Pod Insights Y Combinator Podcast #94 - Joan Lasenby About this episode Joan Lasenby https://www-sigproc.eng.cam.ac.uk/Main/JL is a University Reader in the Signal Processing and Communications Group of the Cambridge University Engineering Department, and is a College Lecturer and Director of Studies in Engineering at Trinity College. Here's a list of her published work http://publications.eng.cam.ac.uk/view/people/jl.html.
In this episode we talk about Joan’s research into 3D reconstruction from multiple cameras and her interest in geometric algebra.
The YC podcast is hosted by Craig Cannon https://twitter.com/craigcannon.
***
Topics
00:32 - What's a tangible example of geometric algebra?
1:52 - What is geometric algebra?
6:47 - What resparked interest in geometric algebra?
7:42 - Why is it important?
11:32 - When did Joan start working on it?
13:27 - Rotations
17:17 - Computer vision in the early 90s
19:32 - Joan's fellowship at the Royal Society
23:32 - What's changed in computer vision since the 90s to allow for Joan's drone research?
30:07 - Machine learning in computer vision
31:52 - How Joan and her students are applying machine learning
35:02 - Unifying qualities of geometric algebra
41:02 - Joan's paper ending up on Hacker News
45:32 - Where could geometric algebra take hold?
47:32 - Running and mobility
48:32 - Where to learn more
Transcript 0:00 Speaker 00 Hey, how's it going? This is Craig Cannon and you're listening to Y Combinator's podcast. Today's episode is with Joan Lazenby. Joan is a university reader in the Signal Processing and Communications Group of the Cambridge University Engineering Department. She's also a college lecturer and Director of Studies in Engineering at Trinity College. In this episode, we talk about Joan's research into 3D reconstruction for multiple cameras and her interest in geometric algebra. If you'd like to learn more about Joan's work, I'll link it up in the description. All right, here we go. So Joan, as we, as we walk through geometric algebra, I think the best place to start might be through a more tangible example. So you're doing a project with drones here at Cambridge. Can you explain that first?
0:44 Speaker 01 Yes, so we're doing a project with drones. This is joint with the Architecture Department. And what we'd like to do is use drones to look at the built environment generally, and the built environment has, is made up of lots of lines. So what we would like to do is to do a lot of our processing, our vision processing with lines. And lines are much more difficult, classically, in computer vision than points. A lot of reconstruction is done with points. You get a point cloud. You can get structure from motion.
1:20 Speaker 00 So motion capture, for instance, when you see someone in the suit with all the ping pong balls, they're connecting points.
1:26 Speaker 01 They are, they are all points, exactly. But then even, no, not with motion capture, just with cameras that are moving, you can get points, you can match points.
1:35 Speaker 00 Okay.
1:36 Speaker 01 So we would like to do this with lines. Lines are difficult and our mathematical framework that we will use for this is geometric algebra.
1:46 Speaker 00 Okay. So let's define it. Let's start there. Yeah.
1:50 Speaker 01 Me, shall I define it?
1:51 Speaker 00 Yes.
1:51 Speaker 01 Okay. And we'll see if it ties in with
1:54 Speaker 00 See if it makes sense.
1:55 Speaker 01 What you've read about it. So, um, do you want some history?
2:00 Speaker 00 Yeah, let's start.
2:01 Speaker 01 So, um, so
2:04 Speaker 01 Grassman was a mathematician and Grassman had something called an outer product. So for example, I can take two vectors.
2:14 Speaker 00 Hmm.
2:15 Speaker 01 I can ha- put a wedge between them, a wedge product or an outer product and I get A wedge B.
2:22 Speaker 00 Hmm.
2:23 Speaker 01 So this quantity is now a different thing and I'll explain what that is. So Grassman had this outer product and Clifford, William Clifford, who was actually at Trinity, before he moved to London.
2:36 Speaker 00 Yeah.
2:37 Speaker 01 Um, so he came along and he extended this outer product. He effectively had um, an inner product plus an outer product. So for example, if I have two vectors.
2:49 Speaker 00 Hmm.
2:50 Speaker 01 A dot B.
2:51 Speaker 00 Hmm.
2:52 Speaker 01 Where the dot is an inner product, giving me a scalar and A wedge B, which gives me this other thing. It's a Clifford product.
3:01 Speaker 00 Okay.
3:01 Speaker 01 So seems like a strange thing to do. And he had an algebra for this, this Clifford product, which is called Clifford algebra. And it's been in the mathematics um, literature and research programme forever. Since Clifford died.
3:21 Speaker 00 Which was?
3:21 Speaker 01 Um, in the eighteen seventies. He died at the age of thirty four, I think, but um, so of TB. So he did a lot, but clearly could have done much more.
3:32 Speaker 00 Right.
3:33 Speaker 01 So, but as a, as a kind of, you know, applied tool, um, it wasn't, wasn't really used. And it was David Hestoness who came in the nineteen sixties and said, gosh, you know, uh, look at this, Clifford called it geometric algebra. I'm going to call it geometric algebra and I'm going to do all these wonderful things with it.
3:53 Speaker 00 Hmm.
3:54 Speaker 01 So effectively, I'll tell, so that's the, the background. The, the basis is, so imagine I have scalars.
4:03 Speaker 00 Hmm.
4:04 Speaker 01 So just numbers, um, vectors, so things with a magnitude and a direction.
4:10 Speaker 01 In 3D, we can think of this.
4:12 Speaker 00 Yeah.
4:13 Speaker 01 Bi- vectors, they're planes.
4:14 Speaker 00 Okay.
4:15 Speaker 01 Right? So things that have had two vectors, that makes a plane.
4:19 Speaker 00 Yep.
4:19 Speaker 01 So a plane would have a position and a magnitude and a handedness.
4:25 Speaker 00 Yeah, what is that?
4:26 Speaker 01 A handedness is, so suppose I have three points that make up a plane.
4:29 Speaker 00 Okay.
4:31 Speaker 01 I can, I can sort of go from A to B to C, or from A to C to B.
4:37 Speaker 00 Got it, okay.
4:38 Speaker 01 So if I take A wedge B to form my plane in vectors.
