Machine Learning: A New Epistemology
As far back as the 1930s, more than a decade before a functioning computer would first be engineered, Alan Turing was not only designing the principles that underpin calculating machines, he was also philosophizing over the possibility of a machine that could “think” (Grève and Warburton). Turing believed that, given an appropriate set of instructions and unlimited computing power (in this case, an infinite length of tape for reading, recording, and deleting instructions and results), a machine could compute anything that a human possibly could (Grève and Warburton). Not only that, but Turing’s famous stored-program design (the aforementioned principle that underpins computing itself) seems to imply a necessary truth: there is nothing that a machine can be programmed to do that it cannot also, in principle, come to learn on its own (Grève and Warburton). Alan Turing was thus contemplating machine learning before there were even machines. By 1947 he was giving the first ever lectures on the subject of artificial intelligence to the London Mathematical Society, when the Colossus (the world’s first ever electronic computer) was not even three years old (Grève and Warburton).
This topic naturally leads to the question as to how far it is possible in principle for a computing machine to simulate human activities. What we want is a machine that can learn from experience… The possibility of letting the machine alter its own instructions provides the mechanism for this. (Turing, qtd. in Grève and Warburton).
The mechanism Turing is referring to here is his stored-program design. Turing could foresee machines that would ultimately beat the worlds best chess and Go players: “there are indications … that it is possible to make the machine display intelligence at the risk of its making occasional serious mistakes. …the machine could probably be made to play very good chess” (Turing, qtd. in Grève and Warburton). Moreover, he also predicted the rise of algorithms so complex that humans (even the programmers themselves) would not be able to understand how the calculating machines came to their results.
Let us suppose we have set up a machine with certain initial instruction tables, so constructed that these tables might…modify [their own contents]. One can imagine that after the machine had been operating for some time, the instructions would have altered out of all recognition, but nevertheless still be such that one would have to admit that the machine was still doing very worthwhile calculations. … In such a case one would have to admit that the progress of the machine had not been foreseen when its original instructions were put in. It would be like a pupil who had learnt much from his master, but had added much more by his own work. When this happens I feel that one is obliged to regard the machine as showing intelligence. (Turing, qtd. in Grève and Warburton).
This sentiment of creative intelligence is echoed by Go world champion Lee Sedol when he remarked on a curious move taken by his machine opponent AlphaGo: “I thought AlphaGo was based on probability calculation and that it was merely a machine. But when I saw this move, I changed my mind. Surely, AlphaGo is creative” (Sedol, qtd. in Kohs).
Though, even in Turing’s relentless optimism for the potential of thinking machines, there was still concern for the impact this may have on society and what it means to be human:
If a machine can think, it might think more intelligently than we do, and then where should we be? Even if we could keep the machines in a subservient position, for instance by turning off the power at strategic moments, we should, as a species, feel greatly humbled. … This new danger … if it comes at all … is remote but not astronomically remote, and is certainly something which can give us anxiety. (Turing, qtd. in Grève and Warburton).
Almost 9 decades and countless evolutions in computing later, we are now dealing with the anxiety that Turing predicted with respect to intelligent machines, or what we refer to now as machine learning algorithms. In the field of science, a black box is any system for which inputs are provided and outputs observed, but for which the inner workings are unknown by the observer; if not unknowable to anyone at all. In computing, a designer may understand a program, whereas its tester is intentionally blind to how it works. This is known as black box testing and is generally the case for simple and straightforward software programs for which the word “intelligent” would never be used. What makes machine learning algorithms different is that they are considered black boxes even to their designers. Essentially, no one can explain how the individual answers or results are arrived at, only how the model was designed to learn in general terms. This is not hyperbole, because it is in the actual nature of machine learning models that this needs to be true.
