CQC2T: Centre for Quantum Computation & Communication Technology

Transcript

My name is Simon Devitt. I'm a founder of a quantum consultancy company called HBAR. I'm also the Chief Science Officer of a new startup called Turing, and I'm delighted to be asked to chair this panel on the quantum application stack and how it might influence development within the private sector and revolutionise the world.

First of all, we have representatives from both the startup community, more established corporate representatives, and from the academic community. I thought initially I'd just run down the panel and let everyone do a minute or so introducing themselves and what their primary focus is. So, I suppose we'll start with Martin down the end from Microsoft.

Thanks Simon. I'm Martin Roetteler with Microsoft Research. Let me just begin by saying it's a very exciting time to be working in quantum computing. You see a lot of emerging devices. We've just heard about the Google effort, there are several other companies and outfits that try to build a quantum computer—some of them allow you to actually play with it. I personally work on the algorithm side. I try to find problems where we can have a tremendous speed-up over what you can do classically. Of course, historically, the first such problem was factoring, which has implications in security and privacy, but we're going to move beyond that and try to find problems that are more of an immediate business value.

You heard in Alan's talk about chemistry. We have a big chemistry effort going on at Redmond. We have Matthias Troyer, who just recently joined our group—he's one of the leading classical experts in the topic. We have material science applications. We are looking into machine learning applications; that's kind of a new area and the applications are not as clearly defined. Personally, I'm interested in costing algorithms. At this point, everything is unclear—what's the final gate set, what we can actually support—but you can still, even at this point, make some statements about how big the machine needs to be roughly to run some applications. I'm interested in that, trying to build the software that allows us to do it, and we kind of cost out the number of gates, the number of qubits. That's kind of part of the Microsoft approach towards quantum computing.

Great, thanks Martin. Matt from QCWare.

Sure. My name is Matt Johnson. I'm the CEO of QCWare. We are a quantum computing software company based in Palo Alto, California. The main points to think about with us are, first of all, we're focused on applications, and the application sets we care about are quantum machine learning, finance, and search. Search for us, in the broadest sense of the word, means finding a needle in the haystack for graph-based or network-based problems.

Another characteristic of our firm is that we're specifically focused on enterprise problems and making sure that the software we're developing is reachable and usable for novice enterprise users. So the composition of our team reflects that. We're computer scientists, quantum algorithm people, and physicists, so we work together as a team to build the technology stack that sits on top of the bare metal or the API that the hardware vendor will expose up through the application side.

Great, thanks Matt. Michael from QBranch.

Hi, good afternoon. My name is Michael Brett. I'm the CEO of a company called QBranch. We're headquartered in Washington DC and have a development team based in Australia. We're a data analytics company; the bulk of our business is focused on applying machine learning and data science to challenging enterprise problems, but we see quantum computing as a long-term strategic differentiator that we want to be involved in. We've been investing some of our equity and some of our partnerships into understanding the quantum computing landscape, the sort of applications that might be enabled by quantum computing that would have an impact on our existing business, and hopefully those two business areas over time will merge together and we'll introduce our classical analytics customers to quantum, and our quantum customers to the analytics work that we do on the classical side. It's a pleasure to be here, thanks for the invite.

Great, thanks Michael. And Gabriele from Volkswagen.

Hi everybody, I'm Gabriele Compostella from the Volkswagen Data Lab here in Munich. Volkswagen meaning one of those companies that we heard is going to be disrupted by new technologies. The Data Lab here is basically a centre of excellence for everything related to data analytics, advanced analytics within the company Volkswagen. That means we work with all the group brands that are owned under the umbrella companies of the Volkswagen Group. Our idea is to bring together business experts, data analysts, technology experts and try to build data-driven innovation for the company.

As part of this effort, we focus on different areas like big data analytics, machine learning, deep learning, but we also have some research topics that go into the direction, for example, of natural language processing and chatbots. One of the research topics we are working on is also quantum computing, to try to prepare us for the new revolutions coming and the new products that we will be able to create with that. So that's why I'm here, and thanks for the invitation.

Great, thank you. And finally, Michael Bremner from UTS in Sydney and the CQC2T now.

That's right. As you've just said, I work for two research centres, both based in Australia. I work at the University of Technology Sydney's Centre for Quantum Software and Information, which is a research centre focused on the theoretical aspects of quantum computing. That goes everywhere from blue sky research into the mathematics underlying quantum algorithms, all the way through to working with teams of experimentalists on developing benchmarking and applications for their devices. I do that in part with the Centre for Quantum Computation and Communication Technology, which is an Australian government-funded research centre.

