Not too long ago I came across a compilation of ChatGPT’s mistakes at reasoning [5]. The fact that a language model of gargantuan dimensions as ChatGPT struggles at reasoning means that despite the overall AI-hype there is still much work to do for AI-researchers and computational neuroscientists like myself. There are still many open questions in the field. For an example of one of ChatGPT’s errors at reasoning look at the figure below.
ChatGPT’s mistakes at reasoning point to the fact that we are studying intelligence and reasoning as two different entities, when in fact, in natural systems, they are tightly connected.
Although we speak of intelligence and intelligent entities, be they natural or artificial, we loosely know what it is meant by the term. Worse yet, we don’t know how it is related to notions such as reason and understanding, so that some times these terms are used interchangeably.
The term intelligence has been defined ad nauseam elsewhere and in very broad contexts. One such efforts is that of Shane Legg and Marcus Hutter [18], where the authors try to formalise as precise as possible the definition of machine intelligence only to get entangled in complex terminology that, far from clarifying or establishing the limits of the term, confuses the reader.
Why do we need other’s people definition of intelligence when we can easily consult a dictionary and agree on what is meant by the term so that whenever we throw the term in a conversation, the involved parties know its meaning? This is exactly where I stand. However, if we needed to agree on a definition of intelligence at this very point I would say that intelligence is the capacity of an agent to solve a given problem. This definition is not too different from those presented elsewhere, and it hides at its core some subtleties that it’s worth taken into consideration. First, the definition implies the existence of an agent and a problem. For whom something is a problem lies on the realm of philosophy. What it is a problem to you might not be a problem to me. Thus, the notion of intelligence has also an aspect of subjectivity. Birds need to solve to problem of seasonal migration, whereas we do not. We need to solve the problem of finding a job in order to sustain ourselves, whereas birds do not. Second, the definition makes no mention about the quality of the problem-solving strategy, it just refers to the capacity an agent has to arrive at it. So that, two agents might solve the same problem differently with one doing it more efficiently than the other in accordance to some evaluation metric.
So what is reason then, and how is it related to intelligence? Are they the same thing? Reason is a mental faculty by which an agent is able to derive conclusions based on premises, or in other words, to derive new statements based on old ones. This again is an express definition which is not too far from others we might find in other places. This definition might imply that language is required for an agent to reason. However, this is not necessarily the case as there are organisms that exhibit some form of reasoning without the need of language, or at least not language as commonly conceived (see this blog post).
The hello-world of reason, that is, the prototypical example of reason is
the following Aristotelian syllogism:
All men are mortal.
Socrates is a man.
Thus, Socrates is mortal.
Before reaching the last statement in the syllogism above, you might feel an itching in your reason, a desire to spit out the conclusion, a prediction somewhere in your mind regarding where are these premises leading to. This is what reasoning is all about.
The above example required us to have the capacity to process language, in this particular example, the English language. This is what understanding is all about. However, I said above that reason not necessarily requires language, or at least not the sophistication that we use for verbal communication. It has been observed that primates and even crows exhibit some sort of abstract reasoning. For instance, it has been observed that Caledonian crows respond to hidden casual agency in their environment [32], as well as exhibiting abstract reasoning for pattern matching [28].
Is reason innate or is it learned? Pioneering studies by Jean Piaget suggested that reasoning emerges during child development, which is the same to say that we are not fully capable of reasoning out of the womb. For instance, consider the theory of conservation in psychology, in which a subject is able to tell that the amount of a liquid is conserved when transferred from a tall thin glass to a small wide glass of the same volume. Jean Piaget observed that children were able to correctly complete conservation tasks before their fifth year. Moreover, although the syllogism presented above might feel universal in the sense that any human being would be able to derive the conclusion based on the first two premises just a percentage of people would be able to correctly do so. The situation looks even more daunting when trying more complex figures of syllogism than the one presented above [3].
However representative of reason a syllogism might be, syllogistic reasoning is more of a formalism of a mental activity that occurs naturally than a mental capacity per se. Consider the following analogy: if I asked you to throw a kick, a martial art kick that is, you would be able to do it because you possess the muscles and you know how to control your body. However your kick would probably have nothing to do with a kick such as the one called yoko geri in karate. Yoko geri is a formalism of a physical movement (a kick) that most people would be able to do. In the same way, a syllogism is a formalism of a mental activity that people would be able to perform naturally. To become proficient at syllogistic reasoning and at throwing karate kicks we require proper training.
Reason follows principles that are innate, and that are further developed through the life of an individual. Intelligence, that is, the capacity an agent possesses for problem-solving relies on reason. Problem-solving depends on the capacity to derive conclusions based on premises, that is, to generate them to form strategies and hypotheses which an agent would test and verify mentally or physically. What about understanding? Reason requires understanding. In particular in a situation where what we need to reason about is expressed within the domains of human language.
Studies have shown that capacity for syllogistic reasoning correlates with short-term memory capacity and language understanding especially during he formative years of early childhood[3, 1, 33]. Thus, stressing the importance of language understanding in the role of reasoning, in particular, the understanding of quantifiers such as all or some, and operators such as negation. Understanding, that is, the capacity to create mental imagery in response to oral or written language, is an essential aspect of reason. However, such understanding does not need to follow the sophistication of human languages. As mentioned before, other species exhibit some forms of reason without possessing languages with the complexity of human languages. I imagine that in such cases these organisms possess some rudimentary form of communication, and thus, of understanding from which their reasoning capacity takes advantage.
In sum, if I say to you “a dog sleeps under the tree”, and in your mind you can picture a dog sleeping under a tree, then you made use of your faculty for understanding. If I say to you “all men are mortal, Socrates is a man”, and you conclude that Socrates is mortal, then you made use of your faculty for reasoning. Lastly, if I say to you “there is a problem, you’re locked inside a room and the door knob just fell apart” and then ask you how would you get out, you would then make use of intelligence to formulate solutions to the problem. In order to create systems that exhibit intelligent behaviour it is necessary to study the mechanisms behind cognitive faculties such as reason and understanding. The alternative is to artificially (i.e., manually) patch systems such as ChatGPT with corrections to its errors at logical reasoning (see the figure above) everytime someone encounters one. A system like this is brittle and non-adaptive, which is contrary to what we observe in natural intelligent systems.
References
[1] Denisa Ardelean. How do primary school pupils think? syllogistic reasoning in primary school children. Procedia-Social and Behavioral Sciences, 203:57–62, 2015
[3] Bruno G Bara, Monica Bucciarelli, and Philip N Johnson-Laird. Development of syllogistic reasoning. The American journal of psychology,
pages 157–193, 1995.
[5] Ali Borji. A categorical archive of chatgpt failures. arXiv preprint arXiv:2302.03494, 2023.
[18] Shane Legg and Marcus Hutter. Universal intelligence: A definition of machine intelligence. Minds and machines, 17(4):391–444, 2007.
[28] Anna Smirnova, Zoya Zorina, Tanya Obozova, and Edward Wasserman. Crows spontaneously exhibit analogical reasoning. Current Biology, 25(2):256–260, 2015.
[32] Alex H Taylor, Rachael Miller, and Russell D Gray. New caledonian crows reason about hidden causal agents. Proceedings of the National Academy of Sciences, 109(40):16389–16391, 2012.
[33] Michael Henry Tessler, Joshua B Tenenbaum, and Noah D Goodman. Logic, probability, and pragmatics in syllogistic reasoning. Topics in Cognitive Science, 14(3):574–601, 2022.
The Thoughtist 