S2E7: Quantum Machine Learning with Dr Kosuke Mitarai (御手洗光祐)

What can quantum mechanics bring to machine learning? Take a listen to Season 2, Episode 7 of insideQuantum to find out!

This week, Dr Kosuke Mitarai (御手洗光祐) tells us about his experience at the cutting edge of quantum machine learning, the development of quantum circuit learning and being a co-founder of the startup QunaSys.

Dr Kosuke Mitarai obtained his Bachelor’s degree from Osaka University, followed by a Masters degree and a PhD at the same university, and is now an assistant professor at Osaka University.



🟢 Steven Thomson (00:06): Hi there, and welcome to insideQuantum, the podcast telling the human stories behind the latest developments in quantum technologies. I’m Dr. Steven Thomson, and I’ll be your host for this episode.

(00:16): In previous episodes, we’ve talked about some of the challenges facing quantum computing, both on the hardware and software sides, and we’ve discussed some of their potential applications. Quantum computing could even prove to be one of the most significant technological developments of our time. We’ll see what the next few years bring. One of the other biggest technological revolutions happening at the moment is the development of machine learning and so-called artificial intelligence. But what about combining aspects of both - can quantum technologies be used to enhance the machine learning revolution?

Today’s guest is one of the pioneers of quantum machine learning. It’s a pleasure to be joined today by Dr. Kosuke Mitarai, an assistant professor at Osaka University. Hi Kosuke, and thank you so much for joining us here today!

🟣 Kosuke Mitarai (01:04): Hi. Hi.

🟢 Steven Thomson (01:06): So before we get into the details of what quantum technologies can bring to machine learning, let’s first talk about your journey to this point, and let’s go right back to the very beginning. What first got you interested in quantum physics?

🟣 Kosuke Mitarai (01:19): Well, my father had a book about quantum physics for middle school students and things, so that was my first contact with quantum physics. And yeah, I read that kind of book at the age of 15 or something. And then, yeah, I was very interested in the concept of the superposition and all those, the experiments and things. So that made me interested in quantum physics and that still kind of excites me. The superposition principle of quantum physics is so really interesting to me still as a researcher in this field.

🟢 Steven Thomson (02:11): I see. And then when did you decide that you wanted a career in quantum physics, that - you know, were interested in it from a very young age, but was it obvious to you from the beginning that you wanted to work in quantum physics or was it an interest that developed over the years?

🟣 Kosuke Mitarai (02:26): So yeah, as I was kind of interested in the field of quantum physics as a middle school student, so I chose some special kind of school in Japan. So in Japan, usually we go to…so at the age of 15, we take exam, entrance exam for the high school, and then we go to that high school for three years and then go to university. That’s the standard route to the university. But I took a different way. There’s something called national technology schools, which is a five year school, and we go there just after we finish middle school. So we take an entrance exam to that school, and then we go to that school for five years. So we go to that school from the age of 15 to 20. So that school is for educating engineers in very industrial things, industrial fields.

(03:55): For example, I was a student of electronics and information science related classes. So that school works half like university and we study about very special and very specialized things in that school. So I was interested in quantum physics, so I took that course to study about electricity, which is kind of related to quantum physics. And information is also kind of related because we know that quantum computers can be made by quantum physics. So that’s kind of related and I chose that route to go. And so in that sense, I decided to go into this field at the age of 15 kind of, but I was not into this quantum computing field until I entered the university. So that depends on how you define this field. I decided to go to a related field when I graduated from middle school and I really specialized in this quantum computing field when I got into this graduate school. So I gradually, gradually went into this field.

🟢 Steven Thomson (05:45): It’s always interesting to hear how people got into quantum computing. Very often people started off in a background like yours where they studied electrical engineering or computer science and then approached quantum computing from the information theory perspective rather than historically I think people maybe came from the many-body perspective. It’s very interesting that in the last 10 years it seems like many more people working on quantum computing come from this computer science, from information theory, from electrical engineering.

🟣 Kosuke Mitarai (06:18): Yeah, yeah, yeah. I was from the electrical engineering, but…so my school told me about semiconductor physics and things, so I’m kind of physics background and also electricity engineering background. So I have really in the very, how do you say this, intermediate? Or how do you say this?

🟢 Steven Thomson (06:47): Yeah, yeah, exactly.

🟣 Kosuke Mitarai (06:48): In the middle of these things.