4:43 Speaker 00 Uh huh.
4:44 Speaker 01 Then B wedge A will give me minus the plane.
4:48 Speaker 00 Hmm hmm.
4:49 Speaker 01 Okay, so it's, it's a, you have to then start to think of these planes as, as geometric objects which have a sign as well.
4:57 Speaker 00 Hmm hmm.
4:57 Speaker 01 So in, so if I take vectors A wedge B wedge C, it gives me a volume. Again, it will be an oriented volume.
5:05 Speaker 00 Hmm hmm.
5:06 Speaker 01 If I live in four dimensions, A wedge B wedge C wedge D will give me a four volume.
5:11 Speaker 00 Okay.
5:13 Speaker 01 So at some point, I get to the highest element in the space. In 3D, that would be a volume, I can't go any bigger. And that has a special, that has a special place in my algebra. So imagine I have in 3D, scalars plus vectors plus bi- vectors plus tri- vectors. So points, lines, planes, volumes. I have an algebra which takes these things as objects.
5:39 Speaker 00 Hmm hmm.
5:39 Speaker 01 And I can add them, I can multiply them, I can differentiate with respect to them. So it's a kind of abstract concept, but it is amazingly powerful.
5:53 Speaker 00 Okay.
5:53 Speaker 01 So that's, that's the, the kind of rough idea of.
5:58 Speaker 00 Right, how Clifford algebra moved to geometric algebra in modern. And, and this is all predating computers, so this isn't being rendered anywhere yet.
6:08 Speaker 01 You mean in hardware?
6:10 Speaker 00 Yeah.
6:11 Speaker 01 So we have, we have a community, not a massive community, but there are lots of people who are very interested in, in, you know, potentially getting instructions for chipsets, et cetera. FPGAs, it is.
6:29 Speaker 00 Okay.
6:30 Speaker 01 So, and of course, we have quite a lot of
6:35 Speaker 01 programs that people have been building so that the community can test these things out.
6:41 Speaker 00 And so what rekindled it in the sixties to make, make it happen now?
6:47 Speaker 01 Um, so David Heston was doing his PhD and he was a physicist and looking at spacetime. So basically, his, his PhD turned into a book called Spacetime Algebra, from which
7:08 Speaker 01 basically in the eighties, people got hold of and started to get interested in. Even though David had been working on it throughout, since the sixties. Um, what started David on the Clifford algebra? I am actually not quite sure.
7:21 Speaker 00 Okay.
7:22 Speaker 01 I'll ask him.
7:23 Speaker 00 Yeah. And then so.
7:26 Speaker 01 But you mean, why? So, but the, he very quickly realized that this algebra simplified a lot of spacetime physics.
7:37 Speaker 00 Right, so that's kind of what I wanted to get at. It's like, why? Yeah. What was the realization that made it like, why is this important now?
7:43 Speaker 01 Why is this important? So in effect, um, um, spacetime physics, um, quantum physics, relativity is extremely complicated and is, you know, i- it's the area where you do have to have a lot of background knowledge and a lot of background mathematical knowledge. You do need to be proficient in, for example, for general relativity, you need differential geometry. You need a lot of these mathematical systems. You need a lot of tensor analysis. So, I mean, David could see that with this algebra, he could work entirely
8:27 Speaker 01 within an algebra of geometric objects, transformations between these objects. Everything stayed in the algebra. Transformations, I linear transformations functions were geometrically intuitive. There were mappings of objects to objects. They weren't just tensors. And we didn't have to go to another space, you know, a sort of dual space as you do in differential geometry. So things became the easy and he started to see that you could interpret a lot of things like Dirac matrices.
9:08 Speaker 01 Well, actually, they're not matrices at all. They were, you know, elements of the algebra and immediately they became easy to deal with.
9:17 Speaker 01 So his big motivation was that here was a unifying language for mathematics and physics, basically. So if you know this language, you can not only do rigid body dynamics in engineering and classical mechanics. You can do linear algebra without matrices and without tensors. And you can do complex things in quantum spacetime physics with the same mathematical system.
9:51 Speaker 00 And this is not theorized, this is proven at this point. It's all good.
9:55 Speaker 01 Oh, it's all good.
9:56 Speaker 00 Just to clarify for everyone.
9:57 Speaker 01 Yeah, it's all good.
9:57 Speaker 00 Yeah, it's not like this idea that maybe this might work someday.
10:00 Speaker 01 No, but the part of the problem is that we all, I mean, you might, your next question might be, well, if it's so wonderful, well, why doesn't everyone use it?
10:09 Speaker 00 Right.
10:09 Speaker 01 Okay. So A, Clifford died when he was thirty four and it was at the point where Gibbs Heaviside came in and produced vector algebra and the cross product. And so have you ever thought about the cross product?
10:25 Speaker 00 Nope.
10:26 Speaker 01 So if I have two vectors and I take A cross B, it gives me the vector which is perpendicular to the plane effectively. Now, that's all very nice, but it only works in three D. It doesn't work in any other dimension. Because in a plane,
10:42 Speaker 00 Right, so you're stuck.
10:43 Speaker 01 You're stuck. In four dimensions, there's no concept of a perpendicular to a plane.
10:47 Speaker 00 Okay.
10:48 Speaker 01 So, but, but of course, we all grow up with vector calculus, linear algebra, matrices. And, you know, then we have our research areas. And it's very difficult to actually, to listen to somebody if they come along and say, I've got this super duper new mathematical system that you ought to, you know, take notice of.
11:10 Speaker 00 Yeah, yeah, yeah.
11:11 Speaker 01 It would say, if someone comes up to me and says, look at this now, I will say, well, you know, I've got my own research and it takes me all my time to, you know,
11:18 Speaker 00 Okay.
11:19 Speaker 01 do what I'm doing.
11:20 Speaker 00 Okay. And so what planted the seed in you that you wanted to follow this path?
11:24 Speaker 01 Um, so you're the truth.
11:27 Speaker 00 Sure, sure. That was the goal. Yeah, okay.
