The marriage of AI and quantum computing holds big promise: the computational power of quantum computers could lead to huge breakthroughs in this next-gen tech. A team led by Stephen Clark, our Head of AI, has just helped us move towards unlocking this incredible potential.

A key ingredient in contemporary classical AI is the “transformer”, which is so important it is actually the “T” in ChatGPT. Transformers are machine learning models that do things like predict the next word in a sentence, or determine if a movie review is positive or negative. Transformers are incredibly well-suited to classical computers, taking advantage of the massive parallelism afforded by GPUs. These advantages are not necessarily present on quantum computers in the same way, so successfully implementing a transformer on quantum hardware is no easy task.

Until recently, most attempts to implement transformers on quantum computers took a sort of “copy-paste” approach – taking the math from a classical implementation and directly implementing it on quantum circuits. This “copy-paste” approach fails to account for the considerable differences between quantum and classical architectures, leading to inefficiencies. In fact, they are not really taking advantage of the ‘quantum’ paradigm at all.

This has now changed. In a new paper on the arXiv, our team introduces an explicitly quantum transformer, which they call “Quixer” (short for quantum mixer). Using quantum algorithmic primitives, the team created a transformer implementation that is specially tailored for quantum circuits, making it qubit efficient and providing the potential to offer speedups over classical implementations.

Critically, the team then applied it to a practical language modelling task (by simulating the process on a classical computer), obtaining results that are competitive with an equivalent classical baseline. This is an incredible milestone achievement in and of itself.

This paper also marks the first quantum machine learning model applied to language on a realistic rather than toy dataset. This is a truly exciting advance for anyone interested in the union of quantum computing and artificial intelligence. About a week ago when we announced that our System Model H2 has bested the quantum supremacy experiments first benchmarked by Google, we promised a summer of important advances in quantum computing. Stay tuned for more disclosures!

For a quantum computer to be useful, it must be universal, have lots of qubits, and be able to detect and correct errors. The error correction step must be done so well that in the final calculations, you only see an error in less than one in a billion (or maybe even one in a trillion) tries. Correcting errors on a quantum computer is quite tricky, and most current error correcting schemes are quite expensive for quantum computers to run.

We’ve teamed up with researchers at the University of Colorado to make error correction a little easier – bringing the era of quantum ‘fault tolerance’ closer to reality. Current approaches to error correction involve encoding the quantum information of one qubit into several entangled qubits (called a “logical” qubit). Most of the encoding schemes (called a “code”) in use today are relatively inefficient – they can only make one logical qubit out of a set of physical qubits. As we mentioned earlier, we want lots of error corrected qubits in our machines, so this is highly suboptimal – a “low encoding rate” means that you need many, many more physical qubits to realize a machine with lots of error corrected logical qubits.

Ideally, our computers will have “high-rate” codes (meaning that you get more logical qubits per physical qubit), and researchers have identified promising schemes known as “non-local qLDPC codes”. This type of code has been discussed theoretically for years, but until now had never been realized in practice. In a new paper on the arXiv, the joint team has implemented a high rate non-local qLDPC code on our H2 quantum processor, with impressive results. 

The team used the code to create 4 error protected (logical) qubits, then entangled them in a “GHZ state” with better fidelity than doing the same operation on physical qubits – meaning that the error protection code improved fidelity in a difficult entangling operation. The team chose to encode a GHZ state because it is widely used as a system-level benchmark, and its better-than-physical logical preparation marks a highly mature system.

It is worth noting that this remarkable accomplishment was achieved with a very small team, half of whom do not have specialized knowledge about the underlying physics of our processors. Our hardware and software stack are now so mature that advances can be achieved by “quantum programmers” who don’t need advanced quantum hardware knowledge, and who can run their programs on a commercial machine between commercial jobs. This places us bounds ahead of the competition in terms of accessibility and reliability.

This paper marks the first time anyone has entangled 4 logical qubits with better fidelity than the physical analog. This work is in strong synergy with our recent announcement in partnership with Microsoft, where we demonstrated logical fidelities better than physical fidelities on entangled bell pairs and demonstrated multiple rounds of error correction. These results with two different codes underscore how we are moving into the era of fault-tolerance ahead of the competition.

The code used in this paper is significantly more optimized for architectures capable of moving the qubits around, like ours. In practice, this means that we are capable of “non-local” gates and reconfigurability. A big advantage in particular is that some of the critical operations amount to a simple relabeling of the individual qubits, which is virtually error-free.

The biggest advantage, however, is in this code’s very high encoding rate. Unlike many codes in use today, this code offers a very high rate of logical qubits per physical qubit – in fact, the number of logical qubits is proportional to the number of physical qubits, which will allow our machines to scale much more quickly than more traditional codes that have a hard limit on the number of logical qubits one can get in each code block. This is yet another proof point that our machines will scale effectively and quickly.

