Transcript | Quantum Machine Learning Using PennyLane

Quantum machine learning, or QML, is one of the three major application categories for quantum computing, along with optimization and simulation. As we’re working with customers to find advantageous use cases in QML, we rely daily on a tool called PennyLane from Xanadu. Find out more about this powerful free software in this episode of The Post-Quantum World.

I’m your host, Konstantinos Karagiannis. I lead Quantum Computing Services at 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.

Our guest today is the head of software at Xanadu. That’s the company behind PennyLane, which is a tool we use regularly here. Nathan Killoran, welcome to the show.

Thanks for having me — so glad to be here.

As I said, we use PennyLane in our Quantum Computing Services projects, so I thought it would be fun to bring my quantum lieutenant into this chat. She’s a senior consultant at Protiviti who puts algorithms to qubits, so to speak, so, Emily Stamm, thanks for joining me today.

Yes. Thank you for having me.

Nathan, tell us a little about Xanadu.

Yes. Xanadu is a Canadian quantum computing company. We’re about five years old now. We have software and hardware — what someone might call a full-stack play, but I don’t really like to think of it that way. I like to think that we touch all the key parts of quantum computing, but they don’t necessarily have to live all in one single monolithic stack. I like more thinking that we build on the software side, we build on the hardware side, and we build also on the education and community side.

One day, I’d love to have someone come on just to dive into the photonic approach that Xanadu is taking with hardware. That will be a whole great episode.

It’s really interesting technology.

Today, obviously, we have you here in your role as head of software, so do you want to tell us about that role?

I’ve been the head of software at Xanadu for just over four years now, leading an amazing team — trendsetters in the field. When I think of software, I think of, in quantum computing, everything between the hardware and the user. So, you might think of software as sometimes just like code, or it’s an app or something, but I think of the whole pipeline, from something that touches hardware to something that touches your keyboard or even something that touches your brain. What does the cloud platform look like that hosts that hardware? What does the local software look like that’s installed in your computer? What are the algorithms that you’re running there on your computer? How do you know what to do there? How does the software know what to do to best use the hardware?

Then, on the community side, how do people know what to run on a quantum computer? How do they know what to do when we give them the software? That is very holistic — the whole pipeline of things that we should be stewards of.

We’re going to drill into a lot of that. But first, we’re going to take a step back. Can you give a high-level of the difference between PennyLane and Strawberry Fields? I have to say I promised my legal department that I wouldn’t even hum during this interview, so I’m not going to for either of those names, but you can if you want to separate them.

You don’t want the copyright strikes.

Obviously, PennyLane and Strawberry Fields are both named after Beatles songs. It’s actually “Strawberry Fields Forever.” They came from the same single — one was the A side, one was the B side, if I remember correctly. They cover two different needs for us here. One is the Strawberry Fields software. Strawberry Fields is our first offering, and for that, we were thinking, if we’re building a photonic quantum computer, people need to be able to access it. There was no tool in 2017, when we started building that. There was no tool in 2017 for accessing cloud-deployed photonic quantum computers, and we couldn’t count on anyone else to build that on our behalf, so we had to build it. Strawberry Fields, you can think of as the API for accessing Xanadu’s photonic quantum computing hardware. Anything that we put on the cloud that’s photonic, very likely you can hit it off using Strawberry Fields.

PennyLane is much more of an agnostic software. Imagine if Xanadu wasn’t a hardware company. What if we didn’t have our own hardware op? As head of software, there’s very little that I can do to influence the hardware, so if I’m running a team and we’re going to build products, what do we build them around? What are the ideas that we use? I said, “Let’s take our best ideas — things that we think are most impactful. Let’s make software around that, and let’s make it connect to every quantum computer.”

We focused, the first days of PennyLane, on quantum machine learning. It was an area that we thought was very up-and-coming, which proved to be a very accurate prediction. We had a lot of expertise on the team, and in those days, there weren’t very many people who knew about QML, so we knew what the community might need. We could anticipate that because we were the first users ourselves. Then we said, “Let’s make sure this tool can connect to any quantum computer out there.” If we were a quantum software startup, we wouldn’t partner with just one hardware provider. We would say, “Bring your own devices. Whatever API keys you have access to, we’ll make sure that you can run the same code on all those different devices.”