11:30 Speaker 01 So the truth is, my husband, um, met David Heston's Antony, um, Lazenby. He, he became fairly obsessed with geometric algebra. He is a cosmologist. He could see that there was this thing which told him what the Cauchy-Riemann equations are, if you, you know, for people who know it. Um, that the Pauli matrices and the Dirac matrices in, in spacetime quantum physics were just, you know, really interpretable.
12:04 Speaker 00 Hmm.
12:05 Speaker 01 Um, he became obsessed with it, totally obsessed. Um, so Antony had a PhD student called Chris Doran. Chris and Antony, they've written a book on, on geometric algebra, geometric algebra physicists. And it was very difficult for me to actually, um, talk to my husband, uh, at that point, without actually finding out something about it.
12:32 Speaker 00 Okay.
12:33 Speaker 01 Because he thought it was one of the, um, most exciting things he'd ever seen.
12:38 Speaker 00 Hmm. And you were at the time just pursuing math, a PhD, what were you?
12:43 Speaker 01 I was pregnant at the time.
12:44 Speaker 00 Okay.
12:45 Speaker 01 So I was just about to have a baby.
12:48 Speaker 01 So I did have some time, actually.
12:49 Speaker 00 Yeah.
12:50 Speaker 01 Cause I was, you know, not that much time, of course, when you've got a small child.
12:55 Speaker 00 Yeah.
12:55 Speaker 01 But yeah, I was in it. I was, I'd gone back as an engineer. I was a postdoc.
12:59 Speaker 00 Uh huh.
13:00 Speaker 01 In engineering, I was doing imaging. I was imaging flames. Um.
13:05 Speaker 01 That was a two year post up position, but I actually then began to realise that these this algebra would be really useful in parts of engineering, particularly things which involve rotations. Hmm. And maybe I should talk about rotations. Mm hmm. Because in engineering and physics, rotation, the way geometric algebra deals with rotations is totally key. So have you heard of quaternions? I've heard of them, yeah. So quaternions are, well, if we go back, people rotate with rotation matrices. Yeah. So you have, say, in three by, in three dimensions, you have a three by three
13:49 Speaker 01 matrix, you act on a vector and it rotates it. Now, a rotation has just three degrees of freedom, which have nine, nine components in a three by three matrix. Mm hmm. And so they're all constrained. So rotation matrices are not numerically nice to deal with because you have to keep them on the manifold. You have to make, make sure if you change, if you update rotation matrices, you have to make sure it, it, um, uh, obeys the constraints. Okay. So, so that's one of the things we use. Now, in graphics, et cetera, and satellite motion, it's long been known that the best things to use is quaternions.
14:34 Speaker 01 Other methods are Euler angles, like a, a rotation about the x axis, the y axis, the z axis. Uh huh. But quaternions have been particularly nice because they are minimally parameterized. They have three components. They are smooth, they don't suffer from singularity problems, and Hamilton created quaternions as an extension of complex numbers. So complex numbers, everybody, people know that complex number i, if you, i effectively rotates in the plane, a multiplication by i. So.
15:13 Speaker 01 Hamilton spent many, many years in his later life trying to extend complex numbers to three dimensions and he came up with quaternions. So quaternions have these elements i, j, k, which all square to minus one. So it's like three imaginaries. So now to start with, that's pretty awful. And but everybody knew they were great, great to these libraries. People have been using quaternions since the, since the early days of satellites.
15:46 Speaker 00 Okay.
15:46 Speaker 01 So if you actually look at code, et cetera, you know, people will use this for rotations. Now, very early on, it, you know, it's in David's book that you see that if I, if I square one of these bivectors in 3D, let's take in 3D for now. If I square it, if it's a unit vector, I get minus one. So I have a real object that squares to minus one.
16:17 Speaker 00 Okay.
16:18 Speaker 01 Which is kind of telling me that I, complete the unit imaginary and this j, k are probably unit planes. And it turns out that the quaternions are just rotations. The i, j, k give you rotations about the unit planes in three dimensions. The x, y, the y, z, and the x, z plane. So immediately you see that you can have complete generalizations of things that do rotations in any dimension.
16:52 Speaker 00 And so at that point, you're doing this post-op position and you realize that you could apply it?
16:57 Speaker 01 Yes. Not to flames, but to
16:59 Speaker 00 No, okay.
17:00 Speaker 01 to mainly computer vision.
17:02 Speaker 00 Okay. And what, what was the stage, like the state of computer vision at this time? What year was this?
17:10 Speaker 01 Ninety three, nineteen ninety three.
17:14 Speaker 00 Okay, so not much happening.
17:17 Speaker 01 Yeah, but a lot of the, you know, it was, it was um really starting to move forward. There was no machine learning in computer vision, but it was all geometry. It was basically all geometry. So people had used project, well, you know, people had been using the ideas of projective geometry in computer vision for a long time, which is a four dimensional
17:41 Speaker 01 space, um, still with matrices. So I rotate and translate things. So I have lots of cameras. I want to find from my images, the rotations and translations between my cameras. Once I've done that, I can triangulate and I can do three D reconstruction.
17:57 Speaker 00 Okay.
17:58 Speaker 01 There was also at that stage, a lot of Bayesian statistics were coming into computer vision. So tracking things in up in images.
18:07 Speaker 00 Yeah.
18:08 Speaker 01 And finding most probable, the most probable tracks in, in, um, you know, crowds.
18:15 Speaker 00 Okay.
18:16 Speaker 01 So computer vision was really starting to take off.
18:18 Speaker 00 Okay. And so you, you're, you finish up your two year, you're creating these flames, these graphics essentially.
18:25 Speaker 01 Yeah.
18:25 Speaker 00 Uh, what do you jump into to actually give it a go?
18:28 Speaker 01 So I was extremely lucky because by that point I had two small children. Um, I, I don't know, have you heard the Royal Society?
18:37 Speaker 00 No.
18:37 Speaker 01 In what you, the Royal Society, um, the body in London.
18:41 Speaker 00 Oh, okay.
