Transcript | Quantum Error Correction with Quantum Machines

[Music] We often hear that the path to fault-tolerant quantum computing will require error correction. How will this technique work? We cover this and other scaling technologies in this episode of The Post Quantum World. I’m your host, Konstantinos Karagiannis. I lead quantum computing services of Protiviti, where we’re helping companies prepare for the benefits and threats of this exploding field. I hope you’ll join each episode as we explore the technology and business impacts of this post quantum era. [Music] Our guest today is the CTO of quantum machines. Yonatan Cohen, welcome to the show.

Hi, thank you for having me. I’m very excited to be here.

Awesome. Yes, let’s dive in. There’s a lot to cover here. But first, I just wanted everyone to get a sense of who you are and how you found your way to quantum computing. I know you have the traditional route, right? A PhD in physics, of course. [Laughter] That’s a good traditional start. So, I guess first question would be like, were you interested in quantum computing when you were doing that, or just did you find your way to it after?

I was interested in quantum computing. Indeed, I think it’s the standard way, at least today in the industry, which is to do a PhD in quantum physics, but I actually started to become interested in quantum computing in my bachelor’s degree, which I actually did in the United States, Seattle University of Washington. I have a very good teacher for - I took this course Quantum Computing 101, and that kind of blew my mind. I saw that it’s a much better way to actually study quantum mechanics than to start with atoms and electrons, just start with these abstract things, which are these qubits, and it really blew my mind. That’s where I started thinking about quantum computing. I was always very interested in entrepreneurship, but it was not until the end of my PhD that I found my partners and started a company in the field.

That’s great. Once again, we’re hearing about this idea of a professor who changed someone’s view. So, I think that’s terrific. I hope to all the teachers out there, they hear that. You can have a really big impact at a key moment in someone’s development.

Huge impact. I mean, this guy’s name is Boris. He’s just an amazing teacher and amazing person, and that’s what impacts you, right? 

Yes, absolutely. Then, it looks like not too much after you got your PhD, it was onto quantum machines, right? So, what was the first gem of an idea you had then? What was the passion for what you wanted to bring to the industry right out of the gate like that?

Taking maybe a step back, so during my PhD, I was working on looking for what’s called topological quantum states, which have a potential for them to allow us to make qubits that are protected from errors. So, that’s kind of where my research was. But at the same time, me and my friend from the PhD back then, Itamar Sivan, were also thinking, “Oh, wow. It would be awesome to start a company one day.” So, in parlance to our PhD, we started an entrepreneurship program at the Weizmann Institute in Israel, which is where we did our PhDs. It was the first time that there was such a program at the institute, so we get a lot of attention and unity from investors and so on. Then so, when we graduated, “So, what do we want to do?” It was clear that we would start a company. We actually didn’t start right away from quantum computing. We had lots of stupid ideas. Machine learning, all that stuff that everybody’s attracted to, and at some point, we saw that investments are starting to rise in quantum computing. It was 2018. That’s where we were kind of “Okay, something’s happened in quantum. Maybe we should look at that.” Because that’s really the only thing that we know. What was very important for us, the one lesson we kept from the entrepreneurship program that we ran ourselves was we wanted to be in a place, in the stack, in the quantum computing stack, where there is a current need for products, not future need. Of course, the need in the future the market will grow significantly once quantum computing becomes super impactful industries and then the requirements for our products are going to grow, but even today, there is a large enough market for us to grow and to become a successful company. So, that was important for us. It might be already today. The other thing that led to what we do with quantum machine was that we knew our third co-founder, who just came back a year before from his postdoc at Yale University in professor Schoelkopf’s group, which is one of the leading superconducting qubits groups in the world today. During his postdoc there, he really developed very, very significant technology there that allowed him to perform the first experiment actually have quantum error correction done on superconducting qubits. So, that was a big milestone. Of course, when we were like, “Okay, there’s investments going into quantum computing. We should call on Nissim.” So, we called Nissim, and then the three of us start the thinking together and we realized that there is a place in the stack where there is current needs, and that there is also a current – so, there’s a current bottleneck and very significant technology to be developed that’s going to be important going forward. That is what’s called the control system of the quantum computing. So, we went to do that.