Yes, and we definitely found that to be one of the most alluring aspects.

As Konstantinos mentioned, we use PennyLane here at Protiviti, and the first use case we had for PennyLane was binary classification. What types of machine learning use cases has your team worked on with PennyLane?

In the early days, we were a bit more exploratory. It was a very developing field, and it wasn’t really clear about, what are the best use cases for using a quantum computer for machine learning? We did a lot of more or less curiosity-driven research in those early days to try and get a sense of the field, and what were the key lessons we could learn?

Right now, our biggest focus is on not a particular, specific kind of use case but on the foundations of the field: What are the ideas that we think are the most impactful things that if we provide the people those tools, those algorithms, then they’re going to go out and they’re going to do research? They’re going to come out with ideas that we never could have thought of. They’re going to apply it to use cases that we never thought about it, and then it’s going to come back and benefit us and benefit the whole field because now, all these new ideas came out without us having to invest the resources into all these possible different ways to tackle something.

Our focus these days, we do do some research for sure, but on the QML side, the focus is on the foundations: How do we build up a field that didn’t exist five years ago into one that in a few short years is running at large scale with very interesting and powerful algorithms doing the interesting stuff on novel data sets? The idea there was giving up on a single use case and focusing on the foundations.

Yes, and you can see some of that foundational work and educational outreach. Both applications have a lot of tutorials, demonstrations, resources. Do you want to talk more about what you’re trying to do there in that other role, like you mentioned, with the educational aspects of PennyLane and Strawberry Fields?

The guiding principle is, it doesn’t matter how cool your product is or your software is if no one knows about it or no one knows how to use it. There’s obviously an ulterior motive to putting these things out, because it helps us to advertise our work. We try not to put any strings attached to it. We try to offer content for free and based on interest and based on what we think is most impactful for the community. We don’t force you to install it. You can view everything in your browser. The focus is on openness and putting that stuff out there, but it has a nice benefit that when people are learning about quantum computing, they’re learning about it, and PennyLane is right there, or Strawberry Fields is right there.

That’s more of the strategic side of things, but it has always been, since day one, for us a huge part of what we do. We started with a sense that community building is important. For a startup in particular, you need to build a user base. You need to build an audience. You need to build that community for your survival, and so it’s part of a survival thing: How are we going to differentiate ourselves from the big players? We do have some expertise that we think the big players could match, but we’re not focused on it, or maybe we’ve got some amazing, unique, once-in-a-lifetime team members who are leaders in the field. Let’s leverage that expertise.

My philosophy is, “Give more than you have to.” Instead of saying, “We’re just going to focus. We’re going to shut out the rest of the world. We’re going to focus on building a quantum computer and building the best software and building the best algorithms,” let’s open up, be friendly, be open to people and be willing to invite them along for the ride. It doesn’t mean we have to spend all of our time teaching people. It means that we can offer our learnings and our educational resources in scalable ways so that more and more people can discover it.

The field moves so quickly. There’s that huge appetite of content. There’s a huge appetite to understand things that, unless you’re a researcher or you’re reading the literature, you’re not going to be able to keep up with. We try to translate what we’re seeing as trends and important things in the field and make it more accessible to people as well.

Who is developing all that material on the educational side? Are they Xanadu developers? Are they independent contributors? How does that work?

We welcome contributions from everyone. As I said, a huge focus on community, and part of the aspect of community, is inviting other people to take part — not having it be too one-sided. On the PennyLane website, anyone can submit a tutorial and be listed on our community page, but there are a number of more polished tutorials and demonstrations that we host on our websites. Those more polished ones sometimes are coming from external partners, but a lot of the time, they’re coming from the internal team here.

The software developers of PennyLane or the researchers at Xanadu are the ones who are translating their knowledge. They might have just finished writing a paper, and they make it more accessible by putting a tutorial up at the same time. Or they might have had a huge new feature that they added that unlocks new powers, and they want to make sure that other people understand how to use it.

The contributions are coming from everywhere, and the more polished ones, definitely, we put a lot of effort into those.

