A brief digression into Kant’s Transcendental Idealism
(Totally relevant to AI, I swear)
This blog is generally about AI. However, it is worth taking a moment to discuss Kant’s notion of the transcendental, because much of the confusion around AI epistemology — how AI learn things — can be traced to a misunderstanding of that word. At the end of the post we’ll connect it back to AI. So here goes.
Kant is widely regarded as one of the most important, if not the most important Western philosopher. His best-known work is the Critique of Pure Reason, in which he discussed the fundamentals of transcendental idealism. Since then, every major philosopher of mind, such as Hegel or Nietzsche, has been working in the shadow of his ideas.
Kant developed his own theories largely in response to the great empiricist philosophers who came before him, such as Locke and Hume. The empiricists had argued that everything you and I learn comes either from sense experience, or from reasoning about those experiences. Kant, and other philosophers like Berkeley (more on Berkeley later) realized that something was missing, and that there are many things you can learn or know that can’t come from experience.
To begin with, if you hope to learn anything at all, you must be a “learning creature”. Chairs don’t learn, regardless of what is happening around them. There is something special about the mind that makes it a great learning machine. This is important, because the kind of learning machine you are will influence what you learn and the shape of that knowledge. All your knowledge must be built out of the functions of your mind, just as an English book must be written in English letters.
The simplest example can be seen in a person who is colourblind. No matter how many red objects you show them, they will never be able to learn about “red”. “Red” must already be part of the structure of their mind before you show them a red object, otherwise it will never make its way into their understanding through experience.
Colours, however, are only the tip of the iceberg. Kant’s idealism went much deeper than that.
Consider the notion of “object-ness”. When you look around you, you see “things”, i.e. objects that persist in reality. For example, when you look at a pen, then look away, then look back, you perceive the same pen. Or do you? All your eyes have seen is a series of colour patterns that are constantly changing, and which never repeat themselves perfectly. That parts of this visual flow represent a unified “object” is something your mind must invent for you out of these experiences. It is not given by the experiences themselves.
Object-ness is an interpretation. This is why you are often mistaken about what is and isn’t an object, as often happens during a magic trick. Even a simple movie or 3D video game can convince you that you are seeing “objects” when there are really just points of light on a screen.
Object-ness is also subjective and personal. For example, most of us think World War 1 is a “thing”. But it’s not an object in any physical sense. Nor is love or the internet. Yet you can conceive of them as “things” in the same way you might a pen. The fundamental act is the same.
It should be clear by now that the objects you stitch together in your mind and the ones in the world are never perfectly correlated. For example, you may look at a tree as a singular “thing”, even though it is losing and gaining atoms and cells all the time. There are no real, objectively definable boundaries of a given tree. Your conception of the tree is your own invention, and you can alter it as you like, if you decide to include or exclude various parts, such as the dead leaves or the dried sap. This is encapsulated in the the famous riddle of the Ship of Theseus, which asks if a ship, after removing and adding planks, is still the same ship. From our new perspective the answer is simply: if you want it to be.
A side-effect of this realization is that it suddenly becomes impossible to prove that objects, like a pen, even exist in the world itself. If separating experiences into objects is an arbitrary mental activity, reality is unlikely to have a definitive set of objects in itself. The universe rather resembles a soup of experience-causing material, which your mind then separates into distinct, unified objects. And even this is inappropriately projecting our own conceptions of substance onto reality.
Now you might object and say: “it’s foolish to argue that objects don’t exist in the real world. A person who went around believing that there are no apples, or bread-loafs would end up starving to death”. And that’s exactly the point. Your mind forms such objects because it is useful for you to do so — because you don’t want to starve, etc. But that is a human bias. The universe doesn’t owe you nourishment. It is self-centred to say that the universe itself, regardless of if humans are present, has defined actual food items in its fundamental ontology just for our benefit — as though truth really cared about our feelings.
There is a certain humility that comes from recognizing that everything you think about is something you invented for yourself, and that reality doesn’t necessarily share your concepts and beliefs. Or at least I hope you’re feeling humble, because the rabbit-hole of Kantian idealism is about to drop the floor out from under you. Startlingly, Kant also argues that even space and time are mental constructs, and exist only in the mind.
“OK”, you might be thinking. “I was following you until now, but this is ridiculous. Space and time are the most clearly, scientifically, incontrovertibly objective entities imaginable”. Well, let’s dig into that claim.