Great. So what I thought I'd do is try to tailor off from what Alan did, but try and give it a little more detail.

First of all, I'd like to get Mick, and maybe Martin might want to jump in on this, to sort of give us an outline of the quantum algorithm and the quantum application landscape. Because we've heard from Alan, there's two very important regimes here. There are things that are non-error-corrected quantum algorithms and what can be done there, all the way up until fully quantum error-corrected universal algorithms, which, as was mentioned, is going to take some time.

So Mick, if you want to talk about this area a little bit, and then Martin might want to jump in on some of the work that Microsoft's done to try and find a few qubit applications and try to get around not requiring error correction.

Okay. So that's going to give me a whole history of quantum computing in 30 seconds, but alright. I guess it's important to understand that quantum algorithm development has been going on now since the early 90s, and to begin with it was really a mathematical topic. It was about studying these mathematical structures, and it was done very hypothetically—imagine if you had this thing called a quantum computer, which maybe one day someone develops. People focused on developing algorithms for universal quantum computers that work, ideally with error correction and everything like that.

In recent years, this has really shifted, and I think Microsoft's played a big role in this, as has Google and a few others too. As hardware has got much better, to the point where we're now actually talking about building really interesting devices, we've focused a bit more on near-term applications, because now we can actually see the thing we need to develop for—we can talk about what is the fidelity, what is the architecture, what's it going to look like. We're now developing applications much more specifically suited to these sorts of tasks.

Now, as was alluded to by Alan, we're approaching this regime which people call quantum computational supremacy—which can be a vaguely defined term—which means we're heading towards building devices which, in principle, could have the capability to, for very specific applications, outperform every classical computer. To begin with, these applications are very academic, and the way they've been developed has been to find some mathematical leverage which lies down in the deep structure of some problem, and then build that up to talk about an actual application or a benchmarking task that a device can do, which is quite different from the traditional way of doing development for enterprise and industry.

Yeah, I very much agree with those points. At Microsoft, we're of course observing very closely the developments around quantum computational supremacy, which means you try to separate the classical world from the quantum world and find a problem that you cannot solve on a classical machine, period. The trouble with it, in a sense, is that none of these problems seem really useful, whereas we, of course, want to set our sights to problems that we care about—the hardest problems people care about—and we cannot solve them in any reasonable amount of time. Even if we covered the entire planet with computers, we could not solve them.

One area where we found such problems is in computational chemistry. Understanding chemical reactions is very important. We heard in Alan's keynote about the problem of nitrogen fixation, for instance. It's done currently—harvesting nitrogen from the air is done currently with a very efficient process called the Haber-Bosch process, which happens under high pressures, high temperatures, and you need catalysts for this, and it has a relatively low yield. A quantum computer could help us—it has been shown on paper that it could—help us innovate there and find better catalysts that help us have higher yield for the process.

Of course, now you have to ask several questions. First of all, how large does a quantum computer have to be to actually implement this? How many qubits? How many operations will you have to do? Can you really solve the optimisation problem of finding a catalyst and then implementing it? All these things—we have done some baby steps towards that. We costed out initially that you will need around 100 logical qubits to understand the process of nitrogen fixation the way nature does it, using a molecule called nitrogenase. Nobody really understands how nature does it, but having a quantum computer, we could at least have a chance of understanding it.

It's maybe interesting for laypersons to know—one of the key applications is just simulating physics. We try to build this machine, this incredibly contrived machine, and one of the key applications is just simulating other physical systems. That's one of the killer applications, presumably, is just to do that. If we had such a machine, we could apply it to problems in material science—maybe find better ways to construct superconducting materials that would superconduct at higher temperatures; that would be a really, really big deal. We could hopefully have the catalysts I mentioned. In machine learning, we could hopefully use a quantum computer to train models in a better way. There are early indications that that might be possible. When you look at deep networks, you can model them by Boltzmann machines, and it's known that a quantum computer can help train Boltzmann machines in a better way. "Better" is kind of tricky to define in that context—it means specifically that a method that is classically used, based on stochastic gradients, can be improved on by having a better quality gradient. If you have a quantum computer, you can point in the right direction, make a step in the gradient.

So, a few very domain-specific applications, that's kind of... But I think one of the key words that you said in there is when quantum chemistry—you're talking about 100 logical qubits, error correction is still going to be required at this level. So, if I can talk to Gabriele about certainly the interest of these bigger companies such as Volkswagen, where do they come in on this in regards to what they're looking for? Are they hoping that there are going to be applications out there at this sort of 50 qubit to 100 qubit level—not logical qubit, but qubit level?