🟢 Steven Thomson (06:52): It’s always fascinating to hear the different backgrounds that people have. Everyone has a different perspective on how they approach quantum. So at the moment, you’re an assistant professor at Osaka University. How did you get to this point? Can you give us a quick summary of your career from graduate school to where you are now?

🟣 Kosuke Mitarai (07:13): Okay, so actually as an undergrad, I was…so in Japan actually, we are assigned to some lab in the fourth year, final year of the undergrad university. But I was not assigned to a lab which is related to quantum computing. I was assigned to a lab which is doing semiconductor research. And I did one year of research in that lab, and I actually have a paper about that field. There’s just one paper, which is very strange from thinking of my other research papers. I actually was doing very different research when I was an undergrad, but I decided to go to the quantum computing field in graduate school. And then, yeah, that was a game changer in my life. So my lab was run by Professor Kitagawa and also there was an assistant professor called Negoro Sensei in the lab too, and I, so I was kind of assigned to Negoro-san’s group to do some experimental quantum machine learning.

(08:53): And there I met Fujii, professor Fujii, who is a professor at Osaka University right now, but he was in a different university when I was a graduate student. But yeah, he and Negoro-san were doing joint research about experiments. So he - Fujii-sensei is the theorist and Negoro-san the experimentalist - and well, they are doing joint research and I was assigned to that project to actually do the experiment. And yeah, that was my start in the field, in this field. And then later Fujii-sensei gave me some ideas, like theoretical ideas to, to develop some algorithms for quantum machine learning. And then, yeah, I started the theoretical research in the end, actually. So I started as an experimentalist, but I kind of switched to the theoretical side of the quantum comput when Fujii-sensei gave me the idea. And after that was very fast, like Fujii-sensei and Negoro-san had the idea of founding a startup called QunaSys, and I was kind of asked to join the startup and I said, just yes, because that sounded very exciting to me. So I just said yes. And yeah, that was why I co-founded that company. And then after I was doing research, doing quantum computing with research as a part of QunaSys and as a part of this lab at Osaka university. And then I wrote some papers and then got the PhD and then, Fujii-sensei has become a professor at this school university.

(11:23): So in Japan, when someone becomes a professor, he usually takes assistant professors in the group and he selected me, chose me in that position. And that’s the brief history of my career after graduate school.

🟢 Steven Thomson (11:50): Wow. So yeah, you started in semiconductor physics, then experimental physics, then got involved in a startup, moved into theory, and now you’re an assistant professor - that’s an amazing number of very different things that you’ve done.

🟣 Kosuke Mitarai (12:04): It is, well, I have very few papers about experiments, so I couldn’t call myself an experimentalist right now, but yes.

🟢 Steven Thomson (12:17): I guess it must give you a good grounding though, in real life, in real experiments.

🟣 Kosuke Mitarai (12:23): Yes, that was a very nice experience to have, doing actual experiments, using quantum simulators and things. So that was a very nice experience to have to become a theoretical quantum computing researcher.

🟢 Steven Thomson (12:44): Yeah, I imagine it’s probably an experience that lots of theorists would benefit from, the chance to understand experiments a little bit more closely and see how they really work, what the real limitations are. So if you weren’t doing your current job, if you were not a theoretical quantum computing researcher, what do you think you would be doing instead?

🟣 Kosuke Mitarai (13:03): Well, that’s an interesting question. So the choice that changed my life was choosing a quantum computing lab at the graduate school. So if I didn’t choose that lab when I got into this graduate school, then I think I would be doing semiconductor research, same in the undergrad school. And in that case maybe I have become a semiconductor engineer at some company, at some private company, maybe for example, Toshiba and for example, Panasonic and - those Japanese big companies or something, I would imagine to be doing such a kind of job.

🟢 Steven Thomson (14:15): I see. Well, I guess luckily for the field of quantum computing, that’s not what you chose. So can you give us an overview of the field that you work in? What’s the big picture goal of your field and how does your work fit into that big picture?

🟣 Kosuke Mitarai (14:31): Okay, our goal is to make quantum computers practical. Right now, quantum computers are not practical. They do not solve very useful tasks. They solve very, well, they just solve very easy tasks which can be computed by classical conventional computers. So yeah, the goal will be making quantum computers practical, and my work is aiming that goal, aiming, aiming to achieve that goal. So I always try to think of some new ideas on how to use quantum computers for practical purposes.