The target audience of these is pretty much anyone? Someone really new to this, someone who’s doing it and wants to expand in their field at their company that’s trying a new use case or a new approach or something, so it’s kind of diverse, it looks like, because there’s just a lot there.

The great thing about the internet is, you can put stuff out and people can find it. We’ve discovered that there is a huge appetite for what I call the enthusiast community. These are people who for the first time are getting into quantum. They want to get up to speed. They’re smart people, but they haven’t had that background. They want to learn about it. They want to know what’s going on, and the latest trends. There’s a huge appetite from that community.

Of course, there are the more established researchers, the old guard, the people who are professors or grad students. Even for them, sometimes they appreciate having the more accessible content rather than having to dive deep into a paper and try to figure it out for themselves.

Yes, and it’s great to keep that spirit of sharing going as long as we can because eventually, I have this fear, the whole industry is going to lock down with IP concerns, and no one’s going to share anything anymore.

It could happen. The way I like to think about it is, right now, everyone is a friendly competitor, and I hope it stays that way for as long as we can.

Yes — same here.

Not to get too technical, but a differentiator of PennyLane from other quantum libraries is the use of the decorator. I was wondering: What inspired the Xanadu team to choose this approach, and is this a question you receive a lot?

Actually, it’s the first time we’ve received a question about why we use a decorator, so congrats. We definitely get questions about the interface and how it came about — the user interface — and it was a very deliberate consideration. Especially when we created PennyLane, the competitor software, the other libraries that were available to people, are very much focused on object-oriented programming, so the user creates these big, monolithic objects and manipulates them.

We decided to take a different approach. PennyLane was initially focused very heavily on quantum machine learning. In modern machine learning libraries like TensorFlow, PyTorch and so forth, the interface has evolved to a state where it’s very much based on functions rather than classes. A function is something where you can have a mathematical representation that takes an input and gives you an output. We can also code up functions in code, so you call a function to do some data processing for you or to evaluate the output given some inputs.

Functional programming is quite different than object-oriented programming. To give away a little secret here, a lot of what we use are what we call functional classes — things that are both classes and functions — but the UI that we present is very much based on functions. You have a quantum circuit. Instead of creating some quantum-circuit object and appending gates to it, you create a function that lists line by line all the gates in your circuit, and that gets called, and you could think of it as somehow being sent out to a quantum computer to evaluate.

Now, the question was about decorators. Why did we end up with decorators? Once we decided to go with the functional approach, Python has this very nice way of hiding away the complexity of manipulating functions. If you want to get deep in the details, when you have functions, you can start talking about higher-order functions, which are functions which take functions as inputs and give functions as outputs. It starts becoming a bit more of a cognitive load to carry all that around in your head.

Python provides this very nice, lightweight decorator approach which says, “Give me a function,” and then it’s just a single line to modify its function to do something. The user never has to worry about that extra burden of thinking about how it is to manipulate functions. There’s a lot of magic happening under the hood when you decorate a function. It’s basically allowing us to abstract away all the complexity and give the users what they want to do without having to think about how it’s done.

That makes a lot of sense. Have you seen any benefits using this approach?

The benefit is simplicity. It allows the user to focus purely on the quantum circuit that they are working on and not worry about how to manage the job of submitting it to hardware. We can hide that all the way with just a function call.

As you alluded to earlier, one of the great things about PennyLane is, it was obviously built not just for your hardware. There are plug-ins. You can run code on many platforms, devices, Qiskit, Braket. Do you want to talk about the plug-ins and how they work?

The philosophy here is, be as agnostic to the actual computational back end as much as possible. Obviously, there are some slight differences in quantum computing hardware. They support different gates. They might have slightly different requirements. But what we want people to do when they’re writing software is to think about the algorithm — what goal do you want to achieve? — and not think too much about the details of execution. If you can get to the point where you swap out one quantum computer for the other and you notice no changes, that’s the ideal there.

There are lots of other competitors in the hardware space. Nobody has one hardware right now, and there’s still a lot of work to do before it’s clear what the potential winners might be there, so it makes a lot of sense to allow users to bring their own device to go where the users are. If they like a particular platform, why would I stop them from using our team’s software if they happen to use hardware that we weren’t expecting? We try to make it that all of the main quantum computing platforms are supported and it’s simple to add your own.

