Back to course documents.

Is machine learning bringing the next revolution in science?

The single greatest transformation in science underway now is the infusion of machine learning into almost all sciences. This infusion is a topic ripe for analysis by philosophers of science. Is something of foundational importance underway? If so, what is it? If not, why is there all the fuss?

The challenge to philosophers of science is provide a sharp formulation of the foundational issue or foundational question. This is not an easy task. Rather it requires considerable creativity by philosophers of science to find these sharp formulations. Only after that has been done do these formulations appear easy or obvious.

Machine learning is infusing into almost all sciences.

An April 2017 study by the Royal Society (UK) called "The AI Revolution in Science" lists technical breakthroughs in methods:

• Convolutional neural networks
• Reinforcement learning
• Transfer learning
• Generative adversarial networks

It gives examples of "AI as an enabler of scientific discovery":

• Using genomic data to predict protein structures
• Understanding the effects of climate change on cities and regions
• Finding patterns in astronomical data

It lists examples of new applications to be expected:

• Satellite imaging to support conservation
• Understanding social history from archive material
• Materials characterisation using high-resolution imaging
• Understanding complex organic chemistry

Recent major discovery:

"'The entire protein universe': AI predicts shape of nearly every known protein,"
Nature, News, July 28, 2022.

There is much celebratory "revolution" talk.

Chris Anderson, "The End of Theory: The Data Deluge Makes the Scientific Method Obsolete,"Science, June 23, 2008.

"The scientific method is built around testable hypotheses. These models, for the most part, are systems visualized in the minds of scientists. The models are then tested, and experiments confirm or falsify theoretical models of how the world works. This is the way science has worked for hundreds of years."

"There is now a better way. Petabytes allow us to say: "Correlation is enough." We can stop looking for models. We can analyze the data without hypotheses about what it might show. We can throw the numbers into the biggest computing clusters the world has ever seen and let statistical algorithms find patterns where science cannot."

Bryan McMahon, "AI is Ushering In a New Scientific Revolution," The Gradient.

"But AI is more than a just powerful tool in the hands of scientists and partner on this search. The technology is also transforming the scientific process, automating and adding to what people can accomplish using it. AI is ushering in a new scientific revolution by making remarkable breakthroughs in a number of fields, unlocking new approaches to science, and accelerating the pace of science and innovation. As partners in discovery, AI and scientists can explore more of science’s endless frontier together than either could alone."

Dan Falk, "How Artificial Intelligence Is Changing Science," Quanta, March 11, 2019

"The latest AI algorithms are probing the evolution of galaxies, calculating quantum wave functions, discovering new chemical compounds and more. Is there anything that scientists do that can’t be automated?"

Is machine learning just glorified statistics?

A Google search on this question produces a profusion of denials, but no early hit on someone saying that machine learning just is glorified statistical analysis:

Search results, July 29, 2022

Why would anyone think that machine learning is just glorified statistics?
Answer: because that is exactly what it looks like!
Illustration through neural nets:

Regression analysis seeks a function from x to y, where the function is specified by parameters ai Linear: y = a0 + a1x 8th order: y = a0 + a1x+ ... + + a8x8	Neural network analysis seeks a function from the variables in the input layer to the variables in the output layer. Variables xi in each layer are related to the variables yi in the next layer by yi = (normalizing factor) (w1x1 + w2x2 + ...), for each i. The concatenation of all these interlayer functions is the overall function sought.
The function is adapted to a data set {<x1, y1>, <x2, y2>, <x3, y3>,...}	The function is adapted to "training" data by "back propagation" methods.
The function chosen is the one that minimizes the difference between the computed values of y and the values of y in the data set. Minimization of sum of squared differences is standard.	They search for the set of weights wi that minimizes the difference between the function output and the training data output. Minimization of sum of squared differences is standard.

Philosophers of science on machine learning

Celebratory:

"Philosophy of Science meets Machine Learning," Conference Program, 2021,

"Machine learning (ML) does not only transform businesses and the social sphere, it also fundamentally transforms science and scientific practice."

Critical:

Suzanne Kawamleh, "Can Machines Learn How Clouds Work? The Epistemic Implications of Machine Learning Methods in Climate Science," Philosophy of Science, 88 (December 2021) pp. 1008–1020.

"...Scientists are now replacing physically based parameterizations with neural networks that do not represent physical processes directly or indirectly. I analyze the epistemic implications of this method and argue that it undermines the reliability of model predictions. I attribute the widespread failure in neural network generalizability to the lack of process representation. The representation of climate processes adds significant and irreducible value to the reliability of climate model predictions."