Simply looking at code can tell you how a simple software program works but this is not the case for machine learning. Models are simply provided with massive amounts of data, for example, images of handwritten numbers, that are labeled correspondingly with what each image represents. The model then teaches itself about the correlations it finds between images that have similar labels. Then, when provided an unlabeled handwritten image of an “8”, it will use statistical analysis and probabilistic weighting to correctly identify the image based on its historical understanding of all the ways that an “8” might look versus all the ways it knows any other number might look (Weinberger and Warburton). It may think there’s a 20% chance that the image is of a “3” but so too might a human that reviews and judges the image in question. And that is the point. What’s been described here is not much different from how a human learns to discern and identify symbols. So if you recognize that your brain is a black box for which you can’t explain how you know when you see an image of an “8” other than to refer to your history of seeing “8”s, then you are beginning to comprehend the anxiety many have over machine learning and our reliance on it for making decisions on our behalf.
Identifying numbers using machine learning is a useful example for being introduced to the concept of machine learning, but these are not the anxiety-inducing decisions previously referred to. Machine learning is also used to drive vehicles, to determine what videos to recommend to you, to make correlations in the health industry and to aid doctors in understanding how a specific body will react under certain conditions, to determine criminal recidivism, to mitigate the spread of disease, to find and flag hate speech and other unwanted material online, and to better understand climate change, just to name a handful of important areas in which machine learning is driving progress (Weinberger and Warburton). All of which come with the black box inexplicability factor. This is where the Turing anxiety ought to set in.
Much in the same way Deepmind’s AlphaGo is not provided the rules for playing Go but is only given thousands upon thousands of played games of Go to analyze (Knight), a healthcare-specific machine learning algorithm is not given a crash course in medicine but is simply given enormous sets of health data on individuals. It then teaches itself to identify connections between individuals with certain health conditions, and in this way it can predict which other individuals are likely to develop the same condition. This is an entirely new epistemological system, one that begins to approach a Laplace’s Demon-esque understanding of the world (Weinberger and Warburton). So is the inexplicability of machine learning black box algorithms simply the price we pay for access to this new epistemology? Or is it a revelation of reality in the sense that machines are simply better at analyzing finer grain details of statistically interrelated data and finding connections that would otherwise be inaccessible to the human mind (Weinberger and Warburton)?
What heightens this concern are the ways in which human bias can sneak into the system. “Garbage in; garbage out” is a programming principle that points out that a program is only as good as the quality of data provided. Which means that if the way in which data is labeled or if the sets of data it receives contain bias, then the algorithm will learn that bias as part of the process, replicating decisions that a biased human would make. For example, there are a number of artificial intelligence projects in development that focus on the use of facial recognition for dubious purposes such as finding suitable job candidates, identifying and tracking the movements of specific ethnic minorities, or identifying potential criminals (Stinson and Davies). The problem with these approaches is that if you provide data of what 10,000 doctors look like and only 30% of them are women, the algorithm is going to learn that men are more suitable to be doctors. Similarly, when an AI is fed the photos of the current criminal population, something we know to be skewed by the systemic over-policing of certain communities, the class inequities that prime for criminal behaviour, and the subsequent lack of equal access to legal representation, you will have created an algorithm that simply regurgitates these structural social problems by targeting these same people based on their appearance (Stinson and Davies). “If it’s programmed to be biased, it will be biased with great efficiency” (Boddington and Warburton). Any machine learning algorithm that is based on data from an analog system that is currently criticized as being racist, classist, or sexist, will necessarily create algorithmically racist, classist, or sexist output. “Garbage in, garbage out” indeed. It is for this reason that many argue that facial recognition algorithms ought to be subject to tight regulation and high public scrutiny (Stinson and Davies).
When viewed this way, these algorithms hardly feel “intelligent”, since they are simply mirroring a pre-existing reality, albeit much more efficiently. But that is simply a reflection of those particular use cases. As seen in the case of AlphaGo, artificial intelligence can also be seen as being “creative”. In 2016, a Princeton student released a soundcloud of music generated by his deep learning algorithm called “deepjazz” (Roeder and Dresser) and the OpenAI group is currently in the testing phase for a program called DALL-E 2 (a play on the film WALL-E and artist Salvadore Dali) which can generate incredible artistic images in multiple styles based only on a user-supplied single-sentence description (Johnson, DALL-E 2). If a Turing test is meant to pass based on whether a computer can make a person believe that they are conversing with a human as opposed to a machine, then surely deepjazz and DALL-E 2 are passing the art version of the Turing test with flying colours (almost literally in DALL-E 2’s case).