A number of plug-ins have been built by users, and in fact, the Qiskit plug-in was initially built by a user because they wanted to run PennyLane on Qiskit devices, and we’ve since taken over adoption, but it just goes to show that sometimes you make it such that you don’t have to think too much about which hardware you’re running on. Then it allows people to very easily transfer to new hardware. It’s not baked into code anywhere.

That’s definitely one of the main advantages of PennyLane. I was wondering: What challenges, if any, have you encountered adding these external libraries?

There are maybe two challenges. The first challenge is, every line of code that you add to your code base is something that you have to test, you have to document. Having a dozen plug-ins is now something where you have to test and make sure everything is up-to-date. Sometimes, we can maybe miss a plug-in that has been updated — a breaking change in Qiskit or Microsoft’s QDK or something — and we haven’t quite noticed it until a user tells us about their use case failing.

That’s one of the challenges: As we scale, especially, our own in-house tutorials, they sometimes break and we have to update them if partners plug into Qiskit. It just becomes more complex internally, but it’s worth it. Again, we’re focusing on providing value to the users, and it’s something we know how to do, and they seem to be quite keen to use different devices, so we can continue to support that.

The other one is encouraging third parties to create plug-ins for their own hardware. There are a couple platforms where we’ve done a lot of the legwork in creating those plug-ins, and the partner would just provide us with an API key to make sure we can test it where we’ve done a lot of the work. It would be great to have more partners developing the plug-ins themselves rather than them through us. There have been a couple out there that have already done it — most notably, Amazon Braket.

Are there any plug-ins in the works you can talk about?

Not that I’m aware of. I don’t think that there’s any that are coming out anytime soon.

There will be when there’s a new machine that surprises us all, right? 

If there’s a new machine that surprises us all, and we’ve done it a couple times, drop everything you’re doing now. We can make sure that as soon as possible, we can connect to it.

Yes, start rolling up. In addition to PennyLane and using it for developing proof of concept for our customers, we have a new offering we’re working on where we’re going to be using it to demonstrate the code that we create, and then we’ll be leaving that code behind with the customer. I imagine that we’re going to be then enticing these customers. They’re going to be seeing PennyLane. They’re going to be seeing its power, and they’ll probably want to start using it in their environments.

Do you have any thoughts about deploying PennyLane in corporate managed environments? Some customers don’t like to just hand things out without having some kind of control over it.

For sure. It’s super cool, as someone who has been working on this for a few years, to see more and more companies adopting PennyLane and building products. There has been a very nice shift in the last year or so where I’m seeing more and more companies building on top of that, and it’s amazing to see. It also gives us some validation that what we’re doing is definitely popular.

The other thing is that we always are very careful not to have too many breaking changes in the code base. Some other libraries in quantum computing, because it’s such a hot field, it’s developing very quickly, the ideas are not set in stone. There might be a lot of breaking changes every release. We have a very impressive release cycle — about eight releases a year. We’re careful to try to limit breaking changes as much as possible.

It’s not always possible to stop breaking things. Sometimes the ideas you had in the past just don’t work anymore. There’s a much better way to do it, but we try to be very mindful that we’re not just doing it for ourselves. We’re releasing new versions, and the users are using them, and they should expect some sort of consistency, or at least a road map for deprecating things. That’s cool to see lots of companies adopting it, and it puts more onus on us to make sure that we’re keeping it as consistent as possible, even in a very hot space.

Sometimes you find that people, when you talk to them, are hacking the internals of PennyLane for their use cases. That’s cool to see, because it means that there’s more and more demand for doing interesting things, and we’d love to talk to partners and see if we can make those hacks more real features. In Python, in particular, there’s no notion of private versus public code. Everything is completely public, and so the way to distinguish code that is meant to be for the users and code that’s meant to be for developers is just by convention. Essentially, you put some underscore in front of it. It’s very easy for a developer, if they’ve got a use case and they can look at the code, to look at the source and say, “This is the function I need to change a little bit of.” Sometimes those are meant to be private functions where we have no guarantees about stability, and people will start using those.

