a mimicking process, that can be modeled with rewards and punishments as Turing (1948) tries to abstract. Then it can be replicated on a computer and in humans. But if our learning is not purely mechanical, we must give proof of what is lacking, something that we mentioned earlier, semantics, and what Turing (1948) adds initiative, as something I think may be similar to intentionality.

To convert a brain or machine into a universal machine is the extremest form of discipline. Without something of this kind, one cannot set up proper communication. But discipline is certainly not enough in itself to produce intelligence. That which is required in addition we call initiative. (Turing, 1948, p.125)

4. Conclusions

While AI programs continue to develop and imitate more and more human activities, these programs without intentionality or initiative will lack innovation and creativity. You may say these adjectives are in the same way relative to an observer as intelligence or consciousness. But, of course, we can say the same for learning. If a student only replicates the teaching of the professor, is the student learning?

In the same way that we doubt machine learning, we may doubt our current learning process. Because if we plan education to be just a mechanical process, with strict discipline and a rewards and punishment system, we are doing the same as machine learning. We may not have schools but factories of paper (human) machines!

What it also lacks in machine learning is what Turing (1948) states about the cultural state, the way in which we interact with our social, cultural, and natural environment. We do not learn in isolation, nor are we born in a Chinese Room. We learn so much not just by our senses, but also by the ideas of others.

The remaining form of search is what I should like to call the 'cultural search'. As I have mentioned, the isolated man does not develop any intellectual power. It is necessary for him to be immersed in an environment of life. He may then perhaps do a little research of his own and make a very few discoveries which are passed on to the other men. From this point of view the search for new techniques must be regarded as carried out by the human community as a whole, rather than by the individuals. (Turing, 1948, p. 127)

Asking whether machines with AI programs can learn, depends a lot on the meaning that we put in learning. If we attach intelligence to learning, we may ask ourselves more questions;

If what we do with our life is just a mechanical process that can be replaced with a Turing Machine, then does it make sense to continue doing it? If the education just follows the same paradigms as our professors, where does the innovation take place? If scientific discoveries are just data analysis, are we making progress?

In conclusion what AI lacks is meaning, intentionality, and connection with our social-cultural environment. We must change the path from AI (Absolute Ignorance) to IA (Intelligence Amplifier), otherwise what we are making is just repetition. Are we going to continue doing just the mechanical process? Or to follow our intuitions and do things with intentionality? Can machines Think? Or Can we think? Now we do have two paths to follow.

References

Floridi, L. (2004). Open Problems in the Philosophy of Information. Metaphilosophy, 35(4), 554–582.

Floridi, L. (2010). Information: A Very Short Introduction. Oxford University Press.

Grier, D. A. (2001). Human computers: the first pioneers of the information age. Endeavour, 25(1), 28–32. doi:10.1016/s0160-9327(00)01338-7

Turing, A. M. (1936). On computable numbers, with an application to the Entscheidungsproblem. Proceedings of the London Mathematical Society, s2-42(1), 230–265.

Turing, A. M. (1948). Intelligent Machinery, in B. J. Copeland (ed.), The Essential Turing (Oxford, 2004; online edn, Oxford Academic, 12 Nov. 2020), https://doi.org/10.1093/oso/9780198250791.003.0016, accessed 28 Apr. 2023.

Turing, A. M. (1950). Computing Machinery and Intelligence, Mind, Volume LIX, Issue 236, p. 433–460, https://doi.org/10.1093/mind/LIX.236.433

Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), p. 417–424.