The Artificial Intelligence (AI) systems that have recently permeated our lives have a serious problem: they are built in a way that makes them very hard - and sometimes impossible - to understand or interpret. Luckily, our team is tackling this problem, and we’ve just published a new paper that covers the issue in detail.

It turns out that the lack of explainability in machine learning (ML) models, such as ChatGPT or Claude, comes from the way that the systems are built. Their underlying architecture (a neural network) lacks coherent structure. While neural networks can be trained to effectively solve certain tasks, the way they do it is largely (or, from a practical standpoint, almost wholly) inaccessible. This absence of interpretability in modern ML is increasingly a major concern in sensitive areas where accountability is required, such as in finance and the healthcare and pharmaceutical sectors. The “interpretability problem in AI” is therefore a topic of grave worry for large swathes of the corporate and enterprise sector, regulators, lawmakers, and the general public. 

These concerns have given birth to the field of eXplainable AI, or XAI, which attempts to solve the interpretability problem through so-called ‘post-hoc’ techniques (where one takes a trained AI model and aims to give explanations for either its overall behavior or individual outputs). This approach, while still evolving, has its own issues due to the approximate nature and fundamental limitations of post-hoc techniques.  

The second approach to the interpretability problem is to employ new ML models that are, by design, inherently interpretable from the start. Such an interpretable AI model comes with explicit structure which is meaningful to us “from the outside”. Realizing this in the tech we use every day means completely redesigning how machines learn - creating a new paradigm in AI. As Sean Tull, one of the authors of the paper, stated: “In the best case, such intrinsically interpretable models would no longer even require XAI methods, serving instead as their own explanation, and one of a deeper kind.”

At Quantinuum, we’re continuing work to develop new paradigms in AI while also working to sharpen theoretical and foundational tools that allow us all to assess the interpretability of a given model. In our recent paper, we present a new theoretical framework for both defining AI models and analyzing their interpretability. With this framework, we show how advantageous it is for an AI model to have explicit and meaningful compositional structure.

The idea of composition is explored in a rigorous way using a mathematical approach called “category theory”, which is a language that describes processes and their composition. The category theory approach to interpretability can be accomplished via a graphical calculus which was also developed in part by Quantinuum scientists, and which is finding use cases in everything from gravity to quantum computing. 

A fundamental problem in the field of XAI has been that many terms have not been rigorously defined, making it difficult to study - let alone discuss - interpretability in AI. Our paper presents the first known theoretical framework for assessing the compositional interpretability of AI models. With our team’s work, we now have a precise and mathematically defined definition of interpretability that allows us to have these critical conversations.    

After developing the framework, our team used it to analyze the full spectrum of ML approaches. We started with Transformers (the “T” in ChatGPT), which are not interpretable – pointing to a serious issue in some of the world’s most widely used ML tools. This is in contrast with (sparse) linear models and decision trees, which we found are indeed inherently interpretable, as they are usually described.  

Our team was also able to make precise how other ML models were what they call 'compositionally interpretable'. These include models already studied by our own scientists including DisCo NLP models, causal models, and conceptual space models.    

Many of the models discussed in this paper are classical, but more broadly the use of category theory and string diagrams makes these tools very well suited to analyzing quantum models for machine learning. In addition to helping the broader field accurately assess the interpretability of various ML models, the seminal work in this paper will help us to develop systems that are interpretable by design. 

This work is part of our broader AI strategy, which includes using AI to improve quantum computing, using quantum computers to improve AI, and – in this case - using the tools of category theory and compositionality to help us better understand AI. 

“Computers are useless without error correction”
- Anonymous

If you stumble while walking, you can regain your balance, recover, and keep walking. The ability to function when mistakes happen is essential for daily life, and it permeates everything we do. For example, a windshield can protect a driver even when it’s cracked, and most cars can still drive on a highway if one of the tires is punctured. In fact, most commercially operated planes can still fly with only one engine. All of these things are examples of what engineers call “fault-tolerance”, which just describes a system’s ability to tolerate faults while still functioning.

When building a computer, this is obviously essential. It is a truism that errors will occur (however rarely) in all computers, and a computer that can’t operate effectively and correctly in the presence of faults (or errors) is not very useful. In fact, it will often be wrong - because errors won’t be corrected.

In a new paper from Quantinuum’s world class quantum error correction team, we have made a hugely significant step towards one of the key issues faced in quantum error correction – that of executing fault-tolerant gates with efficient codes. 