18:42 Speaker 01 Yeah, so you probably, you probably have because it's, it's has um journals and lots of, you know, historical people with big names in history, um, were fellows of the Royal Society. So the Royal Society has something called University of Research Fellows. And I applied to do, it was probably quite a step for them because I probably, I applied to do applications of geometric algebra in engineering. And they gave it to me. So I then had a five year effectively postdoc. And I could choose what teaching I did in the engineering department were very good. You know, I, I didn't, didn't need to do a lot of teaching or admin.
19:25 Speaker 01 But I basically had um kind of five years to try and get this off the ground.
19:30 Speaker 00 Kind of figure it out. And what, yeah, what was an example of an early project?
19:34 Speaker 01 So, um. So an early project was actually with, I think at the time, I mean, the internet wasn't really like it, quite like it was today. But there was some mailing list and there was somebody who called Eduardo Bairo Coritano, who is now in Mexico, but he put this email out on some list saying, anybody working with computer vision and geometrical algebra? So I actually contacted him. And those early days, we did quite a lot of translating all
20:15 Speaker 01 the classic projective geometry, computer vision, which was quite mathematical at the time, into geometrical algebra.
20:25 Speaker 00 Gotcha. Okay.
20:26 Speaker 01 And, I mean, this may sound kind of basic, but one of the really nice things about putting your problem, et cetera, into geometrical algebra is you have an origin, so some inertial frame, everything is with respect to that. So I rotate, translate, I don't have matrices, so I don't have to worry about coordinate systems. I don't stack up coordinate system upon coordinate system and then worry about what on earth that translation vector, what coordinate system is that in?
21:02 Speaker 00 Yeah.
21:02 Speaker 01 And with vision, where you're measuring things in an image, so what coordinate frame is that in?
21:09 Speaker 00 Yeah.
21:11 Speaker 01 It is quite confusing. And I know this for a fact because students are very confused about it all. When you've got a rotation, where is it with respect to in these complicated systems, et cetera? This just makes life very, very easy.
21:25 Speaker 00 Okay.
21:25 Speaker 01 And you, you almost can't go wrong.
21:28 Speaker 00 And was the goal of this five-year period to then apply it to some like product use or just do the basic research and see how it goes? What happened?
21:37 Speaker 01 At that time, it was just really seeing what we could get out of it.
21:44 Speaker 00 Okay.
21:45 Speaker 01 So I did some work with a company called FaceSpace, who's their motion capture company in the States. Not far from Berkeley and, you know, looking at algorithms to calibrate cameras. Because one of the things I haven't mentioned is
22:03 Speaker 01 that there is another aspect to this system, which is extremely useful. So I've said that I have these geometric objects. My rotations are objects, so I can write down coordinate free expressions. But not only can I write them down, I can differentiate with respect to them easily.
22:21 Speaker 00 Okay.
22:21 Speaker 01 So because I have this algebra of objects.
22:23 Speaker 00 Yeah.
22:24 Speaker 01 I can do calculus on them.
22:26 Speaker 00 Okay.
22:28 Speaker 01 And that's quite hard to do conventionally because you've got, people can do it, but you know, you're differentiated with respect to a matrix or a tensor or a vector and all this. So it's a much harder.
22:41 Speaker 01 You can do it component wise, but if I want to get closed form solutions. Um, doing kind of the analytic stage of the calculus is extremely useful.
22:51 Speaker 00 Hmm. And so the, the notion was that you could do it with less compute, like you could render these things using geometric algebra faster.
22:59 Speaker 01 No.
23:00 Speaker 00 Never. Okay.
23:01 Speaker 01 But, but so you can do it pretty fast. It's more, it's much easier to program up.
23:08 Speaker 00 Okay.
23:08 Speaker 01 It's intuitive. You can think of what you want to do and I can program it up at kind of this high level.
23:14 Speaker 00 Yeah.
23:14 Speaker 01 Underlying, you've got an algebra of a much bigger algebra than three dimensional space where I have three vectors. So actually computation.
23:23 Speaker 00 More going on.
23:24 Speaker 01 There's more going on.
23:25 Speaker 00 Okay.
23:25 Speaker 01 Yeah. But at a higher level, you know, I can, I can get code to do all this for me. And at a higher level.
23:30 Speaker 00 Yeah.
23:30 Speaker 01 I can certainly write rotate this object to this object.
23:34 Speaker 00 Okay.
23:34 Speaker 01 Et cetera.
23:35 Speaker 00 And then in terms of like the state of computer vision from then until now, what has progressed to make this, this drone project possible?
23:44 Speaker 01 Okay, so this is an interesting question. Um, so
23:51 Speaker 01 David Heston and Garrett Sobchik wrote a book. Um, this is a sort of
23:57 Speaker 01 real reference book of Clifford algebra to geometric calculus. Yeah. It's got everything in there. It's not a book you'd read. It's a book, it's a book that you go to find out. Like a reference. Yeah. I'm sure they meant it for you to read. And it, that was nineteen eighty four or so. Um, in nineteen ninety nine, David gave a talk at a conference. He'd done a paper with Hung Bo Lee about something which was in the final few pages of this book. Of course, not many of us had ever got to the final few pages of this book, which you call conformal geometric algebra. So this is truly stunning for graphics and for visionary projects, projects which use vision. So
24:43 Speaker 01 conformal geometric algebra is a five dimensional space. So imagine you take Euclidean space, the space we live in. You effectively add on two more vectors. One is the origin, one is a point at infinity. Projective geometry effectively adds on one more vector.
25:03 Speaker 01 So this is this five dimensional space. So you get um this five space and you think, well, okay, what, you know, what is this gonna get me? Well, what this gets you is that points, lines, planes, circles and spheres become objects in the algebra. They're objects. You give me a C, this big C is a circle. It's a trivector in my five dimensional space. And rotors, which are these, this class of objects that rotate, encompass rotations, translations, dilations and in the, uh, rotations, translations, dilations.
25:49 Speaker 01 So you've suddenly got, I, I can.
25:54 Speaker 01 For for example, setup, you give me three points. I immediately get the circle that joins that passes through, you give me four points, I give you the sphere. It's an object. I use my rotor, I rotate it, I intersect them. It's a it's a beautiful language for graphics.
26:12 Speaker 00 Hmm, okay.