Emily Sullivan, "Understanding from Machine Learning Models," The British Journal for the Philosophy of Science, 73, number 1, March 2022.

"...In this article, using the case of deep neural networks, I argue that it is not the complexity or black box nature of a model that limits how much understanding the model provides. Instead, it is a lack of scientific and empirical evidence supporting the link that connects a model to the target phenomenon that primarily prohibits understanding."

David S. Watson, "Conceptual challenges for interpretable machine learning," Synthese (2022) 200:65

"... I argue that the vast majority of IML [interpretable machine learning] algorithms are plagued by
(1) ambiguity with respect to their true target;
(2) a disregard for error rates and severe testing; and
(3) an emphasis on product over process.
... failure to acknowledge these problems can result in counterintuitive and potentially misleading explanations. Without greater care for the conceptual foundations of IML, future work in this area is doomed to repeat the same mistakes.

Deflationary:

Julie Jebeile, Vincent Lam,Tim Räz, "Understanding climate change with statistical downscaling and machine learning," Synthese (2021) 199:1877–1897

"... compare how machine learning and standard statistical techniques affect our ability to understand the climate system. For that purpose, we put five evaluative criteria of understanding to work:

intelligibility, representational accuracy, empirical accuracy, coherence with background knowledge, and assessment of the domain of validity.

We argue that the two families of methods are part of the same continuum where these various criteria of understanding come in degrees, and that therefore machine learning methods do not necessarily constitute a radical departure from standard statistical tools, as far as understanding is concerned."

More Examples in Science

Boyuan Chen et al., "Discovering State Variables Hidden in Experimental Data"  arXiv:2112.10755v1 [math.DS]

"... using video recordings of a variety of physical dynamical systems, ranging from elastic double pendulums to fire flames. Without any prior knowledge of the underlying physics, our algorithm discovers the intrinsic dimension of the observed dynamics and identifies candidate sets of state variables. We suggest that this approach could help catalyze the understanding, prediction and control of increasingly complex systems."

Grey S. Nearing et al. "What Role Does Hydrological Science Play in the Age of Machine Learning?" Water Resources Research, 57, e2020WR028091.

"However, with the accelerating development of modern machine learning (ML) and deep learning (DL) in particular, we know that the reason is the third one listed: watershed‐scale theories (and models) could have been derived from currently available observation data, but the hydrology community simply failed to do so."

"... DL [deep learning] learned to predict in unseen basins better than traditional models..."

This particular paper is remarkable for its heavy engagement with our history and philosophy of science literature:

"On April 27, 1900 William Thomson (Lord Kelvin) gave his “Two Clouds” speech (“Nineteenth‐Century Clouds over the Dynamical Theory of Heat and Light”) at the Royal Institution,..."

"In 1987, Keith Beven gave what might be considered hydrology's version of the Two Clouds speech at a symposium of the International Association of Hydrological Sciences (IAHS) (Beven, 1987). He took a perspective inspired by Thomas Kuhn's theory of scientific revolutions (Kuhn, 1962) ... He proposed that two things would be necessary to push the field of surface hydrology into a new period of “normal science”..."

"As an applied science, hydrology is motivated by both epistm and techn..."

"This is the problem of holist underdetermination (Duhem, 1954; Laudan, 1990), whereby auxiliary hypotheses confound the ability to falsify specific hypothesis."

"Cartwright and McMullin (1984) argued that phenomenological laws not theoretical laws—are the only thing that can actually be tested."

Algorithmic Bias

Topic caution: "low hanging fruit"

This aspect of the modern reliance on algorithmic techniques has attracted more attention, I believe, than other problems. See, for example:

Sina Fazelpour and David Danks, "Algorithmic bias: Senses, sources, solutions," Philosophy Compass. 2021;16:e12760.

"Data‐driven algorithms are widely used to make or assist decisions in sensitive domains, including healthcare, social services, education, hiring, and criminal justice. In various cases, such algorithms have preserved or even exacerbated biases against vulnerable communities, sparking a vibrant field of research focused on so‐called algorithmic biases."

Jeffrey Dastin, "Amazon scraps secret AI recruiting tool that showed bias against women," Reuters, 2018 or here.

"Amazon.com Inc's AMZN.O machine-learning specialists uncovered a big problem: their new recruiting engine did not like women..."