A Glimpse Into The Future of Quantum Machine Learning
By Robert M. Davis, Graduate Technical Writing Intern, IBM Quantum
Quantum machine learning (QML) algorithms have the potential to solve certain kinds of problems much faster than classical algorithms, but how much faster? And which problems will they solve? In a recent panel, a group of experts underscored the fact that much about the science of QML is still unknown — making it an exciting time to be in the field, but an important place to set expectations.
That conversation took place during the Qiskit Global Summer School Commencement, a live digital event marking the conclusion of Qiskit’s annual, two-week online course. This year’s summer school curriculum centered on QML, and the commencement event brought together a panel of QML experts for a wide-ranging conversation about the future of the field. Panelists included IBM researcher Kristan Temme, MIT physics professor Aram Harrow, University of Washington computer scientist Ewin Tang, and Maria Schuld, a senior researcher and software developer at Xanadu Quantum Technologies — moderated by Amira Abbas, organizer and main lecturer of this year’s Qiskit Global Summer School. The question of where and how QML algorithms will best classical algorithms quickly emerged as a central theme.
Academic and industry researchers have already invested considerable resources into the search for applications where quantum computers will outperform their classical counterparts. Quantum computing technology is still relatively new, however, and that search has yet to pay dividends. The field of QML is no exception.
At the most basic level, QML is a research area that integrates quantum computing with machine learning techniques. Since quantum computers have the potential to quickly solve specific classes of problems that might be exceedingly difficult or even impossible for classical systems, researchers have long sought to identify machine learning applications that may get a boost from the power of quantum computation.
According to the panelists at the Qiskit Global Summer School Commencement, however, strong evidence for significant speedups in QML has yet to emerge. “At this point, I’d say it’s a bit difficult to exactly pinpoint a given application that would be of value,” said IBM’s Kristan Temme.
There are proven examples of quantum advantages over classical computing for problems relevant to the field of the machine learning that aren’t yet accessible to today’s hardware. For example, Temme himself recently helped develop an algorithm that relies on quantum computers’ theoretical ability to compute discrete logarithms to provide a new way of looking at datasets in classification problems. Future quantum computers may one day tackle these problems exponentially faster than classical computers. But in Temme’s view, algorithms like these don’t quite meet the definition of true QML. “There, it’s more about actually extracting the valuable features from the data, as opposed to the actual learning part,” Temme said.
As many of the panelists observed, most quantum algorithms that deliver exponential speedups have done so by taking clever approaches to highly structured problems — i.e., problems with clear goals, verifiable answers, and well-defined strategies for finding their solutions. For example, Shor’s algorithm finds the prime factors of large integers exponentially faster than classical algorithms because factoring is a well-structured problem. Shor’s algorithm solves that problem by following many of the same steps that a classical algorithm would, but it leverages the properties of superposition, entanglement, and interference to find workarounds in places where classical computers struggle.
Machine learning, by contrast, usually involves finding patterns in messy, unstructured data. Quantum algorithms can be useful for unstructured problems, but so far, the speedups they offer are more modest. Grover’s algorithm, for example, can find a specific element or record in a messy, unstructured database with fewer steps than a classical search algorithm — but only offers a polynomial speedup over classical approaches. And given the limitations of near-term quantum hardware, and the relative maturity of classical algorithms, some have argued that such speedups shouldn’t be the focus of early-generation quantum processors.
Still, that doesn’t mean quantum computing won’t bring meaningful benefits to the field of machine learning. MIT physics professor Aram Harrow emphasized that there is still a great deal we simply don’t know.
“I would temper our pessimism with a giant amount of uncertainty,” Harrow said. He and the other panelists agreed that exponential quantum advantage is primarily found in well-structured problems, but Harrow argued that this could be for a number of reasons. “That might reflect the fact that only structured problems have quantum speedups, or it might reflect the fact that quantum algorithms have mostly been a theoretical enterprise so far,” he said.
Harrow went on to explain that, until very recently, researchers did not have access to quantum computers that allowed them to run experiments and find out how well their theories worked. Now, with the advent of accessible quantum computing services, researchers are able to test some of those theories on real quantum systems. Still, many more theories cannot yet undergo experimental evaluation due to the limitations of near-term quantum hardware. It’s possible that better hardware will uncover new theoretical advances and even undiscovered quantum advantages.
University of Washington computer scientist Ewin Tang suggested that it might be time to reframe how we think about the advantages of QML. She argued that the quantum speedup, at least as it is commonly understood, presupposes the existence of some classical machine learning pipeline that takes classical input data and transforms it into classical output data. In this context, the core idea of a speedup is that a QML algorithm can replace and improve upon some portion of the classical pipeline. However, according to Tang, this way of thinking essentially ignores the possibility that one may want to run their QML algorithm on quantum rather than classical data. In this case, a quantum computer could perhaps unlock a new computational paradigm, rather than attempt to improve upon state-of-the-art classical methods.
“There’s not as clear a notion of what it would mean to have a speedup if your input data is quantum,” Tang said. “So by talking about speedups in the first place, we’re sort of assuming some regime in which classical computers already have an advantage. And I’m not sure if that’s necessarily what the goal for near-term quantum machine learning is.”
Perhaps, then, there are other important objectives aside from searching for speedups in machine learning. One useful perspective came from Xanadu Quantum Technologies researcher Maria Schuld.
“I wonder if the important question here is: Let’s build infrastructure. Let’s build the tools and the methods and understanding of the mix of mechanisms. Let’s be like proper physicists. Let’s create an Ising model,” a ubiquitous tool for describing complex physical systems, “that generations of people can use to sharpen their thoughts.”
For more insights on the future of quantum machine learning, be sure to check out the entire panel discussion, which is available on YouTube. For more stories like these, follow the Qiskit Medium!
