How can we make AI that reasons?
The past decade or so has been touted as a high point for achievements in Artificial Intelligence (AI). For the first time, computers have demonstrated formidable ability in such areas as image recognition, speech recognition, gaming, and (most recently) autonomous driving / piloting. Researchers and companies that are heavily invested in these technologies, at least, are in no small way lauding these successes, and are giving us the pitch that the current state-of-the-art is nothing less than groundbreaking.
However, as anyone exposed to the industry knows, the current state-of-the-art is still plagued by fundamental shortcomings. In a nutshell, the current generation of AI is characterised by big data (i.e. a huge amount of sample data is needed in order to yield only moderately useful results), big hardware (i.e. a giant amount of clustered compute resources is needed, again in order to yield only moderately useful results), and flawed algorithms (i.e. algorithms that, at the end of the day, are based on statistical analysis and not much else – this includes the latest Convolutional Neural Networks). As such, the areas of success (impressive though they may be) are still dwarfed by the relative failures, in areas such as natural language conversation, criminal justice assessment, and art analysis / art production.
In my opinion, if we are to have any chance of reaching a higher plane of AI – one that demonstrates more human-like intelligence – then we must lessen our focus on statistics, mathematics, and neurobiology. Instead, we must turn our attention to philosophy, an area that has traditionally been neglected by AI research. Only philosophy (specifically, metaphysics and epistemology) contains the teachings that we so desperately need, regarding what "reasoning" means, what is the abstract machinery that makes reasoning possible, and what are the absolute limits of reasoning and knowledge.
What is reason?
There are many competing theories of reason, but the one that I will be primarily relying on, for the rest of this article, is that which was expounded by 18th century philosopher Immanuel Kant, in his Critique of Pure Reason and other texts. Not everyone agrees with Kant, however his is generally considered the go-to doctrine, if for no other reason (no pun intended), simply because nobody else's theories even come close to exploring the matter in such depth and with such thoroughness.
One of the key tenets of Kant's work, is that there are two distinct types of propositions: an analytic proposition, which can be universally evaluated purely by considering the meaning of the words in the statement; and a synthetic proposition, which cannot be universally evaluated, because its truth-value depends on the state of the domain in question. Further, Kant distinguishes between an a priori proposition, which can be evaluated without any sensory experience; and an a posteriori proposition, which requires sensory experience in order to be evaluated.
So, analytic a priori statements are basically tautologies: e.g. "All triangles have three sides" – assuming the definition of a triangle (a 2D shape with three sides), and assuming the definition of a three-sided 2D shape (a triangle), this must always be true, and no knowledge of anything in the universe (except for those exact rote definitions) is required.
Conversely, synthetic a posteriori statements are basically unprovable real-world observations: e.g. "Neil Armstrong landed on the Moon in 1969" – maybe that "small step for man" TV footage is real, or maybe the conspiracy theorists are right and it was all a hoax; and anyway, even if your name was Buzz Aldrin, and you had seen Neil standing there right next to you on the Moon, how could you ever fully trust your own fallible eyes and your own fallible memory? It's impossible for there to be any logical proof for such a statement, it's only possible to evaluate it based on sensory experience.
Analytic a posteriori statements, according to Kant, are impossible to form.
Which leaves what Kant is most famous for, his discussion of synthetic a priori statements. An example of such a statement is: "A straight line between two points is the shortest". This is not a tautology – the terms "straight line between two points" and "shortest" do not define each other. Yet the statement can be universally evaluated as true, purely by logical consideration, and without any sensory experience. How is this so?
Kant asserts that there are certain concepts that are "hard-wired" into the human mind. In particular, the concepts of space, time, and causality. These concepts (or "forms of sensibility", to use Kant's terminology) form our "lens" of the universe. Hence, we are able to evaluate statements that have a universal truth, i.e. statements that don't depend on any sensory input, but that do nevertheless depend on these "intrinsic" concepts. In the case of the above example, it depends on the concept of space (two distinct points can exist in a three-dimensional space, and the shortest distance between them must be a straight line).
Another example is: "Every event has a cause". This is also universally true; at least, it is according to the intrinsic concepts of time (one event happens earlier in time, and another event happens later in time), and causality (events at one point in space and time, affect events at a different point in space and time). Maybe it would be possible for other reasoning entities (i.e. not humans) to evaluate these statements differently, assuming that such entities were imbued with different "intrinsic" concepts. But it is impossible for a reasoning human to evaluate those statements any other way.