🟢 Steven Thomson (15:29): What do you think is the biggest challenge facing the field then? What’s the biggest obstacle in using a quantum computer for more practical purposes?

🟣 Kosuke Mitarai (15:38): So the challenge is the noise on the quantum computer. And yeah, for example, when you think of the current state of the art quantum computer, when you do some operations on the quantum computer, then it fails with probability of one percent or something. So we can only do one hundred operations on the quantum computer, but for doing quantum computing, for doing practical computation, we need more - much more operations on the quantum computer. And that’s the challenge in this field. I think this challenge is very, very intrinsic to the quantum computer because qubits are very fragile. When you want to control the qubits, the qubits should be coupled to the environment like humans. So that coupler makes the qubit very noisy and consequently quantum computing very noisy, so very intrinsic. So we have to solve this problem. I think there is a movement to realize a technique called quantum error correction, which can correct the errors occurring in the computing process on the quantum computer.

(17:28): So if you can realize this error correction, then it can have much, much more computational resources on the quantum computer and then maybe we can solve much, much more practical problems on the quantum computer. But the problem is that if you can build the quantum error correction, this quantum error correction has very large overhead in the hardware requirements. So for example, if you want to realize one logical qubit - one protected, error free qubit - we need to physically implement thousands of cubits to make one error corrected qubit. So in that sense, if you realize quantum error correction, the computing resource number of qubits is reduced by a factor of, for example, thousands. In that case, theoretically we have to give some ideas to employ that kind of very limited number of qubits to be used for practical purposes. So yeah, the biggest challenge is noise and that challenge can be solved with quantum error correction, but that’s not happening very soon. In two or three years, that would not happen. That’s my opinion. But maybe in five years or 10 years, when you have 10,000 physical cubits or something or a hundred thousand physical cubits, then quantum error correction can, I think, be realized. But in that regime, we would have much less physical, logical or error corrected qubits. So we have to think of ideas to use that very limited quantum resource for practical computers.

🟢 Steven Thomson (19:58): The challenge for the moment then is that we have small noisy quantum computers, hopefully in time they will become less noisy and we need to find some practical use for these things.

🟣 Kosuke Mitarai (20:10): Yes, yes. That’s what, yeah, that would be my answer. Yeah.

🟢 Steven Thomson (20:14): I see. Okay. So if I were to summarize your own research work in one single oversimplified phrase, I might say something like quantum machine learning. Now our listeners might be familiar with machine learning as a classical phenomenon and they might be familiar with quantum computers. Can you tell us, what is quantum machine learning? What does quantum technology have to offer the machine learning process?

🟣 Kosuke Mitarai (20:41): As you may know, quantum computing is known to accelerate certain computational tasks like integer factoring and also simulating quantum mechanics. So actually it is not known if this quantum machine learning is very useful or not for existing data sets actually. So there’s no evidence in proving that quantum machine learning is useful for some dataset or some machine learning tasks, like practical machine learning tasks. There’s some very artificial setting where we can prove that quantum machine learning is useful. But in general, in general or in practice, we do not know if this quantum machine learning is useful or not. So I would say we are still trying to make quantum computing to be applied for machine learning, but we actually did not know if that would turn out to be successful or not. So yeah, the answer to the question would be actually, I don’t know.

🟢 Steven Thomson (22:18): See mean it seems like a very new field with obviously a lot of open questions. Is there a reason to believe that quantum technology will help machine learning? You know, said there are some examples where this can be shown to help. Is there a reason to believe that this will become more generally useful for other problems?

🟣 Kosuke Mitarai (22:42): Actually, I am kind of skeptical about using quantum machine learning for conventional machine learning tasks. For example, there’s ChatGPT, right? Yeah. That kind of thing cannot, maybe that kind of thing cannot benefit from quantum machine learning because classical computers are already doing very, very good, very, very good in that task. So in that sense, I think quantum practical quantum computing, I mean the error collected quantum computer would be 10 years away from now, and in that 10 years I think ChatGPT would become more and more powerful. So that would be, even if we can find that quantum computers can help, to do something more than current ChatGPT would need much, much trial and error to realize. So my opinion is that something like ChatGPT cannot be accelerated by a quantum computer. Right now we don’t have actual quantum computers, so we cannot experiment with quantum computers.