Well, I think we are taking the proactive approach that Alan mentioned before. We are trying to see what's already out there and how we could use it for realistic problems. We had an experience just this year, a few months ago—we were able to try our first, let's say, quantum application. I don't know if I qualify as a quantum developer, so I didn't have the guts to stand up before, but basically that's what we did. We had a partnership together with our colleagues at D-Wave, and we tried to test their system—not at 100 qubit scale, but at 1,000 to 2,000. What we did was, with a small team locked in a room for two and a half months without pizza, we tried to build something useful out of it.

We tried to ask ourselves, "OK, so we have a real quantum computer here that is available—this is top of the technology that we have today—so can we use it to solve a real problem that matters for our industry?" Something along the lines that was discussed before. The thing that we did was we created a small application that ran on the D-Wave and that could optimise the traffic flow—so optimal routing for many cars on the street. We had some very interesting results that were presented at CeBIT this year. We are still working on that, but the outcome of this exercise was that, yes, there is something that we can already do. There is something that already works with the current quantum technology.

Of course, the limit is the scale at which we can push this problem. We are now able to perform this optimisation in a few seconds—so from five to fifteen seconds—for a few hundred cars on the street, and we can select between multiple possible routes the optimal route for each and every car based on the behaviour of all the other cars. So, this is, for example, one of the possibilities of some new applications that are not common in our industry yet, that could be differentiators for our company. We are still, of course, exploring other possible applications that are related to optimisation in many other aspects of the things that we do.

But, I mean, in the case of how Volkswagen look at this problem itself, do you reach out to the quantum companies and say, "Hey, we're interested, what do you suggest?" Or do you have your own in-house teams who have some understanding or some expertise to then go, "Well, this might work for us," and so we'll reach out?

Of course, we are not able to do this all ourselves. We have worked in close partnership with D-Wave to perform this exercise. We are also gaining some education ourselves. I'm a physicist by background, so I sort of understand some of the technical problems—not all, of course—but my colleagues too are trying to learn more. Let's say we are going through two different paths: one is to be proactive, again, with respect to the evolution of the technology; the other one is trying to build within the company already the competencies that will be needed once quantum computing becomes really commercially available and more widespread than it is right now.

Yeah, that's something that I've noticed quite a bit, that a lot of these companies are looking to develop competency within this area. So Matt and Michael, feel free to jump in on each other. How do your companies work in this software space—at these non-error-corrected levels—to try and consult with people or develop these applications with various stakeholders? How do you guys tackle that? You both work with financial sectors a little bit as well. How do you approach that? Do they come to you or do you go to them?

Yeah, so building on the experience from Volkswagen, that is the style of project that we typically get engaged in, where a company sees quantum computing on the horizon that's coming up towards them. They're picking up signals, like Google are putting out today, that hardware will soon be ready in a tractable timeframe to start to think about this technology. So organisations are seeking to understand what the impact is going to be on their business. One of the troubles with quantum computing is that the specific application in that business is non-intuitive and non-obvious in an initial meeting. You need to spend some time with that company to understand their particular processes, where they're spending a lot of computational power already, or what are the problems that they're avoiding because it's too computationally intensive, and how quantum computing might be able to have some impact there.

I think it's important to consider the importance of emulators of quantum computers. These are systems that run on classical machines but emulate the behaviour of a quantum computer, at least to some level of fidelity and to some number of qubits. That provides an inexpensive, accessible way for an enterprise to start to develop those capabilities and skills and build the awareness of quantum computing and demystify it in a practical way. Getting hands-on with an emulator is something that can be done ahead of the hardware development, and then once you've got applications in mind, test those on the hardware as and when it becomes available.

I agree with what Michael said. I would add a couple of points, and I think this applies to any member of the quantum computing software ecosystem. The most important thing to do is to be working very closely with enterprises, regardless of the sector they're in, and there's this amount of creativity that's even more important than technical knowledge of the hardware topology or any other aspect of quantum computing. It's that ability to map the problem the enterprise has and to have a quantum representation of that to enable it to be run on a quantum processor. That sounds like a very technical thing to do—and it is—but it actually requires an equal measure of creativity.

So, when we tactically think as a company where we want to focus our time, clearly there's a better scientific argument to say let's focus on quantum chemistry, quantum materials, quantum physics, because those datasets are already in a quantum state and they can run kind of natively on a quantum processor. But we're actually taking a bet—a very clear bet—that the hardware vendors, the ones that are in this room (so that's IBM, Google, D-Wave, Rigetti), we're taking a bet that the engineers there are going to be able to pull forward usable, early-generation quantum processors in a fairly short period of time.