The actual machinery of reasoning, as Kant explains, consists of twelve "categories" of understanding, each of which has a corresponding "judgement". These categories / judgements are essentially logic operations (although, strictly speaking, they predate the invention of modern predicate logic, and are based on Aristotle's syllogism), and they are as follows:
|Categories / Judgements
All trees have leaves
Some dogs are shaggy
This ball is bouncy
Chairs are comfy
No spoons are shiny
Oranges are not blue
Inherence / Subsistence
Happy people smile
Causality / Dependence
If it's February, then it's hot
Potatoes are baked or fried
Sharks enjoy eating humans
Beer might be frothy
6 times 7 equals 42
The cognitive mind is able to evaluate all of the above possible propositions, according to Kant, with the help of the intrinsic concepts (note that these intrinsic concepts are not considered to be "innate knowledge", as defined by the rationalist movement), and also with the help of the twelve categories of understanding.
Reason, therefore, is the ability to evaluate arbitrary propositions, using such cognitive faculties as logic and intuition, and based on understanding and sensibility, which are bridged by way of "forms of sensibility".
AI with intrinsic knowledge
If we consider existing AI with respect to the above definition of reason, it's clear that the capability is already developed maturely in some areas. In particular, existing AI – especially Knowledge Representation (KR) systems – has no problem whatsoever with formally evaluating predicate logic propositions. Existing AI – especially AI based on supervised learning methods – also excels at receiving and (crudely) processing large amounts of sensory input.
So, at one extreme end of the spectrum, there are pure ontological knowledge-base systems such as Cyc, where virtually all of the input into the system consists of hand-crafted factual propositions, and where almost none of the input is noisy real-world raw data. Such systems currently require a massive quantity of carefully curated facts to be on hand, in order to make inferences of fairly modest real-world usefulness.
Then, at the other extreme, there are pure supervised learning systems such as Google's NASNet, where virtually all of the input into the system consists of noisy real-world raw data, and where almost none of the input is human-formulated factual propositions. Such systems currently require a massive quantity of raw data to be on hand, in order to perform classification and regression tasks whose accuracy varies wildly depending on the target data set.
What's clearly missing, is something to bridge these two extremes. And, if transcendental idealism is to be our guide, then that something is "forms of sensibility". The key element of reason that humans have, and that machines currently lack, is a "lens" of the universe, with fundamental concepts of the nature of the universe – particularly of space, time, and causality – embodied in that lens.
What fundamental facts about the universe would a machine require, then, in order to have "forms of sensibility" comparable to that of a human? Well, if we were to take this to the extreme, then a machine would need to be imbued with all the laws of mathematics and physics that exist in our universe. However, let's assume that going to this extreme is neither necessary nor possible, for various reasons, including: we humans are probably only imbued with a subset of those laws (the ones that apply most directly to our everyday existence); it's probably impossible to discover the full set of those laws; and, we will assume that, if a reasoning entity is imbued only with an appropriate subset of those laws, then it's possible to deduce the remainder of the laws (and it's therefore also possible to deduce all other facts relating to observable phenomena in the universe).
I would, therefore, like to humbly suggest, in plain English, what some of these fundamental facts, suitable for comprising the "forms of sensibility" of a reasoning machine, might be:
- There are four dimensions: three space dimensions, and one time dimension
- An object exists if it occupies one or more points in space and time
- An object exists at zero or one points in space, given a particular point in time
- An object exists at zero or more points in time, given a particular point in space
- An event occurs at one point in space and time
- An event is caused by one or more different events at a previous point in time
- Movement is an event that involves an object changing its position in space and time
- An object can observe its relative position in, and its movement through, space and time, using the space concepts of left, right, ahead, behind, up, and down, and using the time concepts of forward and backward
- An object can move in any direction in space, but can only move forward in time
I'm not suggesting that the above list is really a sufficient number of intrinsic concepts for a reasoning machine, nor that all of the above facts are the correct choice nor correctly worded for such a list. But this list is a good start, in my opinion. If an "intelligent" machine were to be appropriately imbued with those facts, then that should be a sufficient foundation for it to evaluate matters of space, time, and causality.
There are numerous other intrinsic aspects of human understanding that it would also, arguably, be essential for a reasoning machine to possess. Foremost of these is the concept of self: does AI need a hard-wired idea of "I"? Other such concepts include matter / substance, inertia, life / death, will, freedom, purpose, and desire. However, it's a matter of debate, rather than a given, whether each of these concepts is fundamental to the foundation of human-like reasoning, or whether each of them is learned and acquired as part of intellectual experience.
A machine as discussed so far is a good start, but it's still not enough to actually yield what would be considered human-like intelligence. Cyc, for example, is an existing real-world system that basically already has all these characteristics – it can evaluate logical propositions of arbitrary complexity, based on a corpus (a much larger one than my humble list above) of intrinsic facts, and based on some sensory input – yet no real intelligence has emerged from it.