(24:21): So that’s my opinion. But on the other hand, I think we can benefit from quantum machine learning when we have quantum data sets. I would say for example, when you have circuits…when you have quantum circuits, that quantum circuit is very hard to simulate classically. So the quantum computing process is very hard to simulate because quantum computers have much more power than classical computers. So when you have quantum algorithms or when you have quantum circuits, then that kind of data set, if you have many quantum algorithms and many quantum programs and quantum circuits, then that kind of data set would be a nice candidate for applying quantum machine learning. Because classical computers are, I think…I believe classical computers cannot analyze those, analyze and learn that kind of data sets because if you can do this, if you can do that, that will imply classical computers are simulating quantum algorithms. So I think the best way to go for quantum machine learning is to look for some data sets that are more quantum.

🟢 Steven Thomson (25:58): I see. Okay. Yeah, that’s interesting. I guess this is kind of how quantum computing fits into the larger computing ecosystem anyway. Quantum computing will probably be used for very specific tasks and will never replace classical computers. And from what you’re saying, it seems like, yeah, yeah, quantum machine learning is the same. It will perhaps enhance some machine learning applications, but it will not replace classical machine learning. They’re both good for different things.

🟣 Kosuke Mitarai (26:26): Yeah, yeah, I agree. Yeah.

🟢 Steven Thomson (26:29): And in your own work in particular, you mentioned their quantum circuits, so you’ve played a key role in developing the field of variational quantum machine learning, and in particular a thing that you called quantum circuit learning. Can you tell us a bit more about quantum circuit learning and what the opportunities and challenges are of using quantum computers to tackle these sorts of learning tasks?

🟣 Kosuke Mitarai (26:55): Yes, yes. So quantum circuit learning is, well, a technique for using quantum circuits for machine learning tasks. And in that paper we proposed to use quantum circuits for…we proposed to use quantum circuits like neural networks where we tune or optimize the circuit structures to do some specific machine learning task. For example, when you train the neural network, we optimize the neural structure of the network and the weights in the network and realize some certain machine learning task, for example. For example, recognizing digits, handwritten digits for example. And so in that paper we proposed how we can do something like that on the quantum computer.

(28:11): Well, so actually that approach was, yeah, I think that approach attracted much attention in the field, but actually we do not find, as I said earlier, we have not found any practical application of this quantum circuit learning approach. And also there’s many other quantum machine learning approaches. But yeah, I think we have not found any very practical approaches to use quantum computers for machine learning tasks. So the challenge in this field is to find some. I think the problem here is we are not short of techniques to realize quantum machine learning, but we are kind of short of tasks that can be accelerated by quantum computers. So actually my recent interest is, is about making some machine learning task that can or that might be accelerated by quantum computers.

🟢 Steven Thomson (29:33): That’s very interesting. I guess particularly for someone from my kind of background who studies many body quantum systems and these, as you mentioned, these are things that cannot be simulated on classical computers. So I guess for me, I see a lot of promise in this sort of approach for trying to solve these kinds of quantum many body systems for I guess quantum chemistry, for drug design, for all of these highly quantum mechanical problems that we don’t have good ways of solving at the moment. Maybe these are areas where quantum machine learning will one day be a big help.

🟣 Kosuke Mitarai (30:07): Yeah, yeah.

🟢 Steven Thomson (30:09): You also mentioned earlier that you were part of a team that co-founded a startup company. So what made you decide to get involved in business as well as an academic career? I guess either one of those careers is very difficult, but doing both at the same time, that sounds like a real challenge.

🟣 Kosuke Mitarai (30:26): So on the company, the goal of the company is making practical quantum software the main goal, and that’s also my interest in the research. So that matches in the, so the company scope and my interest matches actually, and also…so why I decided to involve into this company is that, well, so as I said earlier, my supervisors also co-founded QunaSys, and they kind of asked me to join the founding process. So that made my decision to join QunaSys, but particular motivation involving QunaSys is like, no, I actually didn’t know the difference in the business and the academic career at that stage. I was like a first year graduate student when I founded QunaSys. So I was asked, and that sounded to me very exciting and very fun things to do. So I just said yes, that was the process.

🟢 Steven Thomson (32:01): Yeah, that sounds like an easy decision. Has it been interesting to see the business side of quantum computing as well as the research side? Or for you, have they both always been very aligned?