It’s an interesting problem. You have to be very careful to sunset code, not make too many breaking changes, and your question was about deploying things in corporate environments. I don’t know too much about the security aspect of things and the controlling of things, but keeping things stable is also something that’s very important to customers who are building code on top of things.

Yes, and I’m sure some teams are concerned about API keys and things like that — access to back-end targets. Shots can add up really quickly when you’re running them on quantum computers, and costs can skyrocket, so I don’t know if there’s any thought given to any controls put in place for that in the future or anything like that in the road map?

Right now, everything is done different ways — where you could put your API key, for instance. You can have your configuration file where it’s read from disk. We run a lot of tests on GitHub that are pulling keys out from a secure location. The functionality is definitely there. It takes a little bit of talking to the right people and getting a sense of what their key desires are and making sure that they’re happy with what we’re building. I don’t see any blockers to be able to implement that kind of feature.

A crucial interest with quantum computing is accessibility. How does PennyLane contribute to the accessibility of quantum computing, and how can we make it more accessible?

Our goal is always to make quantum computing more accessible in a couple ways: One, give people free tools: “So, you want to get into quantum computing? You can start right now. It just requires a laptop and an internet connection. You can download the library and get started” — so, putting the tools in people’s hands.

The other thing is, giving them a sense of what to do with those tools. If I give you the tool but you have no idea how to use it, then I wouldn’t say that’s very accessible. Giving people the tools but also the instruction manual or the tutorial videos, or all the supporting pieces that allow you to not only have it in your hands but also do something with it, that really speaks to accessibility.

The third thing is, you don’t want to work in isolation. Accessibility means being able to find someone who can answer your questions or find like-minded people who you want to work on a project with. That’s where our community-building side comes from. We try to not only put out educational resources but also give people opportunities for congregating people of like minds to do get together, to work together.

One of the ways we’ve done it is, we just had this event, QHack. That’s our big developer conference for PennyLane. It’s very open. It’s very welcoming. It’s a great way to get introduced to the field and to meet others. We had about 65 teams that are working on hackathon projects, and some of these teams were not necessarily people all living in the same city or going to the same university, but they found each other through the PennyLane community or the quantum computing community.

Yes, and I attended QHack. That was a great conference.

Glad to hear it.

Yes. It was terrific. I do think PennyLane makes quantum computing very accessible, and definitely the fact that it’s free also lowers that barrier which a lot of people face getting into quantum computing. I’m glad that accessibility is something that Xanadu is thinking about.

It’s one of our biggest priorities.

It’s a free product, but we’re still using it even though we have access to a lot of others. It’s that obvious compliment: When something works, why wouldn’t you want to use it?

I’m curious. Maybe this is my opportunity to turn the tables on you. What do you guys use PennyLane for?

Mostly different quantum machine learning use cases, especially in finance. We’ve been working on one using the quantum support vector machine, another using Variational Quantum Classifier. Those are both for that binary classification use case I mentioned earlier, but also exploring it for a lot of other quantum machine learning use cases as well.

Yes, and that’s why I wanted to ask you about your recent paper, Is Quantum Advantage the Right Goal for Quantum Machine Learning? What are some of the more meaningful questions to ask, and why is this approach beneficial?

This is a recent paper by Maria Schuld and me. Even the title went through a few iterations, and we ended up with Is Quantum Advantage the Right Goal for Quantum Machine Learning? because we really wanted to force the question. The way we see it, it’s amazing that we have a big ecosystem in quantum computing and all ideas are on the table. We don’t want to have people gatekeeping or people saying, “No, don’t look into that.” At this early stage of the technology, it’s important to keep all ideas on the table and let them simmer and let people with passion push ideas that we weren’t expecting to end up as successful. I never want to tell people what you should work on and what you shouldn’t work on. But, of course, we did write a paper with a very provocative title, so, obviously, we have our own thoughts about this.

The way we see it is that there’s a bit of overemphasis on trying to prove quantum advantage for quantum machine learning. It’s important, but maybe there are too many resources being devoted toward that compared to other areas where we could potentially be putting our time and energy, so let’s see.