This work explores the use of “genon braiding” – a cutting-edge concept in the study of topological phases of matter, motivated by the mathematics of category theory, and both related to and inspired by our prior groundbreaking work on non-Abelian anyons. 

The native fault tolerant properties of braided toric codes have been theoretically known for some time, and in this newly published work, our team shares how they have discovered a technique based on “genon braiding” for the construction of logical gates which could be applied to “high rate” error correcting codes – meaning codes that require fewer physical qubits per logical qubit, which can have a huge impact on scaling.

Stepping along the path to fault-tolerance

In classical computing, building in fault-tolerance is relatively easy. For starters, the hardware itself is incredibly robust and native error rates are very low. Critically, one can simply copy each bit, so errors are easy to detect and correct. 

Quantum computing is, of course, much trickier with challenges that typically don’t exist in classical computing. First off, the hardware itself is incredibly delicate. Getting a quantum computer to work requires us to control the precise quantum states of single atoms. On top of that, there’s a law of physics called the no cloning theorem, which says that you can’t copy qubits. There are also other issues that arise from the properties that make quantum computing so powerful, such as measurement collapse, that must be considered.

Some very distinguished scientists and researchers have thought about quantum error correcting including Steane, Shor, Calderbank, and Kitaev [9601029.pdf (arxiv.org), 9512032.pdf (arxiv.org), arXiv:quant-ph/9707021v1 9 Jul 1997].  They realized that you can entangle groups of physical qubits, store the relevant quantum information in the entangled state (called a “logical qubit”), and, with a lot of very clever tricks, perform computations with error correction.

There are many different ways to entangle groups of physical qubits, but only some of them allow for useful error detection and correction. This special set of entangling protocols is called a “code” (note that this word is used in a different sense than most readers might think of when they hear “code” - this isn’t “Hello World”). 

A huge amount of effort today goes into “code discovery” in companies, universities, and research labs, and a great deal of that research is quite bleeding-edge. However, discovering codes is only one piece of the puzzle: once a code is discovered, one must still figure out how to compute with it. With any specific way of entangling physical qubits into a logical qubit you need to figure out how to perform gates, how to infer faults, how to correct them, and so on. It’s not easy!

Quantinuum has one of the world’s leading teams working on error correction and has broken new ground many times in recent years, often with industrial or scientific research partners. Among many firsts, we were the first to demonstrate real-time error correction (meaning a fully-fault tolerant QEC protocol). This included many milestones: repeated real-time error correction, the ability to perform quantum "loops" (repeat-until-success protocols), and real-time decoding to determine the corrections during the computation. We were also the first to perform a logical two-qubit gate on a commercial system. In one of our most recent demonstrations, in partnership with Microsoft, we supported the use of error correcting techniques to achieve the first demonstration of highly reliable logical qubits, confirming our place at the forefront of this research – and indeed confirming that Quantinuum’s H2-1 quantum computer was the first – and at present only – device in the world capable of what Microsoft characterizes as Level 2 Resilient quantum computing. 

Introducing new, exotic error correction codes

While codes like the Steane code are well-studied and effective, our team is motivated to investigate new codes with attractive qualities. For example, some codes are “high-rate”, meaning that you get more logical qubits per physical qubit (among other things), which can have a big impact on outlooks for scaling – you might ultimately need 10x fewer physical qubits to perform advanced algorithms like Shor’s. 

Implementing high-rate codes is seductive, but as we mentioned earlier we don’t always know how to compute with them. A particular difficulty with high-rate codes is that you end up sharing physical qubits between logical qubits, so addressing individual logical qubits becomes tricky. There are other difficulties that come from sharing physical qubits between logical qubits, such as performing gates between different logical qubits (scientists call this an “inter-block” gate).

One well-studied method for computing with QEC codes is known as “braiding”. The reason it is called braiding is because you move particles, or “braid” them, around each other, which manipulates logical quantum information. In our new paper, we crack open computing with exotic codes by implementing “genon” braiding. With this, we realize a paradigm for constructing logical gates which we believe could be applied to high-rate codes (i.e. inter-block gates).

What exactly “genons” are, and how they are braided, is beautiful and complex mathematics - but the implementation is surprisingly simple. Inter-block logical gates can be realized through simple relabeling and physical operations. “Relabeling”, i.e. renaming qubit 1 to qubit 2, is very easy in Quantinuum’s QCCD architecture, meaning that this approach to gates will be less noisy, faster, and have less overhead. This is all due to our architectures’ native ability to move qubits around in space, which most other architectures can’t do. 

Using this framework, our team delivered a number of proof-of-principle experiments on the H1-1 system, demonstrating all single qubit Clifford operations using genon braiding. They then performed two kinds of two-qubit logical gates equivalent to CNOTs, proving that genon braiding works in practice and is comparable to other well-researched codes such as the Steane code.

