> One of my problems was to provide statistical advice to the people who were developing metals to be used in the blades of turbines. I had an enormous amount of data, and I had to construct a regression with five or six different variables having to do with the chemical composition of the metals.
> We estimated that it would take us three months to solve this problem using our desk calculators. In the whole country there was only one calculator—one computer, if you want to call it that—which could do this problem more quickly.
> It was up at Harvard. It wasn’t electronic. It was a whole collection of IBM card sorters. It was in a big, air-conditioned gymnasium, a tremendous collection of sorters all linked by wires. It did our problem for us in forty hours.
As he mentioned elsewhere, it did not work as expected back then.
The prediction algorithm is cool. But call me a tad skeptical. Repeated application of elastic strains results in creep occurring after 1000s of cycles does it not?
AI research into the Materials Genome gets attention. But the problem of discovery still seems secondary to the primary issue. Getting innovations out of laboratories and into manufacturing.
The sorts of fatigue damage you get from cyclic loading aren’t necessarily a problem for the materials applications they’re targeting. Take the diamond transistor the my talk about — the idea is to apply an elastic strain that’d nominally constant through the lifetime of the device, giving you nominally constant (engineered) electronic properties.
I don't know much about materials engineering, but it's not clear from the article how Artificial Intelligence (which I'm assuming means Neural Networks) is used. Is it to simulate different strains? Or to correlate a set of parameters to final product properties?
I'm aware of efforts in statistical product design, where known mixtures, manufacturing parameters, etc. are collected into a database, and design-of-experiments is used to extrapolate into new products with properties that have never been made before. But this is mostly accomplished using statistics and design of experiments.
Agreed - this just seems to be an extension of the calculator story listed above, except using machine learning. Why wouldn't you use state of the art technology (including those labeled as AI) to help push any field forward?
looking at the paper in PNAS I think that there is a clever arrangement and application of deep nets in play here. As ever the trick is to couple deep domain understanding to clever application of AI techniques.
From what I gathered from the paper -- again I'm not an expert on what they're doing -- could the NN model have been replaced by any number of more parsimonious nonlinear regressors? (Not that an NN isn't a valid choice -- it is often used as a nonlinear regression model, but it is one of many valid choices in this scenario.)
Which prompts the next question: does anything that uses NNs warrant the term AI?
The use of NN here seems to be that of a model surrogate (i.e. model reduction).
There's no perception problem, logic problem, decision problem, or anything that we commonly associate with "intelligence".
Well. Maybe it could be neater (in an information theoretic sense) but DNNs are a bit fuzzy on that, and if you have modern compute, who cares? I think that there's many questions about the validation of machine learning models which mean that your question "could the NN model have been replaced by any number of more parsimonious nonlinear regressors?" is open, in the sense - we don't have a sharp way of deciding which is best, because one out of 100 million isn't informative for statistical reasoning.
The other question "does anything that uses NNs warrant the term AI?" is very, very difficult. Because as Marvin Minskey said "intelligence is a portmanteau term" by which he meant overloaded. It's full of meaning and non specific, so I like to say that AI is about technology and capability and not an association with human or animal cognition - which is the domain of Artificial Intelligence proper.
To a practitioner, "better" can be defined along well-known dimensions. Suppose you know your data lies more or less on a straight line -- you could fit a NN model or a run a linear regression. In this scenario, linear regression would be the "better" choice for some widely-accepted measures of "better" (interpretability, computational efficiency, parsimony, regularizability, etc.).
One almost never chooses NN just for the sake of choosing NN... there are well-understood trade-offs.
For instance, it is widely known that NN's generally require a larger amount of data than classical/statistical algorithms for weights to converge -- mostly because it's fitting a more general function than most classical ML algorithms. (The reason NNs have begun to show the results they have (vs in the 1990s) isn't just because we have more compute than before or that the theory has advanced significantly; it's also because we have more a lot more data to fit the more general function with.)
To the second point, I can see the point, but I wonder if the semantic meaning of the term is eroded by being overly encompassing. We could say a calculator implements AI.
In my opinion fundamental NN theory hasn’t really advanced significantly since the 1990s. I guess it was phrased ambiguously in my comment
But there has been a great deal of new techniques that make NN work better in practice like dropout, ReLU for vanishing gradients, CNNs and GANs for specific problem types, transfer learning, etc. The recent work with Neural ODEs show that the field is advancing in terms of ideas.
This is a good trajectory IMO. Many engineering fields work this way — find out new ways of doing things that work and then take a step back to see if there’s anything fundamental that links everything together. Practice precedes theory.
My undergrad is in Materials Science and I am very uninformed on the NN/AI stuff, but I completely agree that the term AI seems to be applied to all sorts of techniques with zero relation to true intelligence, seems more to be just working with patterns that have been fed into them.
The way they’ve formulated the problem, yes, I agree with you. There’s nothing really special about the deep learning model they used, and in fact they also tried gradient boosted decision trees.
"Although ab initio calculations such as those involving many-body corrections can provide accurate energy-band results, the scope of such calculations is somewhat limited to about 1,000 strain points because of high computational cost. On the other hand, by discretizing ε with a regular grid comprising 20 nodes separated at each 1% strain interval over the strain range of −10 to +10%, the computational model would entail about 108 band structures, up to five orders of magnitude higher computational requirement than what can be reasonably achieved presently. To overcome these difficulties, we present here a general method that combines machine learning (ML) and ab initio calculations to identify pathways to ESE. This method invokes artificial neural networks (NNs) to predict, to a reasonable degree of accuracy, material properties as functions of the various input strain combinations on the basis of only a limited amount of data."
Further down:
"We aim to describe the electronic bandgap and band structure as functions of strain by training ML models on first-principles density-functional theory (DFT) data. This approach leads to reasonably accurate training with much fewer computed data than fine-grid ab initio calculations and a fast evaluation time."
--
If I'm reading this correctly, it sounds like they already have a high-fidelity 1st-principles model that is computationally intractable to solve at scale, so they are using ML techniques to create an surrogate model that is computationally more tractable -- the AI modeling is a model reduction exercise.
That's a pretty good summary. There are many different sets of approximations and simulation techniques that make different trade-offs of accuracy/scale.
They're essentially using outputs of a higher-fidelity model to tune the free parameters of a lower fidelity model, and using ML to explore the parameter space efficiently.
The higher fidelity model is itself still an approximation, but there's a lot of interest in this approach and quite a few groups doing work like this targeted towards various material properties, since there just isn't nanoscale experimental data to train models that depend on nanoscale material features.
> If I'm reading this correctly, it sounds like they already have a high-fidelity 1st-principles model that is computationally intractable to solve at scale, so they are using ML techniques to create an surrogate model that is computationally more tractable -- the AI modeling is a model reduction exercise.
So in the end is this really different from simulated annealing (which appealingly actually sounds like materials science process)?
> One of my problems was to provide statistical advice to the people who were developing metals to be used in the blades of turbines. I had an enormous amount of data, and I had to construct a regression with five or six different variables having to do with the chemical composition of the metals.
> We estimated that it would take us three months to solve this problem using our desk calculators. In the whole country there was only one calculator—one computer, if you want to call it that—which could do this problem more quickly.
> It was up at Harvard. It wasn’t electronic. It was a whole collection of IBM card sorters. It was in a big, air-conditioned gymnasium, a tremendous collection of sorters all linked by wires. It did our problem for us in forty hours.
As he mentioned elsewhere, it did not work as expected back then.
[1] https://miltonfriedman.hoover.org/friedman_images/Collection...