So, we are not just doing that kind of application set, but finance—working with D. E. Shaw, one of our investors—working with Airbus on engineering design and test problems. Those are areas that may be somewhat under-addressed right now, but we're taking the bet that quantum processors are going to be more powerful sooner than the conventional wisdom suggests they are, and so that's really where we want to put our footprint down.

So, a few-qubit quantum algorithms that have a marketable impact—something that is demonstrably better than what we can do classically, has an application that can be marketed or commercialised—this has been the holy grail of the academic community to a certain extent for the last 20 years and still really doesn't exist yet. So, Michael, you talked about emulators. Emulators only take you so far—arguably, you can do a lot of it on pen and paper. Once you get to 20, 30 qubits, and then when you're pushing past 30 qubits, you find your classical computer can't keep up. So, within what QCWare and QBranch are doing, how much of it is app development at this sort of non-error-corrected level?

We're going to have a really tough time as a community if we can't get quantum programming into the hands of classical computer scientists and software engineers, and to make it normal for a graduate from computer science from Stanford who's never thought about a quantum computer before to pick up a library and start to use it and think about it as a classical problem. That means building out the libraries and the toolkits and the emulators that exist within a normal development environment that can run on-premises at their location, whether it's at a financial institution or university or wherever, to get access to the datasets that they have.

So, for our company at least, the way we think about our role here is to do the top-down work—to look at the applications that might be relevant and to drill downwards in building out those libraries and compilers and toolkits that go into the hands of regular developers. The hardware teams and the people close to them do the bottom-up—build the bare metal, the control systems, and the abstraction layers on top of that. Building out that workforce that has some understanding of quantum computing and can use those libraries—that is one of the enablers to making sure we can find the applications down the track.

Just really quickly to add to that, you brought up a very interesting point. You talked about, for circuit model machines at least—let's talk about adiabatic or annealing processors in a second—you talk about the lack of development, so to speak. I mean, it's a hard thing to do, as Michael said. Certainly, there are six or seven quantum algorithms that have provable end-to-end speed-up potential. The issue is those algorithms were developed in the scientific and research community and they really weren't expanded or put to the test with real datasets. I think the interesting challenge will be to take these algorithmic primitives that are out there and to mash them up against enterprise problems, and to force the algorithm community—the practical algorithm community—to try to shoehorn those real problems onto those algorithmic primitives. Those algorithms that do have provable speed-up unfortunately do require extensive error correction to implement.

Yeah, I was just going to jump in and say, it's really important to understand that quantum computing speed-up is not just a matter of something going faster. It's really hard to map problems into a subroutine which will deliver you a speed-up, and it really deeply depends on, in some ways, the architecture—the way in which, especially when we're talking about near-term devices, not everything's going to work perfectly. Sometimes those imperfections can swallow your speed-up completely, or it turns out that a classical computer can do just as good a job.

This is a thing which I think everyone on the stage here is kind of focused on trying to get to the bottom of from different perspectives. I think it's interesting that people are looking at enterprise development and how do we map those problems onto the existing subroutines. From an academic perspective, there's really a lot of work on finding those points of differentiation between classical and quantum computers and trying to leverage that up. I think that's a bit of a summary of what I think is going on.

It is the interesting approach of Microsoft—and Martin, you're going to have to say it real quick because we're getting short on time—in this method that people might have heard of called topological quantum computing. There's a bit of a distinction: there are topological quantum error correction codes which work on qubits, but what Microsoft's approach is, is these systems with intrinsic topological protection. It's a bit of a gamble, because it would, in one sense, mitigate the error correction requirements, but it's also based on particles that don't really exist yet.

Right. Okay, let me try to say that in a few remaining minutes. The problem of decoherence has already been mentioned. That's a very tough problem that the quantum computer has to deal with. Intuitively, it's kind of like a negative interest rate—like if you put money in the bank and after a year you have, instead of $100, you have $99, and then it keeps compounding the negative interest rate. You can see that quickly kills the computer. If you have a number of steps you want to do, it just drops exponentially the amount of signal you get from the machine, and it makes it hard to scale up a quantum computer. It's as simple as that.

There are several ways out of that. We could just go bare and work with the bare qubits and hope that our circuits never have to be very deep, and that might work with the great proposals on quantum variational eigensolvers that actually might give you good answers. The other approach is to add error correction, which comes with overhead. Classically, you have redundancy—triple redundancy, stuff like that. We have that in the quantum world too, but our factors are larger. They're not three, they're maybe seven in the most optimistic case, maybe ten to the four in a more pessimistic case.