What does this all mean? This work is a great example of co-design – tailoring codes for our specific and unique hardware capabilities. This is part of a larger effort to find fault-tolerant architectures tailored to Quantinuum's hardware. Quantinuum scientist and pioneer of this work, Simon Burton, put it quite succinctly: “Braiding genons is very powerful. Applying these techniques might prove very useful for realizing high-rate codes, translating to a huge impact on how our computers will scale.”

Doing mathematical physics with diagrams instead of traditional formalism allows researchers to tackle difficult problems in an intuitive and mathematically strict way that opens the door to new insights and solutions. The new calculus we are developing that we refer to as ZX calculus, also known as Penrose Spin Calculus, has applications in fields as diverse as quantum chemistry, condensed matter physics, and loop quantum gravity.

In a recent paper on the arXiv, Quantinuum researchers Harny Wang, Razin A. Shaikh, and Boldizsár Poór have proven the “completeness” of this ZX calculus in finite dimensions, meaning that one can now use diagrams instead of linear algebra to perform calculations in finite dimensional quantum mechanics. This is a remarkable achievement.

“Now very complicated formulas in quantum chemistry and loop quantum gravity can be derived by diagrams,” said co-author Harny Wang.

Physicists have used graphical calculus for a long time. They are used widely in quantum field theory, in the form of Feynman diagrams, or in gravitational theory, in the form of Penrose diagrams. Graphical calculation strategies are generally very well appreciated as they replace a lot of difficult and tedious ‘formal’ mathematics with a simpler, more intuitive, but no less accurate diagrammatic approach.

Our researcher’s work on ZX and ZXW calculus (a near cousin to ZX) is the latest but most innovative shift from “shut up and calculate” to “depict and rewrite”, a shift that many researchers are sure to welcome.

ZX calculus was initially developed by scientists as a tool for working on problems in quantum mechanics that require intricate calculations. ZX calculus, created by Professor Bob Coecke and Dr. Ross Duncan, both of whom are senior scientists at Quantinuum, has developed over the course of 15 years, leading to a growing global community of researchers. This most recent paper marks the transition of important parts of ZX from ‘a work in progress’ to something that is fully formed.

Both ZX and ZXW calculus are known for efficiently expressing quantum relations such as entanglement. It is hoped these new formalisms may uncover connections between some of the most challenging problems in science and quantum computing.

Distinguished physicist Carlo Rovelli has already expressed interest in using ZX and ZXW graphical calculus for his work, stating “Indeed, there are concrete steps in place to translate quantum gravity problems into quantum computing problems, and I have hope that the powerful conceptual and technical tools developed by Bob [Coecke], Harny [Wang] and their collaborators could play a major role in this.”

In addition to interest from the gravity community, ZX is being adopted in the wider quantum computing community. Dr. Peter Shor recently worked with colleagues to develop an algorithm that maps Clifford encoders to graphical representations in the ZX calculus.

A key step in many quantum algorithms is setting everything up: you need all your dominoes in place before you can do much else. This is called “state preparation”, and it’s a trickier problem than it might seem. 

Our team has developed new protocols that can help – and in a big way. Specifically, the team worked on preparing “multivariate” functions, which just means functions that are used to explore problems with more than one variable, or in more than 1 dimension. One-dimensional problems do exist (think of a path that only goes forwards or backwards – we can call the variable “x”) but in the real world it’s much more common to have problems with many dimensions, or variables (think instead of a landscape where you can go forwards, backwards, left, right, up, and down – we can call the variables “x”, ”y”, and “z”).

Our new multivariate function quantum state preparation protocols don’t rely on some commonly-used and computationally expensive subroutines - instead they expand the desired multivariate function into well-known mathematical basis functions, called Fourier and Chebyshev functions. This makes our protocols simpler and more effective than previous options. 

Generally, state preparation is a hard problem, and costs exponentially many resources to prepare an arbitrary state. By expanding the functions in a Fourier or Chebyshev series, one can truncate the series to create good approximations, which instead uses only polynomially many resources – meaning that this method has better asymptotic scaling than many other non-heuristic methods (which are often designed to work in only one dimension anyways). 

Our team used their protocol to prepare a commonly used initial state on our H2 trapped-ion quantum processor, the bivariate Gaussian. Bivariate Gaussians are used everywhere from physics to finance, underscoring the practicality of these new protocols. They also analyzed examples potentially useful for quantum chemistry and partial differential equations.

A very nice feature of this work is that it is broadly applicable, generic, and entirely modular – meaning it can be plugged in to the beginning of almost any quantum algorithm, giving our customers and users even more flexibility and power.