Okay, so let’s dive into some of these points in order. First of all, you’re interested in topological quantum computing, and so you must have been excited by Microsoft’s Majorana Fermion topological gap paper last year.

Look, to be honest, this was what we did in our research, it is not exactly what we’re doing in the company, but definitely looking at the physics of topological quantum system is one of the most exciting physics that’s happening today in the world, I believe. Seeing such great progress then with this system is definitely exciting. I do think that there is still a long way to go to become the base. Based on this topological…

Yes, it’s going to be a while before we have Majorana qubit. [Laughter] I think it’s going to be a little bit. Of course, that would be virtually error free, so that’d be great, but because we don’t have that, the more logical approach right now is to also be working on error correction. So, I definitely wanted to spend a few minutes digging into that. If you want to maybe take a step back and explain some common error correction methods and what your company’s working on, so that way our listeners can get a sense of what that really means, because I’m sure they conceptually have an idea, but just to dig in a little bit.

Sure. First of all, why? Why do quantum error correction? So, we have already quantum computers, and they didn’t work, at least to some extent they work. We are in what we call the NISQ era or noisy intermediate scale quantum devices era where our current quantum computers, they have a very high error rate. So, if I run the same experiment twice or same circuit on my quantum computer, I can get different results, and not just because of quantum statistics, but also because of errors. There are certain algorithms which might be useful running on these NISQ type quantum computers, but for one, they’re heuristics. So, we don’t know whether they will work or not, we just need to try that and see. Hopefully, that would give us some advantage, but we don’t know that. Also, we do know about quantum algorithms that will give us a significant computational advantage if we could run a quantum computation that doesn’t have any errors. So, that’s the reason why - and that’s called a fault-tolerant quantum computer. Quantum computer where I could run my entire circuit without anything else. Now, the mainstream way to get to a fault-tolerant quantum computer is to perform quantum error correction, which is a way to actively correct for errors. So, it’s you could just say, “Let’s just build the qubits that doesn’t have,” right? The topologic of qubits can disappear, qubits that’s super isolated from its environment. That’s one way. But today, the current error rates are too high. So, the way people – the least that we will be able to lower the error rate is not just by just lowering the physical elements, but by taking many physical qubits and holding with those many physical qubits a single logical qubit. This logical qubit, by performing certain control protocols on it, meaning that we measure some of the qubits by taking these measurements in real time, we can calculate what were the errors, what were the physical errors, and even correct those errors. This is called quantum error correction, and then we can make our logical qubit have a much lower error rate than the physical qubits. 

That’s a great basic…

This error correction - by the way, Google, just a week ago, published a new result that are very important because for the first time in superconducting qubits, they show that when you grow the number of qubits - so until now, many people have tried quantum error correction and you can see some building blocks, right, but the key here is really that you wanted, as you add more qubits, more physical qubits to your logical qubits, the error rates go down. Until now, it was the opposite. Now, it’s starting to become - it can go in the right direction.

Because if you cram too many qubits in one space, all of a sudden you’re introducing new noise and that becomes its own issue.

Exactly. 

So, we need to find that optimum size. I was going to ask you about that Google thing too. [Laughter] So, if you want to explain for our listeners what’s significant about the Google achievement and how you view the number of logical qubits by technology. Do you have an overview of what you’re seeing in the industry? What would be good guesses at how many physical translates a logical by, let’s say, type like transmon, trapped ion, and any kind of use like that?