The advent of “computer-generated art” makes some uncomfortable, of course, since when art is meant to be a commentary on the human condition of existence, it can be surreal when the “artist” lacks this all-important condition (Roeder and Dresser), regardless of the quality of the art being produced. However, it is all a matter of perspective. Lee Sedol wrestled with AlphaGo over five games partially to “defend” the idea and supremacy of humanity, but the DeepMind team argues that AlphaGo’s success is due to the same human ingenuity that makes Lee Sedol a world champion Go-player. Which is to say that beautiful music or complex illustrations generated by algorithms are still human art — just with extra steps (Roeder and Dresser). Whether it ought to be considered art is perhaps the less important question, since, if these products are indistinguishable from traditionally designed art, the market will inevitably accept them without thinking much about it. The bigger question may be how this impacts the role of the artist and what it means to be one going forward (Roeder and Dresser). Art is not at risk here, but the economics of art may be about to shift wildly (Johnson, DALL-E 2).
Unfortunately, the world of AI-creativity too has a dark side (and you may be noticing a theme here), which is the problem of perpetuating societal stereotypes in the images it produces. It seems that it is impossible to have clean input data when the starting point is a world that itself is unclean. DALL-E 2 is not currently available to the public, and that’s partially because of its tendency, when asked to generate photorealistic images of people, to default towards white men when given positive or neutral image descriptions (like “CEO”) and a preference for depicting minorities when given image descriptions with negative adjectives like “angry” or negative nouns like “terrorist” (Johnson, DALL-E 2). Moreover, its images of women are often overly sexualized and are more likely to be generated in response to prompts like “assistant” or “flight attendant”, etc. (Johnson, DALL-E 2). There are more subtle problems as well such as having a clear western bias when presented with prompts regarding food or cityscapes, or in the different ways it will respond when asked to depict a European classroom versus an African classroom (Johnson, DALL-E 2). For this reason, there is a strong push, both externally and internally, to release DALL-E 2 without the ability to produce images of human faces at all (Johnson, DALL-E 2). The good news is that the only reason we know this is because the OpenAI team is aware of and publicly communicative about this problem as opposed to trying to hide or disregard it. That acknowledgement alone is a step in the right direction since the industry of artificial intelligence has a history of releasing products before understanding the potential harms they could cause or perpetuate (Johnson, DALL-E 2). Researchers at MIT claim that the measurement and mitigation of bias in datasets used in the development of artificial intelligence is “critical to building a fair society” (Johnson, DALL-E 2).
The problems of “garbage” input data and the black box mystery inherent in the outputs of deep learning models make for a potentially dangerous combination when we begin to put our blind trust in the guidance of these systems that can calculate answers to problems that a human mind never could. It is already a known problem that people will implicitly accept decisions made by artificial intelligences (Johnson, LaMDA), but the key to solving this may lie in empowering them with two-way communication. When lead researcher of DeepMind David Silver was asked about the importance of being able to query these black box systems about the decisions they make and the conclusions they come to, he said “That’s an interesting question. For some applications it may be important. Like in health care, it may be important to know why a decision is being made” (qtd. in Knight). Artificial intelligence systems ought to be designed to be collaborative, which is to say that they will need the power of understanding language (Knight). That is the only way we will know if an AI understands what is at stake when it makes a choice.