The best way to explain Kant’s argument about space is to look at an example from modern robotics. Imagine a house-vacuuming robot that builds a two-dimensional floor-plan of the room it is in. Such a robot cannot, and will never be able to perceive the world in three dimensions, despite the fact that it ultimately does its labour in three dimensions. So when it topples down the stairs, it will have no way to conceive of what happened to it. You could say the same about us humans. Perhaps the real world has many more dimensions (many string theorists claim it does), but our mental capabilities are constrained to only function within three of those. Perhaps when unexpected things happen to us, it is simply our version of toppling down the trans-dimensional stairs. Either way, the three-dimensional aspect of space is a function of your brain.
In recent decades, computers have given us many novel ways of thinking about space. There is a video game called Hyperbolica, in which you can stroll around in a non-Euclidean space. During development, the game developers conceived of the world in regular (Euclidean) space — they had no choice, that is how their minds worked. They then translated it using mathematical equations into a non-Euclidean space. If you can imagine that, you can also imagine reversing that process; it’s entirely possible that real space is non-Euclidean, but your mind creates a simulation of it that is Euclidean¹.
The same can be said of time. You don’t actually experience time directly, you reconstruct temporal sequences based on your experiences and memories. Henri Bergson discussed this process in depth in his book Time and Free Will (section 2). I also discussed this at length in a previous post, so I won’t repeat myself here.
The bottom line is: how you experience space and time are functions of your mind. Real space and time, as they are in themselves, may be radically different. I emphasize those words because we are getting into what Kant called “the thing in itself”. The thing in itself is reality as it actually is, separate from human experience. Kant’s conclusion was that the thing in itself can’t ever be known; external reality is permanently closed off to us. This conclusion may be unsatisfying, since most of us want to at least try to understand the world as it really is. Perhaps we can find some clever way around this roadblock, some tool or logical device that gets beyond human subjective limitations. We’ve gotten around the problem of subjectivity before, by inventing the scientific process. So why not do it again? Kant’s point was that when you try to know reality as it is in itself, what you are asking for is a way of understanding the world outside human understanding.
That, obviously, is a self-contradiction. The problem isn’t your tools or your process, the problem is you. Your faculty of understanding itself is limiting. It needs to be, otherwise it wouldn’t be “understanding” as we know it, it would be something else. What you learn is shaped by mental functions like space, time, unity, etc. These are the “form” of knowledge, so to speak. They are the train-tracks of your mind; and also the train itself moving along them. The things you learn through your senses are only the stations along the way of the train. Or perhaps the functions of the mind are more like the walls and boundaries of what you can learn. All these metaphors feel insufficient. It’s hard to put into words what the functions of understanding themselves are, as seen from a subjective point of view. Except, we do have a word: transcendental.
The “transcendental” is anything that is outside sense perception; it wraps around and moulds your perception. Thus it is impossible to learn about the transcendental through the senses. For example, I can show you an apple, and you can learn about apples. I can even show you something abstract like a footrace, and you can learn about footraces. But how can I show you time? I could put a clock in front of you, but that’s just a clock. I could make you wait in a room, but then you will only sense your own irritation. Unless you came to the experience already knowing what you are looking for (i.e. time), and already having the ability to “detect” or conceptualize time, nothing I can show you will get you to see time. It would be like showing a red object to a colour-blind person.
The above definition of “transcendental”, namely as anything you can know that does not come through your senses, can easily lead people to a misunderstanding. You can see how a person would jump from that to a notion of God, or to spirituality in general. It is worth noting that Kant himself, though a Christian, flatly denied this to be the case. He spend a large portion of the Critique of Pure Reason arguing that God, or a first mover, or free will, or the infinite, cannot be proven through the system in question². All his transcendental idealism could do was put limitations on what people can know.
His stance was different from that of a previous philosopher, Berkeley; or to use his full title, Bishop Berkeley. Berkeley developed many similar ideas before Kant did. Early reviews of the Critique even argued that the book was simply “Berkeleyism” — which is an unfair simplification. However, Berkeley went the extra step and argued that since reality can’t be known, and that all reality is created in the mind, therefore all minds were joined in the mind of God. This was a step too far for Kant.
These days, the word “transcendental” still gets used to mean something spiritual, and you can see why. Philosophically, however, it is more precise than that. It is a recognition that what humans know — and what AI will ultimately know — is shaped by their faculty of understanding.
As of yet, we don’t have any way of seeing into the subjective state of an AI, and transcendental idealism requires that the whole system be viewed from the inside. So we can’t easily do a transcendental analysis of an AI’s mind. However, this challenge is soon coming. A previous post on this blog was intended to do just that — to figure out how an AI could understand itself and its own transcendental reality. If we’re ever going to take the notion of AGI seriously, then the field cannot be exempt from such analysis.
¹ It is impressive that Kant came up with his own theories long before computer simulations existed, using nothing but thought-experiments.
² These arguments are the so-called “antinomies of reason”.