26:14 Speaker 01 So, but because I'm a lot, a lot of night from nineteen ninety nine, lots of people, but fair number of people.
26:21 Speaker 00 So, yeah.
26:22 Speaker 01 Um, Leo Dawson, Stephen Mann and Daniel Fontaine, um, in Steve in Canada, um, Leo Atal in Amsterdam. Um, wrote a book, Geometric Algebra of Computer Scientists, and a lot of that was based on this conformal, just what you could do.
26:38 Speaker 00 Yeah.
26:38 Speaker 01 Just how easy it was to do things. So. Um, because lines are just objects, because I know how to sort of compare one line to another. Uh, because I can intersect lines and planes, keeping this nice algebra.
26:57 Speaker 00 Hmm.
26:57 Speaker 01 Um, you know, it becomes quite. Almost easy to see what I have to do to implement a variety of algorithms. So you've, I mean, it's quite hard to intersect two spheres. I mean, people can do it, but in.
27:16 Speaker 00 But it's easy to imagine how it works, but yeah, for, yeah.
27:18 Speaker 01 Yeah, but you know, you've got equations, but here we just do have operators that do it, we're between objects.
27:25 Speaker 00 Hmm.
27:26 Speaker 01 And one of the beautiful things is, and we live in a Euclidean, well, to all extents and purposes, you know, and what from what we see, it's Euclidean. Now, if you have a different underlying geometry.
27:37 Speaker 00 Yeah.
27:38 Speaker 01 So if you have. Um, hyperbolic or spherical geometry. Uh, then in this algebra, you have to change, you have it in conformal algebra. Euclidean geometry is the thing that keeps the point at infinity invariant. Then if I keep other things invariant, I get these other geometries. And I can just use my standard apparatus and do exactly the same things. Rotates, rotate my objects in hyperbolic space and rotate my objects in spherical space, move them around. So it is a beautiful language.
28:22 Speaker 00 Hmm.
28:22 Speaker 01 The geometry.
28:24 Speaker 00 Okay.
28:24 Speaker 01 And I haven't even touched on physics because a lot of, you know, de Sitter space, which a lot of the cosmologists work in is a different geometry.
28:34 Speaker 00 Okay, gotcha.
28:35 Speaker 01 Which you couldn't use this for.
28:37 Speaker 00 Right, well, which is why it was appealing to your husband thirty years ago.
28:41 Speaker 01 So exactly. So he, you know, he started out by being amazed at how things like Pauli and the Dirac matrices, spinners were all just trivial in this algebra. Then he started to realize that these complicated transformations tense, which are all written in tensor notation, are actually, if you put them in geometric algebra, they are mappings between real things like bi- vectors to bi- vectors or bi- vectors to vectors and things like this. And as soon as you see it in this way, it enables you to sort of, you know,
29:26 Speaker 01 interpret things and then maybe move, move on. And so, for example, so Anthony and Steve Goll, Chris Doran and David Heston are interested in a theory of gravity in flat space, which produces all. So, you know, their theory of gravity, you could understand. If you understood the algebra.
29:49 Speaker 00 Yeah.
29:50 Speaker 01 Basically.
29:50 Speaker 00 Okay, gotcha. And so the, as you were saying before, like the reason why people aren't picking up geometric algebra is that you become kind of in a certain track and then you know what you know. But, but right now, what are people using for modern computer vision to do comparable work?
30:10 Speaker 01 So computer vision is really massively advanced.
30:16 Speaker 00 Of course, yeah.
30:17 Speaker 01 So.
30:21 Speaker 01 Today, so people are really now moving from geometry to machine learning. You know, so you use, uh, using a little bit of geometry, but you are learning to segment things, recognize things by giving it lots and lots of images. But you know, you still have, you, we still have lots of geometric problems. So we still have to extract things from images if we've got moving cameras.
30:43 Speaker 00 Yeah.
30:44 Speaker 01 And things like that. But we have, I suppose it's a case of geometric algebra won't really give you anything that you can't do conventionally. But what it might enable you to do is to see how, how to do that thing. So for example, if I ask you how close is one line to another?
31:07 Speaker 00 Hmm.
31:07 Speaker 01 I have a way of doing that in my algebra.
31:09 Speaker 00 Hmm.
31:10 Speaker 01 If I could, I could sit down and write it in conventional.
31:13 Speaker 00 Oh, of course. Yeah.
31:14 Speaker 01 So, but, but could I have actually thought of that conventionally? Uh, probably not.
31:21 Speaker 00 Okay.
31:21 Speaker 01 Cause I'm not clever enough.
31:22 Speaker 00 Yeah.
31:22 Speaker 01 So, okay. So, you know, I, I need, I need a set of tools, which makes sense to me geometrically and physically. And I can then think, oh, me and, you know, other people can
31:38 Speaker 00 Of course, yeah.
31:39 Speaker 01 Think about how to extend that.
31:41 Speaker 00 Okay. And so in, in your day to day research, how are you then applying machine learning? Because many of your, your PhD students are working on exactly that.
31:51 Speaker 01 Okay. So, um, it's, it's almost impossible to avoid machine learning.
31:55 Speaker 00 Yeah.
31:55 Speaker 01 No, you can try, but you can't avoid it. So I have, I have two, um, a sort of strands. So some students are applying conventional machine learning techniques.
32:06 Speaker 00 Uh huh.
32:06 Speaker 01 Or conventional, you know, they, they change all the time. Neural networks, um, recurrent neural networks, LSTMs, two classical data like um, medical time series data. And I, I, I don't know how you can use geometric algebra for that.
32:22 Speaker 00 Right.
32:22 Speaker 01 Okay. So, um, and, you know, doing image segmentation, et cetera. So classical image segmentation, well, you know, um, I'm not sure how you can get geometric algebra into that. But as soon as it comes to.
32:40 Speaker 01 Anything involving like a moving camera, moving drone, multiple moving cameras, having streams of images, you want to match things, you want to triangulate, et cetera, then it is almost the only way I know how to do it.
32:57 Speaker 00 Okay, gotcha.