So Microsoft has set its sight on a particular approach where we try to build a sort of self-correcting memory. Classically, you know, like hard disks, for instance—they're pretty resilient, right? If one of the magnetic spins flips, it flips back because the neighbouring spins just force it back. There are similar things in topological insulators—structures that have topological materials. In fractional quantum Hall effects, they have been observed. They show extreme resilience against actions that the environment might do to your system. So in a sense, it's very protected against the environment.

On the flip side, it actually doesn't want you to talk to the system anymore. So the only thing you can really do is some topological moves—you can braid it around each other, and hopefully by doing so, you can do operations. To answer the question, we have set our sight on that particular approach to build a topological quantum computer. The experiments running right now in Copenhagen and Delft—the two experimental groups that are now part of Microsoft—they try hard to build a small computer at this point.

Once we go beyond the two-qubit gate, which will be a crucial step in that roadmap, there seems to be no limit from a physics point of view. There's nothing that, on the theory side, prevents us from building a large-scale computer that needs only very little error correction. When we scale up that machine, we could reach a practical quantum computer quite quickly, and we have to worry about actually compiling into that architecture. So that's kind of our approach. It's definitely an engineering challenge. It's an interesting bet from Microsoft, that's for sure.

So I suppose that's all we really have time for on this. I've got a couple of minutes for questions if people want to throw their hands up before somebody cuts me off.

I've heard it said a number of times that the basis for Microsoft's quantum computers—these quasi-particles—haven't been proven to exist, and yet I've also heard it said that there is no explanation other than these particles exist for why these machines are doing what's happening right now. So I was wondering if we can officially get to the resolution of this question right here today. Is it true that the Microsoft machine essentially utilises an effect that there is no plausible explanation other than you have these quasi-particles?

It seems like it. It was first observed in 2012 by Leo Kouwenhoven in Delft. Back then there was a big debate and there were several papers written that had stuff like "smoking gun evidence: yes or no" in the title. I think by now there's so many signatures of these so-called Majorana zero modes that have been observed that people kind of settle on, yes, this is a zero mode, this can be used to encode a qubit. The questions are now of a different nature. The questions are like, can we actually do it in a meaningful way so that we can operate on it? Can we actually couple several qubits together? That's kind of where we move to at this stage, and there is a white paper out there on the arXiv that lays out a particular architecture, how we can scale, but on the engineering side there's several really tough problems that we have to solve. But I think it's pretty much settled that we have observed the Majorana zero modes, yes.

I should do what Alan did and get all the condensed matter theorists and experimentalists in the room to stand up, because there's no condensed matter theorists or experimentalists on this panel. Take it with some rock-sized grain of salt.

Any other questions?

Yes, you spoke about this idea of classical emulators of quantum computers. Could you tell what is the state of the art—how many qubits can be emulated presently with some computers?

Thanks. It depends on the computation. It depends on the depth, it depends on the layout and the actual specific thing that you're trying to do. For the quantum supremacy stuff, which Alan alluded to earlier, which I've worked on with Google, I know that the best classical simulators can, specifically for those classes of problems, get to around 45-ish qubits. There are real problems in going much further than that, because of RAM limitations and other things. But for different classes of problems, you can go a lot further—you can simulate a lot more qubits, but...

Wasn't there the Nature paper from IBM about factoring an arbitrarily large number with two qubits?

Yeah. A quick advertisement for QBranch: we have developed an emulator that looks at a particular hardware system that has been deployed on premises at the Commonwealth Bank of Australia. The purpose for that is to get the emulator into the hands of their developers inside the bank, so that they can start to broaden their awareness and understanding of how quantum computers work and what the applications might be. That particular emulator does at least 30 qubits, but again, it's very dependent on the problem, and we'll continue to push that state of the art as we go forward.

Okay, I've got Jason standing up. Sorry, can I have one more? Very quick one.

Just a quick question for Matt, because you brought up that you guys are looking at potential financial applications. Is that just because D. E. Shaw is an investor, or what do you guys see as—are the big algorithmic capital allocators investing in this, or...?

Yeah. The big reason—and we can say this on a comparative basis—algorithmic traders, principal traders, have the ability to be really, really supple in terms of kind of being a solution looking for a problem, whereas in most cases you need to identify a problem and find a solution to it. So again, there's this creativity thing that quantitative traders look for to try to get an edge. You can look at problems, both that would be run efficiently on annealers and circuit model machines, that could be translated into trading strategies.

I do actually think that there's an argument to be made—forget about the underlying maths and whether you can mathematically argue that there's a particular advantage to a particular industry for quantum computing—but I think that the psychology of principal traders and their creativity and flexibility makes them an obvious place to explore implementing this technology, yeah.

Thanks. Well, thanks again, and please, a round of applause for all our panellists.