Join the philosophy faculty and other interested parties for the next installment of the 22–23 the Strange Philosophy Thing. The Strange Thing meets in the Strange Lounge (room 103) of Main Hall. We gather there most Mondays of the fall and winter terms from 4:30–5:30 for refreshments and informal conversation around a given topic. It’s a lot of fun!
I take my inspiration for this week’s Strange Thing from this article from The Guardian, by Laura Spinney. Reading it is optional. I’ll hit the most important points, and highlight the most important questions the article raises, quoting some useful passages along the way.
A traditional and familiar strategy for carrying out scientific research is to explain a type of phenomenon by formulating a theory—a hypothesis—which predicts some particulars of that type of phenomenon, and run experiments to confirm those predictions are accurate. Newton, for example, hypothesized that gravity operated in the absence of air resistance according to his famous inverse square law. And sure enough, he was able to derive Kepler’s empirically well-confirmed laws of planetary motion from it. It’s easy to miss the fact that a lot of scientific practice—especially recently, with the use of AI—differs drastically from this picture. As Spinney notes,
“Facebook’s machine learning tools predict your preferences better than any psychologist [could]. AlphaFold, a program built by DeepMind, has produced the most accurate predictions yet of protein structures based on the amino acids they contain. Both are completely silent on why they work: why you prefer this or that information; why this sequence generates that structure.”
The reason they are silent on why they work is due to the opaqueness of what is going on inside the neural nets which generate the predictions from ever-growing datasets.
Spinney grants that some science is, at this stage at least, merely assisted by AI. She notes that Tom Griffiths, a psychologist at Princeton, is improving upon Daniel Kahneman and Amos Tversky’s prospect theory of human economic behavior by training a neural net on “a vast dataset of decisions people took in 10,000 risky choice scenarios, then [comparing it to] how accurately it predicted further decisions with respect to prospect theory.” What they found was that people use heuristics when there are too many options for the human brain to compute and compare the probabilities of. But people use different heuristics depending on their different experiences (e.g., a stockbroker and a teenage bitcoin trader). What is ultimately generated via this process doesn’t look much like a theory but, as Spinney describes, ‘a branching tree of “if… then”-type rules, which is difficult to describe mathematically’. She adds,
“What the Princeton psychologists are discovering is still just about explainable, by extension from existing theories. But as they reveal more and more complexity, it will become less so – the logical culmination of that process being the theory-free predictive engines embodied by Facebook or AlphaFold.”
Of course there is a concern to be noted concerning bias in AI, particularly those that are fed small or biased data sets. But can’t we expect the relevant data sets eventually to become so large that bias isn’t an issue, and thus accuracy of their predictions abounds? Will science look more and more like this in the future? What does that mean for the role of human beings in science?
This last question can be addressed with the help of the notion of interpretability. Theories are interpretable. They involve the postulation of entities and relations between them, which can be understood by human beings, and that explain and predict the phenomena under investigation. But AI-based methods of prediction preclude interpretability due to the opaqueness of the AI processes by which they are generated. Human’s don’t observe the process of the weights of the nodes of the neural net changing as new data is fed in. AI-based science appears to prevent us from having an explanation of why scientific predictions are accurate, and thus presumably precludes us from providing one another with explanations of the phenomena under investigation. There are at least two sorts of reasons to be leery of non-interpretability. One is practical, the other theoretical.
On the practical side, Bingni Brunton and Michael Beyeler, neuroscientists at the University of Washington, Seattle, noted in 2019 that “it is imperative that computational models yield insights that are explainable to, and trusted by, clinicians, end-users and industry”. Spinney notes a good example of this:
‘Sumit Chopra, an AI scientist who thinks about the application of machine learning to healthcare at New York University, gives the example of an MRI image. It takes a lot of raw data – and hence scanning time – to produce such an image, which isn’t necessarily the best use of that data if your goal is to accurately detect, say, cancer. You could train an AI to identify what smaller portion of the raw data is sufficient to produce an accurate diagnosis, as validated by other methods, and indeed Chopra’s group has done so. But radiologists and patients remain wedded to the image. “We humans are more comfortable with a 2D image that our eyes can interpret,” he says.’
On the theoretical side, is such understanding important for its own sake, in addition to or solely because of practical considerations like the one mentioned above? Spinney suggests that AI-based science circumvents the need for human creativity and intuition.
“One reason we consider Newton brilliant is that in order to come up with his second law he had to ignore some data. He had to imagine, for example, that things were falling in a vacuum, free of the interfering effects of air resistance.”
She adds,
‘In Nature last month, mathematician Christian Stump, of Ruhr University Bochum in Germany, called this intuitive step “the core of the creative process”. But the reason he was writing about it was to say that for the first time, an AI had pulled it off. DeepMind had built a machine-learning program that had prompted mathematicians towards new insights – new generalisations – in the mathematics of knots.’
Is this reason in itself to be leery of AI-based science? Because it will reduce the opportunity for creative expression and the exercise of human intuition? Would it do this?
Hope to see you Monday for discussion with the philosophy department and refreshments!