It’s a very good question. First of all, the significance of the Google experiment I think that, yes, it’s very significant. Very significant to show that as you scale up your code, what’s called the code distance, which basically means how many physical qubits are used to my logical qubit. I can lower the errors and not - that means that I could continue to scale in theory. Again, in theory, I could continue to add more and more physical qubits and lower my error rates as much as I want. That is how I build a fault-tolerant control. So, of course, this is in theory. The fact that they went from close distance three to five and the error rates goes down, does that mean that it will work when they go from five to seven or seven to 700? It’s a fairly significant first step. Very important. Actually critical in order to continue to believe in this route. So, that was that. On the other end, yes, they think the road is still very long. So far, the qubits people talk about with current errors rates, about 1,000 Physical qubits to make a logical qubit.

The terrifying IBM number. [Laughter]

Exactly. That is still quite far. You need - and that’s for a single logical qubit, and you need any logical qubits in performing gates. The good thing for quantum machines, by the way, is that means you need a very large control system to control these qubits. [Laughter] In fact, one of the key aspects that you need to perform quantum error correction effectively is what’s called real-time feedback. That means that I could measure a qubit while the algorithm’s running, when my circuit is running. I measure a qubit. I could do classical calculations on these measurements. I can measure bunch of qubits and I could, based on the results, do some classical decision making in my classical hardware that’s [processed then] and even responds in real-time, say, “Okay, if all of these physical qubits were in this state, maybe I shouldn’t operate on my logical qubit this way or that way, and that in this way, I could correct the errors.” So, this specific capability, and it’s called real-time feedback, of the control systems is one of the key advantages that we offer in our control system with quantum machine.

Okay. So, your control system doesn’t just do a cookie cutter approach. It doesn’t just say, “Oh, I have 10 qubits. I’m going to try and turn them into one logical one exactly the same way every time.” It actually looks at what’s going on in a particular environment. It makes adjustments. Is that correct? To try and to get closer.

Yes, make some measurements - exactly. The user could make measurements of subsets of qubits. Based on this subset of qubits realize what is the state of - that is partially what is the state of some of the other qubits, and based on that, make decisions in real-time, what gates do I apply to certain qubits to fix my errors? It’s actually very interesting, because we started from topological qubits, and I think one of the very interesting areas that are being explored these days is can we make a topological state from regular qubits? Because regular qubits could simulate any other states, right? So, I could use them to simulate a topological qubit or to actually create a topological qubit and then this topological qubit is protected. So, in fact, taking a bunch of physical qubits and making a topological qubit out of them is a form of error correction by itself.

Yes.

That area of research is now very trendy. This one also requires a lot of real-time measurements and feedback from the control system.

Yes, that’ll really hurt some people’s heads quickly. They’re familiar with the idea of taking a classical computer to simulate a qubit. Now, you’re simulating a different type of qubit [Laughter] with a qubit that doesn’t meet that criteria.

That’s true. 

Yes, that gets pretty trippy, I suppose. So, are you actively looking at that right now also at your company as something to offer? That’ll get us to the next question too, because I also want to ask like, who your typical customers would be for this control platform.

I will start from the second question actually, because, yes, our customers are everyone who was developing quantum computers. We’re selling a component of the computer. Okay. So, for example, if you’re developing your quantum processor, you need a control system, or actually, a full stack quantum computer, you need a control system. That’s going back to the beginning. We wanted to be in a place where people who are working in this field today, they need control systems. So, whether they buy it from us, whether they make it themselves, or whether they buy from competitors, it needs to be done. So, we’re working with many of the leading players, start-ups, big companies, academic groups, that are developing these qubits. We work all kinds of qubit types. Superconducting qubits, spin qubits, trapped ion, atoms. With many of our customers, we’re also exploring such as these, as, for example, demonstrate and edit. They’re just.

Are you then seeing all the different types of error correction play out on your platform? Do you have to modify? The name of this, this particular thing, that’s the quantum orchestration platform, right?

Right. Correct.

So, are you then able to see, are you getting feedback from customers about the different types of error correction they’re working on? Or does that kind of like stay with them at that point?