One of the most important missing ingredients, is the ability to hypothesise. That is, based on the raw sensory input of real-world phenomena, the ability to observe a pattern, and to formulate a completely new, original proposition expressing that pattern as a rule. On top of that, it includes the ability to test such a proposition against new data, and, when the rule breaks, to modify the proposition such that the rule can accommodate that new data. That, in short, is what is known as deductive reasoning.
A child formulates rules in this way. For example, a child observes that when she drops a drinking glass, the glass shatters the moment that it hits the floor. She drops a glass in this way several times, just for fun (plenty of fun for the parents too, naturally), and observes the same result each time. At some point, she formulates a hypothesis along the lines of "drinking glasses break when dropped on the floor". She wasn't born knowing this, nor did anyone teach it to her; she simply "worked it out" based on sensory experience.
Some time later, she drops a glass onto the floor in a different room of the house, still from shoulder-height, but it does not break. So she modifies the hypothesis to be "drinking glasses break when dropped on the kitchen floor" (but not the living room floor). But then she drops a glass in the bathroom, and in that case it does break. So she modifies the hypothesis again to be "drinking glasses break when dropped on the kitchen or the bathroom floor".
But she's not happy with this latest hypothesis, because it's starting to get complex, and the human mind strives for simple rules. So she stops to think about what makes the kitchen and bathroom floors different from the living room floor, and realises that the former are hard (tiled), whereas the latter is soft (carpet). So she refines the hypothesis to be "drinking glasses break when dropped on a hard floor". And thus, based on trial-and-error, and based on additional sensory experience, the facts that comprise her understanding of the world have evolved.
Some would argue that current state-of-the-art AI is already able to formulate rules, by way of feature learning (e.g. in image recognition). However, a "feature" in a neural network is just a number, either one directly taken from the raw data, or one derived based on some sort of graph function. So when a neural network determines the "features" that correspond to a duck, those features are just numbers that represent the average outline of a duck, the average colour of a duck, and so on. A neural network doesn't formulate any actual facts about a duck (e.g. "ducks are yellow"), which can subsequently be tested and refined (e.g. "bath toy ducks are yellow"). It just knows that if the image it's processing has a yellowish oval object occupying the main area, there's a 63% probability that it's a duck.
Another faculty that the human mind possesses, and that AI currently lacks, is intuition. That is, the ability to reach a conclusion based directly on sensory input, without resorting to logic as such. The exact definition of intuition, and how it differs from instinct, is not clear (in particular, both are sometimes defined as a "gut feeling"). It's also unclear whether or not some form of intuition is an essential ingredient of human-like intelligence.
It's possible that intuition is nothing more than a set of rules, that get applied either before proper logical reasoning has a chance to kick in (i.e. "first resort"), or after proper logical reasoning has been exhausted (i.e. "last resort"). For example, perhaps after a long yet inconclusive analysis of competing facts, regarding whether your Uncle Jim is telling the truth or not when he claims to have been to Mars (e.g. "Nobody has ever been to Mars", "Uncle Jim showed me his medal from NASA", "Mum says Uncle Jim is a flaming crackpot", "Uncle Jim showed me a really red rock"), your intuition settles the matter with the rule: "You should trust your own family". But, on the other hand, it's also possible that intuition is a more elementary mechanism, and that it can't be expressed in the form of logical rules at all: instead, it could simply be a direct mapping of "situations" to responses.
Is reason enough?
In order to test whether a hypothetical machine, as discussed so far, is "good enough" to be considered intelligent, I'd like to turn to one of the domains that current-generation AI is already pursuing: criminal justice assessment. One particular area of this domain, in which the use of AI has grown significantly, is determining whether an incarcerated person should be approved for parole or not. Unsurprisingly, AI's having input into such a decision has so far, in real life, not been considered altogether successful.
The current AI process for this is based almost entirely on statistical analysis. That is, the main input consists of simple numeric parameters, such as: number of incidents reported during imprisonment; level of severity of the crime originally committed; and level of recurrence of criminal activity. The input also includes numerous profiling parameters regarding the inmate, such as: racial / ethnic group; gender; and age. The algorithm, regardless of any bells and whistles it may claim, is invariably simply answering the question: for other cases with similar input parameters, were they deemed eligible for parole? And if so, did their conduct after release demonstrate that they were "reformed"? And based on that, is this person eligible for parole?
Current-generation AI, in other words, is incapable of considering a single such case based on its own merits, nor of making any meaningful decision regarding that case. All it can do, is compare the current case to its training data set of other cases, and determine how similar the current case is to those others.
