At some point in time between the rapid jump in facial recognition abilities of my iPhone and the Google DeepMind defeat of world champion Lee Sedol in the ancient game of Go, I began paying attention to developments in artificial intelligence.

Both of these achievements occurred years ahead of predictions by computer scientists who are familiar with the extraordinary challenges posed by machine learning.

Pictures of faces are packed with a huge amount of complex information – information that is changing over time, and involves different lighting conditions, image quality, and camera angles. The extraction of a simple quantitative feature (e.g., my name) from a database of photos seems like a Herculean task.

The game of Go is another complex challenge. Each move presents a vast number of strategic advantages and disadvantages. It was long assumed that such decisions are best made by an intuitive human, evolved by nature to be adept at the abstract arts of pattern matching, imperfect prediction, and hunches.

But as Sedol discovered when he was beaten 4-1 by the AlphaGo artificial intelligence (AI), a new breed of computer algorithms is beginning to dominate these and similar tasks.

These algorithms can be loosely categorized as “machine learning,” as opposed to “machine being explicitly told what to do by a programmer.” Instead of hard-coding which “features” in a photo are important, programmers have started to teach machines how to work that out themselves. Much like the old adage “if you teach a man to fish…,” if you teach a computer how to auto-encode features in a photograph, it will feed you images of fish for life.

Paradigm Shift

This is a massive shift in paradigm, and it could prove fruitful for theoretical physics in more ways than one.

I am a researcher in computational quantum many-body physics, studying matter, materials, and artificial quantum systems. Complex problems are our milieu. Indeed, condensed matter physics contains the most complex object in all of nature: the quantum wavefunction of a many-particle system.

If I wanted to use a computer to mathematically represent the electron wavefunction for a minuscule, nanometre-sized chunk of dust, it would require a hard drive containing more magnetic bits than there are atoms in the universe. To get around this, physicists have a grab-bag of tricks that allow us to extract the useful properties of some wavefunctions, using only the modest computer hardware currently available.

Yet many other important problems in quantum physics are veiled behind a dark cloak of infinite complexity. I began to wonder: what if machine learning could be harnessed to make a dent in this complexity? Overcoming the impasse in even a small area of research could produce unknown breakthroughs.

In late 2015, I excitedly delved into the exploding field of machine learning research. In this world of unfamiliar conferences, codes, and academic literature, I found a field complete with its own quirks, fashions, and trends – yet solidly grounded in the real-world success of what it is creating.

My first surprise should not have been one at all: physicists have long been exploring crossovers between machine learning and quantum mechanics.

However, unlike my desire to use existing machine learning algorithms and computer hardware to tackle difficult quantum physics problems, many early adopters from physics set their eyes on a different prize: translating machine learning algorithms for use on a quantum computer.

What better way to precipitate rapid advances in all facets of AI than to marry a disruptive software technology (machine learning) to a potentially disruptive hardware technology (quantum computers)? This is an appealing idea – but there are substantial hurdles in the way.

Quantum hardware offers speed-ups only for very specialized tasks. There are no concrete proposals for, say, a quantum neural network or a quantum deep learner – only bits and pieces of various algorithmic ideas.

And even if these building blocks could lead to a real-world improvement in machine learning speed or quality, the quantum hardware simply does not exist to make it a reality. We have not even fully agreed on a construction material for the qubits that could be used to produce quantum computers at an industrial scale. Decades of scientific and engineering dead ends, innovations, and breakthroughs are destined to intervene between now and development of an artificial quantum brain.

Nonetheless, today’s research landscape offers remarkable opportunities. There’s no denying that machine learning technology works; proof of that is only a smartphone away. But can quantum physicists harness this technology, and maybe even contribute to it?

The Birth of Quantum Machine Learning

That was the driving question behind the “Quantum Machine Learning” conference at Perimeter in August, 2016, organized by myself and some friends who also wanted to explore the potential of machine learning in quantum research.

It turns out we were far from alone. The depth of interest and breadth of perspective from both academia and industry surprised us. Close to 100 people came, from universities; information industry giants like Microsoft, Google, and Intel; Silicon Valley startups; and government research interests from the United States. It was a heady week spent comparing research, exploring common ground, and defining what we hope will become a richly rewarding field.

Perhaps quantum machine learning could apply face-recognition protocols to quantum physics. One of the many tantalizing examples proposed by students and postdocs was to have state-of-the-art neural networks (running on conventional hardware) identify the subtle signatures of a phase of matter, using the imperfect snapshots of the wavefunction.

Since we know how to simulate these fleeting snapshots on our banks of massive supercomputers, we can program open-source machine learning code from Google and others, and use it to identify the “face” of a superconductor, or an insulator, or even an exotic topological phase of matter. In essence, we could create “Phasebook” software.

Once in place, we could then reverse the software in order to generate unique, machine-learned “mugshots” of new materials or phases. We could even use such a technology to analyze the wavefunctions of artificial quantum devices, simulated on our (classical) supercomputers as prototype quantum computers.

While most early adopters in the physics community saw machine learning as something to do with a quantum computer, it could first help us actually design a quantum computer.

We live in a remarkable time. Some physicists hope that the current revolution in AI research may unearth a rare algorithmic gem that could serve useful to quantum research. Even a small improvement in the understanding of a particular wavefunction could crack a problem wide open. In condensed matter physics, the stakes are as high as the design of a room-temperature superconductor or a new, exotic replacement for silicon.

But regardless of whether or not we make such an algorithmic discovery, this path of enquiry is rich with potential.

As physicists, computer scientists, and AI researchers, we’re still looking for what unifies us. But by the end of the conference, it felt like we were forging the beginning of a community. By learning each other’s language, and listening to ideas tangential to our own areas, we are creating space for remarkable crossovers, and laying groundwork for game-changing discoveries.

Perhaps the most exciting thing about the new field is that it will be driven by people much younger than me, who – unencumbered by the current, largely artificial barriers between the old fields – will forge a new path where quantum machine learning will flourish.

Roger Melko is the Canada Research Chair (Tier 2) in Computational Quantum Many-Body Physics. He is an Associate Professor at the University of Waterloo and an Associate Faculty member at Perimeter Institute.

Watch all the talks from “Quantum Machine Learning” at pirsa.org/C16017.