Yes. It really depends on the customer. Certain customers are very, very collaborative; certain customers are less. It really depends on what they do and what’s kind of important, but yes, we do have a lot of visibility to what’s happening today in the field. The difference between these different kinds of qubits. What types of error correction protocols people are trying to run and what are the different requirements, because they could vary significantly. So, in simpler language, qubits, for instance, this is the most demanding requirements from a control system if you want to run quantum error correction because everything happens very fast. So, you need to do this measurements and processing and feedback. They mentioned you need to do it insanely fast. While, for example, in trapped ions, they live much longer, so you have much more time to, measure the qubits, think what you want to do and then performing the qubits. So, we do see significant differences and significant differences in the requirements from the control system.

Now, I believe one of the paths to scalability is to identify the optimal number of qubits for any technology and then just build a lot of those modules and then interconnect. Do you see your hardware and software sort of living in the interconnect space as well down the road? Are you working towards…?

Yes, definitely. I think that that’s one of the advantages of our system, that it’s very versatile and very programmable. That’s why we managed to work with all these different qubits types, and that’s why you can also combine qubit types. Some of our customers do that. In many cases, it’s actually quantum networking that people are combining different types of qubits, because you have the photons, the spin qubits, and trying to move information from here to here. So, photons are very good to move information from one place to another while a spin qubit maybe is very good just to preserve information. Then, we see it, and people are actually using our systems very successfully to kind of combine different types of qubits.

That’s pretty exciting. We’re seeing a lot in that. Like IBM System Two is a great example. They’re claiming that with that new infrastructure, they’ll be able to support something like 16,000 qubits in one giant array, but they’re not really telling us what that interconnect is. So, do you see your type of approach being used in something like that, where like multiple machines are already going to be connected?

Yes. I think what’s going to be important also in these types of when you start connecting different quantum chips where you could relatively easily create an entanglement locally, but it’s harder to then spread the entanglement around to enable a chip. There, I think, having this real-time classical logic, classical processing and feedback, is actually going to be even more important because it can help you basically lower the overheads, the quantum overheads by taking certain measurements and then responding on other qubits based on the measurements that you took with your control system. So, it’s kind of like how much can I offload from my quantum processor to my classical compute system, which is this control system. That becomes more and more important as we scale up and as we get this weak links between different islands of qubits.

So, how would you describe your role, your company’s role, in things like hybrid? This idea of using classical and quantum together? Is it just in that control area that you just mentioned? Or are there other areas where you actually facilitate more major parts of a computation to be on a classical machine and then hand it off? 

It’s a great question. So, there are two pillars to this. First is the control system itself. We were, in fact, the first to really embed general classical processing units into the control system in such a way that you can really combine quantum operations and classical operations in the same code, which is our programming language, QUA. So, QUA really was the first to allow you to do general classical processing. You can define variables X, Y, Z and say Z= X+Y or X²-Y or X/Y. Anything that you want to do with classical processing, like in any other programming language, but the same operations, same commands, classical processing commands can be embedded now in the quantum controls sequence. So, with the gates, actually, the policies that we play to the qubits like shape the processes and they can shape it based on this classical processing. It’s a very, very low-levelled stuff in quantum computing, but that now mixes quantum and classical processing together. That’s kind of control. That’s critical, again, for quantum error correction, quantum error mitigation, calibrations that you want to do. Today, you are going to be - fix the errors with quantum error correction, we have huge drifts in the systems, of the system drift. Just by doing measurements and real-time feedback, we can correct those drifts. So, that’s kind of hybrid already on that it’s very lower level of the controller. Then, of course, you want to also get out the controller because there is a limit to how much compute power you can put so close to the qubits. So, it’s very important to start building the right integration between the quantum computer to the general classical compute systems that we have, whether it’s cloud or high-performance computers. That’s something that we are now with quantum putting a lot of emphasis on. We already have some few projects where we are integrating quantum computers into more scalable compute. However, specifically in Israel, we’re building the Israeli Quantum Computing Center, where we’re going to have a small HPC, but nevertheless, a high performance computer that’s tightly… 

They’re cheap. 

Yes. The more important for us is less how much compute power we put there to begin with, but this tight integration, the fact that we can really integrate the quantum computer to become another resources in an integrated classical and quantum system, and you can run these hybrid algorithms very efficiently on this on this system.