32:58 Speaker 01 So, so it's that aspect of it. And then can we extend it? So can we then, you know, we know how, how to parameterize in this algebra on my lines, my planes, et cetera. Can I learn them? So can I learn these geometric objects? So we were keen on analyzing both sort of moving images, which we're going to extract lines, planes, and also motion. So motion, can I actually parameterize my problem in terms of my geometric objects and learn them?
33:33 Speaker 00 Hmm, okay. And so for instance, like, you know, someone doing computer vision with a self-driving car, like, are they applying the same techniques that you're applying to get these lines? Or what would they use?
33:46 Speaker 01 I think, um, no. So, so, you know, there's huge amounts of research.
33:52 Speaker 00 Yeah, yeah, of course.
33:54 Speaker 01 And who knows what they're doing. But primarily, it's, people are using, if it's not Li, if it's not LiDAR.
34:03 Speaker 00 Right, yeah.
34:03 Speaker 01 It's single camera, multiple camera, lots of data.
34:10 Speaker 00 Yeah.
34:10 Speaker 01 Bayesian methods for segmentation. Following lines is easy. You know, it's not, that's not matching them or trying to reconstruct them. You're just kind of following them. It's really recognizing, um, if it's a person, if it's a road, if it's a tree, and, and you've got multiple sensors. So you know exactly where you are. GPS is.
34:33 Speaker 00 Right.
34:34 Speaker 01 Is really accurate these days. So.
34:39 Speaker 00 No.
34:40 Speaker 01 No, is the answer.
34:41 Speaker 00 Yeah.
34:41 Speaker 01 I think, I don't know.
34:43 Speaker 00 Well, I mean, so the question I'm kind of getting around to is like, where are the other applications, right? So like, you, in your instance, like, you're rotating in a camera, you want to map something, makes a lot of sense, like, and then you can move it around, but because you have the lines, you can recreate the shape. What are the other use cases?
34:59 Speaker 01 Well, you know, I do think that its real use is this, the fact that it unifies, it's a unifying language. So if I know this, I can just work things out more easily. Instead of trying, I mean, if people have worked with computer vision, they will know that often things don't work. So instead of a rotation matrix R, they try R transpose. And instead of a translation vector T, they try R transpose T. And they mess around till it works, because it's kind of confusing.
35:32 Speaker 00 Okay.
35:32 Speaker 01 Um, you've got no such problem here. It's very straightforward. Now, okay, that's not, that's not a good reason to use it, if you're proficient enough with classical techniques. But I, I can go into feels like thin shell elasticity.
35:52 Speaker 01 You know, and there is a student who is here at the moment doing this, that is quite an involved field. But if we put it in terms of, like, I have these rotors, I have a surface, I can stretch them, I can translate them, then it's just in the language of computer vision. It's not in differential geometry, it's not in anything else. It's, it's a totally understandable, um, process.
36:19 Speaker 00 Right, okay. So you could, you could see a world where then, you know, an artificial intelligence understands this, and then can apply it anywhere, and then you can then understand it.
36:30 Speaker 01 I, yeah, yeah, I think that's where its strength lies.
36:34 Speaker 00 Yeah.
36:34 Speaker 01 And in, and in new physics. I, I think, I think, you know, that because it unifies lots of different, um,
36:45 Speaker 01 Quantum mechanics and relativity and, you know, it hasn't, it's not, it's not unifying quantum gravity yet, but you can see that it's a, it is a, a system whereby you might be able to think of different ways forward.
37:01 Speaker 00 Right, okay, so it's actually a better tool.
37:03 Speaker 01 Better tool. It's a better tool.
37:06 Speaker 00 Okay.
37:06 Speaker 01 Because people are extremely clever, so they will find ways of doing things that are stunningly clever.
37:15 Speaker 00 Hmm.
37:15 Speaker 01 But, um, are complicated.
37:18 Speaker 00 Right, and so where do you see limits right now?
37:23 Speaker 01 I think, I don't think it's, it's not computational anymore because we're building up more and more tools. So I can, I can give you a, a website and, you know, you can try it out.
37:32 Speaker 00 Yeah.
37:33 Speaker 01 And you won't have to install anything because it's an online, you know, Python notebook. So you can, you can have a play with it. So we are, you know, we are a community which is certainly moving forward. Um, there still isn't, there aren't a lot of us. Um, so it's not taught, it's not taught, so, you know, teach my fourth year's image processing, but I don't teach them geometric algebra.
38:01 Speaker 00 Okay, and are, is there an interest there to learn or they're just like, I don't know?
38:07 Speaker 01 Uh, when you don't know about it, you, you can't really be interested in it.
38:12 Speaker 00 Okay.
38:12 Speaker 01 So it's a, it's a, and students have no hangups. They learn it and they think, great, you know, it's another tool.
38:20 Speaker 00 Yeah, yeah, yeah.
38:21 Speaker 01 And use it.
38:22 Speaker 00 Okay.
38:23 Speaker 01 So, um, it really, it enables you. I mean, I have a lot of confidence that if you give me a paper on, uh, cinch shell elasticity, it's going to be tough because there's tensors everywhere and I've got frames and frames and frames and dual frames and things, but I can eventually understand it.
38:46 Speaker 00 Hmm.
38:46 Speaker 01 Um, so these different fields, in fact, there's a colleague, Alan MacRobey, who's using a sort of form of it in structures.
38:55 Speaker 00 Hmm.
38:56 Speaker 01 So.
38:58 Speaker 01 There are fields, electromagnetism is also, I should have, I should have mentioned this, what is a field whereby you really get huge simplification. So I am fairly confident I can model electromagnetic fields and do maybe some new engineering
39:17 Speaker 00 Hmm.
39:17 Speaker 01 things using this, if I had time.
39:20 Speaker 00 Okay, well, I was, we were talking about programming before we started recording at lunch.
39:24 Speaker 01 Yeah.
39:24 Speaker 00 And about going from MATLAB to Python.
39:26 Speaker 01 Yeah.
39:27 Speaker 00 Right. Have people tried to create a port? In other words, like, oh, you have this traditional equation, we can port it over, and then you can understand it and see the value. Has that happened yet?