Quantum computing and quantum machine learning, they’re like the classic moonshot. This is a technology that you anticipate is going to change the world — it’s a huge undertaking technologically to get there, scientific challenges to overcome. If you’re going to the Moon, you don’t look up the Moon and say, “I’ll just jump higher and higher and higher,” because you’re not going to get there. That’s how I see quantum advantage. Some of the papers that we see, some of the works we see, are people looking at the Moon and saying, “Well, I know how to jump, so I’ll just try to see if I can push jumping as high as I can get.” It’s like saying, “I know how to do something today with a classical computer. I’m going to try it with quantum and try to see if I could prove advantage there.”

You don’t get to the Moon by jumping higher and higher. You have to at some point learn about the physical world and propulsion and learn to build rockets, and engineering, and train your astronauts. There are all these important, interesting steps along the way that you’ll need to put the energy into if you want to get to that final stage.

We’re proposing that people to put more and more energy into those foundational steps, those key building blocks, those key tools that we’re going to leverage to get to that goal of quantum advantage. We all believe that they will be able to show some quantum advantage for quantum machine learning. It’s already known for general quantum computing how to show quantum advantage. There are some interesting papers the last couple of years about the QML case. The consensus is that that’s not a mirage, so if that’s not a mirage, then let’s get started on building those rocket ships so we can get there. That’s a very storylike way to summarize the position.

Do you think that there’s going to be some developments that we haven’t thought of that will get us there? Machines might become really powerful, and then code might be really more efficient, but is there something else, some other synergy we’re missing still?

There are definitely going to be some unknown unknowns — things that we don’t know we don’t know that we’ll have to solve when we get there. We don’t know where they’re going to be, and we don’t know what the solution is, but at some point, we’ll be faced with those questions, for sure.

I also think that there’s probably stuff that’s known today, and a few people are probably recognizing, or maybe they’re fortunate to have been in the right area at the right time. They’re recognizing that this is the way it’s going to have to go: We’re going to have to solve this problem. We’re going to have to push this idea in order to move the field forward.

There are going to be questions like that along the hallway. There are some that are here today. We might even have answers now, but it’s just not clear that they are the right answers, because we haven’t probed them enough. That’s a lot of what our focus is on at Xanadu at the research side — trying to identify those key questions and try to solve them, and use that to move the whole field together.

With those optimistic words, this is the decade of quantum, so you have until 2030 to figure it all out.

Yes. No pressure.

Yes, no pressure. Thanks a lot.

As I said, it’s a community effort, so if we get there, it’s because of the community.

Yes, I agree. Nathan, thank you so much for joining us today. I really enjoyed this. Thank you.

My pleasure. Thanks for having me.

Thank you to you, Emily, for joining me as a cohost today.

You’re welcome.

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.

Xanadu is truly a full-stack quantum computing company. They’re working on a photonic quantum system and focusing on all the software layers involved in accessing and programming their machine, as well as systems from other manufacturers. In addition to creating the PennyLane tool, Xanadu is focused on making learning about quantum coding more accessible. They have a community of users providing tutorials, in addition to staff helping create content.

PennyLane is a powerful tool for coders interested in quantum machine learning. Users can select from different back-end devices and platforms. For example, you can run code on Amazon Braket or Qiskit. As the library of plug-ins grows, the community of users helps provide real-time feedback to detect any issues that may appear — an excellent model for an excellent free piece of software that we use daily at Protiviti.

Xanadu is thinking about what it takes to take QML to the next level. There are still unknowns in the past to making this field advantageous to multiple use cases. I highly recommend you check out Nathan’s coauthored paper Is Quantum Advantage the Right Goal for Quantum Machine Learning?, which I’ve linked in the show notes.

That does it for this episode. Thanks to Nathan Killoran for joining to discuss PennyLane and Xanadu, and to Emily Stamm for being a special cohost. Thank you for listening.

If you enjoyed the show, please subscribe to Protiviti’s The Post-Quantum World and leave a review to help others find us. Be sure to follow me on Twitter and Instagram at KonstantHacker. You’ll find links there to what we’re doing in Quantum Computing Services at Protiviti. You can also DM me suggestions or questions for what you’d like to hear on the show. For more information on our quantum services, check out Protiviti.com, or follow ProtivitiTech on Twitter and LinkedIn.

Until next time, be kind, and stay quantum curious.