I do feel that an HPC needs to be physically located very close to quantum computer. Do you think that latencies and things come into play?

That’s a very good question. So, it really depends on what is the application that you’re trying to run. If, for example, you’re trying to perform the quantum error correction, and you want to use the HPC for doing the coding of the errors, then you need very low latency and you need to be close, very close with qubits. In some of the other hybrid applications, like algorithms, VQE and QAOA, you might be able to get away with having not such close physical connection to begin with, but then sometimes, if you’re trying to do certain optimizations, for example, as part of this hybrid loop or you want classical compute, quantum compute, but then all of a sudden you have this very heavy classical computer that you want to run inside the loops for doing some optimizations, then you might, again, want to lower the latency. What’s most important in those types of applications is maybe less to begin with, that we’re physically close, but more that we can schedule the resources of the HPC and the quantum computer in a unified way. Okay. Because typically, what you do is you run the classical processing, then maybe someone else is taking the quantum computer and then you need to wait for the quantum computer to be ready, then you want something in the quantum computer, but then your classical computer is waiting because maybe someone else is using it. So, being able to co-schedule the resources so that they get - like I do an HPC, I asked for a few CPUs and a few GPUs, and I run my entire application in these resources that I scheduled for myself. That is something that we’re working on hard with quantum. 

So, the current model for logistics purposes is that the big cloud providers, Azure Quantum, Bracket, AWS, they typically have the quantum computers elsewhere. They reach out to them. So, for certain applications in the future, it might benefit them to have one of these boxes sitting inside their cloud stack, somewhere physically, so that way, they can do some of these more optimized runs, right?

Yes, yes. I think at the end, people will want to set up quantum computers in their physical data centers. It’s not just for performance, but for a lot of reasons, but performance is definitely one. At the end of the day, that’s going to be important. Yes.

Okay. Going back to QUA for a second. For those listening, it’s Q-U-A. That programming language is really about the pulse-level control. It’s not about building a whole circuit and then sending it, right? It’s not some new interface for an end user to send a program, let’s say.

Right. It’s about pulse-level control, and then it’s about this integration of classical and quantum processing in the same code. When we say positive, and we mean that we don’t just say, “Yes, operate. Apply this gate to this qubit,” but really say play this pulse shape to this channel of the control system that goes to this qubit and then play another part, and we also can control the timing, when are we planning, because the quantum computing is because of the error rate that we have, the timing of these operations can become important. Now, coupled to that the classical processing. So, the fact that I could say, “Play this pulse, but change its shape according to some classical calculation that I performed.” Okay, will be on the measurement results, so maybe change the shape of the pulse based on some measurements on the qubits just 200 nanoseconds before, because remember, I can control timing. So, I can really do all this super low-level things. Today, you can say in some sense, unfortunately, this makes more interesting, it’s really critical to optimize on that level in order to optimize the performance of quantum computers and to push forward towards getting up quantum computers. 

Yes. So, do you anticipate in the future having some kind of like API hooks or something that let coders talk to QUA and say, “I need a little more in this area. I need the qubits that are optimized for this operation,” anything like that? Like a little more interplay?

Yes. I think that will be - like we already have integrations with higher level programming languages. For instance, OpenQASM 3. We’ve done a lot of work to integrate with OpenQASM 3 because we believe the standards are now becoming very important. They’re actually going in the direction that we’ve pushed for now four or five years to integrate classical processing with the quantum circuit. So, OpenQASM 3 has done a lot of these things. So, in some sense, except for the fact that it’s more gate level, even though it does allow us for some pulse-level, it starts to resemble in many aspects in that it allows you to integrate classical processes into your code, into quantum code. That’s why we’re integrating with those higher-level languages that also are becoming now standards to even higher-level frameworks, such as Qiskit and even higher level companies, like Classic and others. Once you’ve integrated with openQASM, you can now enjoy the power of the ecosystem.