39:38 Speaker 01 Um, no.
39:41 Speaker 00 Okay.
39:41 Speaker 01 Not really, not really. I mean, there are, I mean, most of the code is
39:52 Speaker 01 enabling you to do things in a kind of transparent way, like I can wedge together two vectors, or I can multiply two vectors together.
40:00 Speaker 00 Yeah.
40:00 Speaker 01 I can in conformal, I can form my sphere, visualize it.
40:04 Speaker 00 Yeah.
40:04 Speaker 01 And I can do numerical, I can do lots of numerical computations with it, fast. So we can now do it quickly. But porting things is a kind of difficult one. Yeah, like you're saying, well, you know, here's my equations in terms of powerly spinners.
40:24 Speaker 01 What does it look like in geometric algebra? Well, it kind of looks the same, except your spinners aren't spinners. They're, you know, uh, so it's a difficult, that, that's a very hard.
40:34 Speaker 00 Yeah.
40:35 Speaker 01 Um, question to answer.
40:37 Speaker 00 Right, because it's not so simple as like, oh, this is how you call this function. It's not the same thing at all.
40:43 Speaker 01 No, not really.
40:43 Speaker 00 Right, okay. So this is sort of a weird, random tangent, but before, when we met, you were telling me that a couple of years ago, someone posted on Hacker News, one of your papers.
40:53 Speaker 01 Yeah.
40:53 Speaker 00 What was it, what was it about? How did all this happen?
40:57 Speaker 01 So it was, it was an invited paper, um, the Millennium Edition, so it was 2000, of Philosophical Transactions in the Royal Society. So we, we wrote a paper which was telling, just saying how.
41:14 Speaker 01 Geometric algebra was a unifying language and look at all these great things we can do in computer science, engineering and physics.
41:21 Speaker 00 Yeah.
41:21 Speaker 01 And this is the way the world works and this is, you know, you should, you should do it. So it did, you know, some citations and it sat there until bizarrely it was posted there. And it was a friend of mine who reads these things texted me and said, well, your paper is number one read on this hack and use. And I said, well, what's that? You know, is that good or is that bad? I don't know. So hard to know how it, how it emerged and hard to know how many, I mean, whether some people who looked at that and now are in our geometric
42:03 Speaker 01 algebra community who, you know, started to get interested and came along. Because throughout the world now, there are groups and, you know, in the 1990, there were not many.
42:17 Speaker 00 Yeah.
42:18 Speaker 01 There were little groups who were really keen on it. But now there are groups, oh, almost everywhere, they're not big.
42:25 Speaker 00 Sure.
42:26 Speaker 01 But.
42:28 Speaker 00 Interesting. Did you, did you read the comments when it went up there?
42:32 Speaker 01 I did. Yeah. I mean, they ranged from, wow, this looks really cool to this looks crazy, you know, and impossible to understand. Well, you know, I thought, oh no, that doesn't, you know, that wasn't the.
42:45 Speaker 00 We're not trying to communicate that.
42:47 Speaker 01 I know, I know. We didn't want to say that to, gosh, this looks as though it can do everything, you know. So they weren't, they weren't very detailed. It was more people beaming in and saying, do we short things? Although I don't know whether I read them all.
43:01 Speaker 00 That's the nature of the internet. You have to comment quickly. The post just falls off the front page and then your comments.
43:05 Speaker 01 Because of course, yeah, in the next five hours, it was probably.
43:08 Speaker 00 Yeah.
43:09 Speaker 01 Gone.
43:09 Speaker 00 Right, exactly.
43:10 Speaker 01 Sunk down to 100.
43:11 Speaker 00 Right. But, but what do the folks who are, are there naysayers? Is that a community?
43:18 Speaker 01 Sorry, what?
43:19 Speaker 00 Are there naysayers of geometric algebra?
43:21 Speaker 01 Oh, sorry.
43:22 Speaker 00 Not going to be a thing.
43:23 Speaker 01 Yeah.
43:27 Speaker 01 I wouldn't say so. I mean, there are people who, there are lots of people who think, yeah, that looks interesting, but really I can do it anyway. You know, and I have, I am, I am the world's expert in, you know, powerly matrices or whatever.
43:41 Speaker 00 Yep.
43:41 Speaker 01 Why do I need to put my matrix in terms of a vector? Um, and so, you know, people think in different ways. Um, to them, that is the way of doing it. That's always been the way of doing it. Um, so I don't think there are people who say, oh, this is complete rubbish. There are just people who say, well, yeah, but you know, what's it gonna get me?
44:04 Speaker 00 Okay.
44:05 Speaker 01 Um, and I don't think people, it's not the kind of thing where you could say it's wrong. Cause it's not wrong.
44:13 Speaker 00 Right.
44:13 Speaker 01 It's just a, a tool which, you know, depends whether you want to invest the time. Cause you think about multiplying vectors together and, and things like this.
44:24 Speaker 00 Yeah.
44:24 Speaker 01 It's, it's anti-commutative. It's not a commutative algebra. So immediately you throw away everything you've learned.
44:31 Speaker 00 Hmm, mm hmm.
44:31 Speaker 01 As a kid and through school and through university. So it, it makes perfect sense once you're into it, but there's a little, you know, there's a little hump.
44:42 Speaker 00 It's a learning curve. Yeah.
44:43 Speaker 01 There's a learning curve.
44:43 Speaker 00 You have to throw some things away. Yeah.
44:44 Speaker 01 And it's much easier for younger people to do that.
44:47 Speaker 00 Sure. That makes sense.
44:48 Speaker 01 Cause they have no real prejudices.
44:50 Speaker 00 Right. Okay.
44:51 Speaker 01 They, they say, yeah, it's another, another thing, you know, it's like machine learning. Oh, yeah. Another algorithm, I'll program it up now.
44:57 Speaker 00 Right.
44:58 Speaker 01 Um, and so, yeah, I don't, I think probably there are people who, who do think, well, why should I bother?
45:06 Speaker 00 Hmm.
45:07 Speaker 01 Um, but I don't think they would say, no, it's the wrong thing to do. You know, one shouldn't do that. One should use matrices and.