What makes AlphaGo’s defeat of Lee Sedol so much more impressive than Deep Blue’s chess defeat of Larry Kasparov in the ’90s is in the nature of the two games themselves. Chess is a game of thinking steps ahead and about how your opponent might respond, something that a deep learning model can be programmed to assign probabilistic weights to. Go on the other hand does not have “strategy” in the same computable sense but is much more intuition-based. A chess player can explain the reasoning for their move to the extent that moves and certain sets of moves have famous names because they are known to be effective. Go does not have known “gambits” and even its greatest players admit that there is difficulty in explaining why they chose one particular move (Knight). This is the crux of why people were so skeptical that an artificial intelligence could be any good at Go, let alone defeat world champions. AlphaGo’s success reveals that deep learning can calculate its way to intuition and creativity (Knight). And if artificial intelligence can master the intuition and creativity required to excel at the game of Go, that may be a starting point for it to use language not just as output but to actually understand what it is communicating (Knight). There are many linguists and philosophers of mind who argue that language is the key to bootstrapping self-awareness, and while the potential for truly conscious machines is outside of the scope of this paper, the point is that a machine that can actually understand the nature of its work and how to genuinely communicate with those it is providing answers to allows for a way for us to query the black box. A deep learning model that can be asked about how and why it came to the answer that it did would be extremely useful for justifying trust in its responses and/or may supply us with reasons to reject it in other cases. This would also give us the opportunity to update the model with the reasons we have for rejecting its response simply by conversing with it.
It is disconcerting to peer into a black box and have to wonder how trustworthy its decisions are and how infected with bias they may be. But the plain fact is that these algorithms do work, in the sense that they produce something of great value. In all cases they are producing faster and more efficiently than would otherwise be possible, and in many cases are producing something that could not be produced otherwise at all. “The success of our technology is teaching us that the world is the real black box” (Weinberger and Warburton). This new epistemology is possible explicitly because it is an intelligence that is free from the generalizations that human thought relies on (Weinberger and Warburton), giving us access to the complexities that make up our world. What remains to be seen is whether we will truly understand the answers we are given or if we will have to accept them as if given by a divine, unknowable oracle. “This technology gives us increased mastery, but not understanding” (Weinberger and Warburton).
References
Boddington, Paula, and Nigel Warburton. “What are the Values that Drive Decision Making by AI?” Aeon, 21 March 2019, https://aeon.co/essays/what-are-the-values-that-drive-decision-making-by-ai. Accessed 12 June 2022.
Grève, Sebastian Sunday, and Nigel Warburton. “Why We Should Remember Alan Turing as a Philosopher.” Aeon, 21 April 2022, https://aeon.co/essays/why-we-should-remember-alan-turing-as-a-philosopher. Accessed 29 May 2022.
Johnson, Khari. “DALL-E 2 Creates Incredible Images — and Biased Ones You Don’t See.” WIRED, 5 May 2022, https://www.wired.com/story/dall-e-2-ai-text-image-bias-social-media/. Accessed 11 June 2022.
Johnson, Khari. “LaMDA and the Sentient AI Trap.” WIRED, 14 June 2022, https://www.wired.com/story/lamda-sentient-ai-bias-google-blake-lemoine/. Accessed 14 June 2022.
Knight, Will. “AI’s Language Problem.” MIT Technology Review, 9 August 2016, https://www.technologyreview.com/2016/08/09/158125/ais-language-problem/. Accessed 12 June 2022.
Kohs, Greg, director. AlphaGo — The Movie. Moxie Pictures, Reel As Dirt, 2017, https://www.youtube.com/watch?v=WXuK6gekU1Y.
Roeder, Oliver, and Sam Dresser. “There is No Such Thing as Computer Art: It’s All Just Art.” Aeon, 20 July 2016, https://aeon.co/ideas/there-is-no-such-thing-as-computer-art-it-s-all-just-art. Accessed 11 June 2022.
Stinson, Catherine, and Sally Davies. “Algorithms Associating Appearance and Criminality Have a Dark Past.” Aeon, 15 May 2020, https://aeon.co/ideas/algorithms-associating-appearance-and-criminality-have-a-dark-past. Accessed 11 June 2022.
Weinberger, David, and Nigel Warburton. “Our World is a Black Box, Predictable But Not Understandable.” Aeon, 15 November 2021, https://aeon.co/essays/our-world-is-a-black-box-predictable-but-not-understandable. Accessed 4 June 2022.
