From Philosophy to Machine Learning with Bruce Buchanan

Tom Mitchell (0:00)

Welcome to Machine Learning How Did We Get Here? I'm Tom Mitchell, your podcast host. Today's episode is Conversation with Bruce Buchanan, a philosopher turned machine learning researcher, who was responsible for a major milestone of machine learning advance in the 1970s. Bruce developed a symbolic learning program called Metadendron, that worked in the domain of chemistry. In particular, it worked on the problem of discovering how molecules would fragment once they are put in a mass spectrometer. The milestone is that this was the first symbolic learning program that discovered valuable enough knowledge that the knowledge it discovered was published in a scientific journal, in this case, the chemistry journal, in the domain that it was working in.

I hope you enjoy this conversation with my Ph.D. advisor, Bruce Buchanan.

Okay, today I have with me Bruce Buchanan, a early pioneer in AI, originally trained as a philosopher, spent most of his career as an AI scientist and researcher. Bruce, welcome. It's great to have you here.

Bruce Buchanan (1:39)

Thank you, Tom.

Tom Mitchell (1:40)

Let me just start by asking you about Mededendron. I should probably point out for the audience that Bruce was also my PhD advisor at Stanford University when I was PhD student in the 1970s. I was attracted to work with Bruce because he was working on a machine learning project which he called Mededendron.

Maybe first, Bruce, you could give us a very quick brief of what the Mededendron program itself did.

Bruce Buchanan (2:15)

Well, Tom, I think you should mention that you escaped from electrical engineering.

You were working with Bernie Widrow, as I recall. So I was very lucky to get to know Bernie because of you.

Tom Mitchell (2:36)

That's true. So Mededendron?

Bruce Buchanan (2:40)

Well, let's see, Tom, that story starts with the Dendrel Project, which was a classic search program given a generator of the descriptions of molecular structure, the planar graph descriptions. And the question was whether we could interpret the structure of unknown molecules from the mass spectra that were produced. And in the mass spectrometer, huge amounts of energy are fired at a small sample of a new chemical. And the electron bombardment shatters the molecule. And those, the particles that are produced, then are pieces of the molecule that show up as a bar graph. And the bar graph is really a plot of the molecular weight of a fragment plotted against the number of or the intensity of the signal of particles with that weight. So the DENTRO project was an attempt to look at a new mass spectrum and determine the structure. We did that primarily through knowledge engineering, and what came to be known as expert systems. I was interviewing a post-doc in Carl Gerassey's mass spectrometry lab, Alan Duffield, whom you may have known. Alan and I would sit looking at known mass spectra, and Alan would talk through his process of interpreting the spectrum to determine what was the structure of the molecule that produced it.

So that was a very labor intensive project, and the knowledge acquisition bottleneck, as it came to be known, was a major stumbling block in creating expert systems.

So years before we had even thought about the general problem of expert system, Joshua Lederberg was certainly, he and Ed Fagermann were principals for the Dendrel Project. And Josh had the bright idea that we could be learning the rules of mass spectrometry that produce the spectra, and we could bypass the knowledge engineering bottleneck. So he named the program MetaDendrel because it would sit above the Dendrel code and instantiate the rules that I was working out by hand with Alan Duffield.

Does that all make sense?

Tom Mitchell (6:10)

That's a great description. So then what if you isolate, then as you described the machine learning problem, which is the key part of MetaDendrel, is given these mass spectra that we observe, and these known molecular structures that go with the various mass spectra, given that kind of data as input and the output is a collection of rules that are discovered automatically by MetaDendrel of the form, if you have this substructure, then you can expect the molecule to fragment along these bonds and give rise to the calculatable mass spectral peaks. Is that correct, Seth?

Bruce Buchanan (6:58)

Yes, that's correct. But bear in mind that in order to get new data for one new molecule, required about one graduate student, one year to put the interpretation onto a new molecule. So we were very much limited to the number of spectra that Professor Karl Gerasi had in his notebook. And we could, with any single class of molecules, and he was very much interested in the steroids, of course, there were maybe a few dozen known spectra, but more likely on the order of 10