45:13 Speaker 00 Okay. So it, in your, you know, educated opinion, where do you see this really taking hold in the next, I don't know, like for thinking about it in the, in the practical sense? Um, so many people who are listening to Y C podcasts are like entrepreneurs or engineers or studying. Um, where do you see the people applying things like this?
45:34 Speaker 01 Um, so, um, clearly in the, in the fields of, um,
45:43 Speaker 01 Fundamental theoretical advances. So, you know, like physics or the aspects of physics that we're not quite sure about.
45:51 Speaker 00 Yeah.
45:51 Speaker 01 That's the big, um, as I, where I see it will have some effect. Um, because it enables you to think and it enables you to think in different ways. Um, in engineering, I think there are lots of fields and I don't know whether the, is there a killer application? Where if I could do it, everyone would say, whoa, this is the way to go.
46:16 Speaker 00 Right.
46:16 Speaker 01 I don't know. Um, as I say, I, I, I would, what I would like to see is that people, people had it in their toolbox.
46:27 Speaker 00 Yeah.
46:28 Speaker 01 Because, um, you know, there are lots of very, very clever people around who can cope with very hard, sophisticated physics and maths. There are a lot of people who are maybe not quite so clever. And they need the right tools in order to do these complex things. And I see it as actually this provides these people with a real, you know, a lot of people have a lot of geometric insight, but maybe not the mathematical sophistication.
46:56 Speaker 00 Hmm.
46:57 Speaker 01 And I think this was certainly sort of give them, um, a big advantage. Um, because it seems to me to be the way the world works. This, if it's a unifying language, it's what we should be writing.
47:11 Speaker 00 Yeah.
47:11 Speaker 01 Our equations in.
47:13 Speaker 00 Okay.
47:14 Speaker 01 Um, so I'm not sure that's answered your question.
47:18 Speaker 00 Oh, no, that's a great, that's actually a great answer. Yeah, I believe that. And I, I like it a lot too. Um, if you weren't working on this, do you have thoughts on where you might apply your, your energy?
47:31 Speaker 01 Well, it's interesting because I have always had, I'm, I'm a runner.
47:36 Speaker 00 Mm hmm.
47:37 Speaker 01 So I've been a runner since I was a kid. I.
47:42 Speaker 01 I am absolutely convinced that even if you don't run, you know, as you get older, you need to move and you need to keep your body moving independently. You need to make the muscles move independently. You need to keep healthy. And I, I probably would be doing something which tried to get everybody out and moving. And isn't that, that was not what you were expecting.
48:07 Speaker 00 No, that's great. Yeah.
48:09 Speaker 01 Yeah. But, you know, I look at people and as they get older and life could be so much better for them if they kept mobile and they, they looked after their body.
48:21 Speaker 01 That's quite obsession of mine, but anyway, yeah.
48:23 Speaker 00 Yeah, no, that's great. That's totally awesome.
48:26 Speaker 01 Yeah.
48:29 Speaker 00 I guess then my last question is if someone wants to learn more about this and actually start trying it out.
48:34 Speaker 01 Yeah.
48:35 Speaker 00 Uh, where should they go?
48:36 Speaker 01 Well, so lots of books available and I, I, people will probably be annoyed with me if I don't mention their books. But the, so the, David Hessens has three books, Spacetime Algebra, Probably Not the Place to Start, New Foundations of Classical Mechanics and Geometric Algebra, Clifford Algebra to Geometric Calculus. Now, they're, they're great books. My husband, Anthony, and Chris Doran have a book, Geometric Algebra Physicists. Leo Doherty, Stephen Mann, and Daniel Fontaine have a book, which is Geometric Algebra
49:14 Speaker 01 of Computer Scientists. Eduardo Baracarnachano has books which are more focused on robotics. So there are lots of books out there, lots of review articles, lots of conference proceedings, lots of code now. So people, you can kind of get code for.
49:34 Speaker 00 Yeah.
49:34 Speaker 01 MATLAB, for C, for Python. We've been using now a path, a package by an American guy called Alex Orsenovich, who wrote a Clifford package. And we've been integrating it into,
49:51 Speaker 01 I know my one of my students Hugo who you met
49:54 Speaker 00 Yeah.
49:55 Speaker 01 Has created a web version. So the big problem with a lot of these things, you have to download the package. You've got the package, you've got to get the number, you've got to, you've got to get all these other things.
50:07 Speaker 00 You've got to get it all working.
50:08 Speaker 01 Got to get it working. And if you're on Linux, it tends to be not available. If you're on Windows, it's like impossible.
50:12 Speaker 00 Yeah.
50:13 Speaker 01 Um, not impossible, but your average person would probably say, you know what? No.
50:17 Speaker 00 Yeah.
50:17 Speaker 01 But we, we're working to try and get a web version. So people can actually go on.
50:24 Speaker 00 Hmm.
50:25 Speaker 01 Try it out with some readme files and, and a bit of graphics. So you can see it. So there's loads of stuff out there at the moment.
50:31 Speaker 00 Where can they find that if they want to? Maybe, is it on your site or?
50:34 Speaker 01 Um, it is Hugo. So Hugo and Alex have been working on this and I can give you the web address.
50:44 Speaker 00 Okay, we can put it in the blog post.
50:46 Speaker 01 Yeah, yeah.
50:46 Speaker 00 There we go.
50:47 Speaker 01 They're just re- releasing it to the world.
50:50 Speaker 00 Okay.
50:51 Speaker 01 Yeah.
50:52 Speaker 00 Perfect.
50:52 Speaker 01 So it'd be good to have people te- test it out and say, email back and say, this doesn't work.
50:57 Speaker 00 Yeah. All right. Thank you so much for your time.
50:59 Speaker 01 Okay. Been a pleasure.
51:02 Speaker 00 All right. Thanks for listening. So as always, you can find the transcript and the video at blog.ycombinator.com. And if you have a second, it would be awesome to give us a rating and review wherever you find your podcasts. See you next time.
Previous #93 - Peter Reinhardt Next #95 - Eli Schwartz Page updated October 13, 2026 · transcript and summary are machine-generated