That’s great. Yes, I want our listeners to get a sense of how they might actually be touching this and not knowing it. [Laughter] It could already be.

Exactly. So, you might program something high level and you don’t know, but under the hood, it’s running QUA and quantum illustration.

Yes, I would say it’s a full stack, and any performance you can get on any level, that stack greatly impacts performance. So, I guess one last question to bring this home. If you could just put your futurist hat on for a moment and give me your sense of when you think error correction will start actually showing up in the industry. Maybe if you think it was one particular type of processor, that’s fine, but just like a sense of the timeline, when people might start seeing this be advertised and showing up like you log into one of the experiences and say, “Oh, no, right now, you’re getting 16 error corrected logical qubits.” When people might start seeing that kind of thing.

Wow. Okay. So, maybe I’m not the right person to ask because I’m not the most optimistic one when it comes to predicting. I would assume that playing around with several will error corrected qubits, that is something that we might see by the end of a decade. I think that that’s realistic, but really having a full fault-tolerant quantum computer that can run and with the most grandiose applications, I think that will take even longer. In the meantime, however, I’m optimistic that there is a lot of value we could extract some of these NISQ computers hopefully, and so that we continue to push the feet forward.

Yes. For all we know, we might just get better qubits before we get better error correction. [Laughter]

Yes, that’s true. I think it’s also a gradual thing. So, by the way, in some sense, even when you just take a current quantum computer and use it to simulate another quantum system, which is starting to happen now, for purposes that previously physicist would use their university cluster to do that, right? You have an HPC in your University and you use it to simulate your quantum system. So, okay, it’s not the application that we dream of that would change the world, but it does cost a lot of money to run your quantum simulation on your university HPC cluster. It means if you can do it with a quantum computer for cheaper, for less resources essentially, and have already shown some. So, I think that’s going to be the very first you will see. From there, it’s going to get us gradually progress.

Yes, that’s a great refreshing approach. It’s cheaper and I guess better for the environment too if you can run it on very, very low power comparatively for the amount of time it runs, a millisecond or whatever. [Laughter]

 Exactly.

That’s terrific. Thank you so much. I really appreciate your insights and explaining some of these core concepts to our listeners, and giving them a sense of what goes down that stack. I really appreciate it.

Thank you so much for having me. Happy to be here.

Now, it’s time for coherence, the quantum executive summary where I take a moment to highlight some of the business impacts we discussed today in case things got too nerdy at times. Let’s recap. [Music] Quantum machines works on control systems for quantum computers. Their Quantum orchestration platform can control a hybrid of classical and quantum processors to run algorithms. It should scale to over 1000 qubits. They also developed a program called QUA, spelled Q-U-A, which is a cross-platform software development gate that handles classical variables and operations and passes them to classical and quantum systems. On the quantum side, QUA allows for pulse-level type programming that supports trapped ion, topological, superconducting, and other types of QPUs. The company is also working on error correction. As the name implies, noise is present in the noisy intermediate scale quantum or NISQ era. Error correction will be required to get to a fault-tolerant quantum computer that can run deep circuits without introducing errors from compounded noise affecting each qubit. We will have fewer errors as qubit technologies improved, but we still expect to contend with some. Error correction involves encoding information across multiple physical qubits. A system of measurements then allows the detection of errors in the qubits, yielding one reliable, logical qubit. How many physical qubits will be needed for each logical qubit will vary significantly by technology type. That does it for this episode. Thanks to Yonatan Cohen for joining to discuss quantum machines. Thank you for listening. If you enjoyed the show, please subscribe to Protiviti’s The Post Quantum World and maybe leave a review to help others find us. Be sure to follow me on all socials, @KonstantHacker, that’s Konstant with the K, hacker. You’ll find links there to what we’re doing in quantum computing services at Protiviti. You can also DM me questions or suggestions for what you’d like to hear on the show. For more information on our quantum services, check out protiviti.com or follow Protiviti Tech on Twitter and LinkedIn. Until next time, be kind and stay quantum curious. [Music]