A human deciding parole eligibility, on the other hand, does consider the case in question based on its own merits. Sure, a human also considers the numeric parameters and the profiling parameters that a machine can so easily evaluate. But a human also considers each individual event in the inmate's history as a stand-alone fact, and each such fact can affect the final decision differently. For example, perhaps the inmate seriously assaulted other inmates twice while imprisoned. But perhaps he also read 150 novels, and finished a university degree by correspondence. These are not just statistics, they're facts that must be considered, and each fact must refine the hypothesis whose final form is either "this person is eligible for parole", or "this person is not eligible for parole".
A human is also influenced by morals and ethics, when considering the character of another human being. So, although the question being asked is officially: "is this person eligible for parole?", the question being considered in the judge's head may very well actually be: "is this person good or bad?". Should a machine have a concept of ethics, and/or of good vs bad, and should it apply such ethics when considering the character of an individual human? Most academics seem to think so.
According to Kant, ethics is based on a foundation of reason. But that doesn't mean that a reasoning machine is automatically an ethical machine, either. Does AI need to understand ethics, in order to possess what we would consider human-like intelligence?
Although decisions such as parole eligibility are supposed to be objective and rational, a human is also influenced by emotions, when considering the character of another human being. Maybe, despite the evidence suggesting that the inmate is not reformed, the judge is stirred by a feeling of compassion and pity, and this feeling results in parole being granted. Or maybe, despite the evidence being overwhelmingly positive, the judge feels fear and loathing towards the inmate, mainly because of his tough physical appearance, and this feeling results in parole being denied.
Should human-like AI possess the ability to be "stirred" by such emotions? And would it actually be desirable for AI to be affected by such emotions, when evaluating the character of an individual human? Some such emotions might be considered positive, while others might be considered negative (particularly from an ethical point of view).
I think the ultimate test in this domain – perhaps the "Turing test for criminal justice assessment" – would be if AI were able to understand, and to properly evaluate, this great parole speech, which is one of my personal favourite movie quotes:
There's not a day goes by I don't feel regret. Not because I'm in here, or because you think I should. I look back on the way I was then: a young, stupid kid who committed that terrible crime. I want to talk to him. I want to try and talk some sense to him, tell him the way things are. But I can't. That kid's long gone and this old man is all that's left. I got to live with that. Rehabilitated? It's just a bulls**t word. So you can go and stamp your form, Sonny, and stop wasting my time. Because to tell you the truth, I don't give a s**t."Red" (Morgan Freeman)
The Shawshank Redemption (1994)
In the movie, Red's parole was granted. Could we ever build an AI that could also grant parole in that case, and for the same reasons? On top of needing the ability to reason with real facts, and to be affected by ethics and by emotion, properly evaluating such a speech requires the ability to understand humour – black humour, no less – along with apathy and cynicism. No small task.
Sorry if you were expecting me to work wonders in this article, and to actually teach the world how to build artificial intelligence that reasons. I don't have the magic answer to that million dollar question. However, I hope I have achieved my aim here, which was to describe what's needed in order for it to even be possible for such AI to come to fruition.
It should be clear, based on what I've discussed here, that most current-generation AI is based on a completely inadequate foundation for even remotely human-like intelligence. Chucking big data at a statistic-crunching algorithm on a fat cluster might be yielding cool and even useful results, but it will never yield intelligent results. As centuries of philosophical debate can teach us – if only we'd stop and listen – human intelligence rests on specific building blocks. These include, at the very least, an intrinsic understanding of time, space, and causality; and the ability to hypothesise based on experience. If we are to ever build a truly intelligent artificial agent, then we're going to have to figure out how to imbue it with these things.
- Immanuel Kant: Aesthetics
- The Mismatch Between Human and Machine Knowledge (1994)
- Gödel, Consciousness and the Weak vs. Strong AI Debate
- Transforming Kantian Aesthetic Principles into Qualitative Hermeneutics for Contemplative AGI Agents (2018)
- Recognizing context is still hard in Machine Learning — here’s how to tackle it
- Towards Deep Symbolic Reinforcement Learning (2016)
- Generality in Artificial Intelligence (1987)
- The Symbol Grounding Problem (1990)
- Aristotle’s Ten Categories
- Computational Beauty: Aesthetic Judgment at the Intersection of Art and Science (2014)
- Philosophy of artificial intelligence
- A proposal for ethically traceable artificial intelligence (2017)
- Commonsense knowledge (artificial intelligence)
- Kant's Critique of Pure Reason
- Schopenhauer on Space, Time, Causality and Matter: A Physical Re-examination (2018)
- A Brief Introduction to Temporality and Causality (2010)
- Sequences of Mechanisms for Causal Reasoning in Artificial Intelligence (2013)