🟣 Kosuke Mitarai (32:13): Well, I think it has been aligned in the same direction because it has started and now also, now that company is a research company actually, and well, they do a lot joint research with the big companies in the chemistry field and also in, well…chemistry field companies are main customers of the QunaSys, but their main job is to do research. So in that sense, it’s very aligned to my daily jobs or daily research, but well, other business side, yeah, when I was working QunaSys, I did some presentation to the company to get that joint research contract or something. So that experience was very nice too. We have to…so when you are an academic, we usually don’t have very frequent interaction between this business side of the company, but we sometimes have as interaction with the research side of the company, but when you do a startup, we have to involve this business side of the company company. So that was, I think that was a very nice experience to have.

🟢 Steven Thomson (34:12): Just a few final questions to wrap up with then. One is a question that I ask every guest on this podcast, which is that physics historically has been a field dominated by white cisgender men, and I think hopefully things are improving slowly over time, but there is still a long way to go before we reach any kind of equality. So I wanted to ask, over the course of your career, have you seen attitudes towards diversity changing or improving, and have you seen a difference in attitudes in business and in academia towards diversity?

🟣 Kosuke Mitarai (34:48): Yeah, I think in Japan we are, there’s many efforts in making this inequality more, how is this…?

🟢 Steven Thomson (35:00): Reducing the inequality?

🟣 Kosuke Mitarai (35:00): Yes. Yeah. But actually the progress is very slow. My classmates are all actually all men. Actually, I think there was one female student in my class when I was an undergrad, and also when I was in the high school there, there’s only - as I said already, I went to a special kind of school where we only do engineering stuff. So there’s one, only one female student there too, out of 40 in the high school and out of 50 or 60 in the undergrad to school. So I think they progress very slow, but I think it’s making progress, actually. I’m not very sure about that. I think it’s making progress, and I think it’s very hard to motivate the female students to come to the field of this kind of engineering, the physics field.

🟢 Steven Thomson (36:17): Do you think it’s hard to motivate people to come into a field where there is already a lot of inequality and it perhaps doesn’t feel very welcoming?

🟣 Kosuke Mitarai (36:24): I guess so. I guess so. Yeah. Yeah. I think there’s many, there’s a very strong bias in the people’s minds to not to go, not to, so for girls, they don’t go to engineering stuff and also to physics stuff. We should be making some efforts, some much more efforts to make that bias smaller or something.

🟢 Steven Thomson (37:01): Yeah, definitely. It seems like a very important issue, I guess particularly we want as many people as possible to be working on quantum computing. We want the best people. We want to make people from all different backgrounds welcome, so that we have as big a workforce as diverse a workforce as possible.

🟣 Kosuke Mitarai (37:20): Yes, yes, yes.

🟢 Steven Thomson (37:22): Okay. One final question then, which is, if you could go back in time and give yourself just one piece of advice, what would it be?

🟣 Kosuke Mitarai (37:32): Well, I think I’m enjoying my life, so I would say just take the choices that I took in my life. Actually. I’m kind of…I’m really satisfied with my life, actually.

🟢 Steven Thomson (37:53): Perfect. Then you made all the correct choices and wouldn’t change a thing.

🟣 Kosuke Mitarai (37:56): Yes. Yeah. Well, I don’t know the other consequences of my other choices, so I don’t know about the other possibilities, but yeah, I’m quite satisfied with my life right now.

🟢 Steven Thomson (38:11): Fantastic. I think that is a great place to end this. So if our audience would like to learn a little bit more about you, is there anywhere they can find you on the internet or on social media, anywhere like that?

🟣 Kosuke Mitarai (38:27): Yeah, you can find me on Twitter and also on my website. I think you can Google my name on the internet to know about me. For example, you can Google me on the Google Scholar to look at my papers, yeah.

🟢 Steven Thomson (38:50): Okay, perfect. Now we will leave some links on our own site, insidequantum.org. Thank you very much, Dr. Kosuke Mitarai for your time here today. Thank you also to the Unitary Fund for supporting this podcast. If you’ve enjoyed today’s episode, please consider liking, sharing and subscribing wherever you’d like to listen to your podcasts. It really helps us to get our guest stories out to as wide an audience as possible. I hope you’ll join us again for our next episode, and until then, this has been insideQuantum. I’ve been Dr. Steven Thomson, and thank you very much for listening. Goodbye.