Booming interestImagine the Avengers of futuristic computing, merging machine learning and quantum computers — a captivating alliance capturing attention akin to the superhero dream team in comic books and films. However, just like fiction, a compelling plot is crucial in technology.
The potential of quantum computers, if built at sufficient scales, lies in efficiently solving specific problems, utilizing the subatomic world's unique properties. A long-standing question in research revolves around whether this efficiency could extend to machine learning, an artificial intelligence (AI) subset detecting patterns in data, learning rules for unfamiliar situations.
The recent introduction of the high-profile AI system ChatGPT, relying on machine learning for eerily human-like conversations by deducing word relationships, coincides with quantum computers' rapid growth in size and power. The convergence of these technologies raises anticipation, but the question lingers: will meaningful advancements emerge from their fusion?
Booming interest
Numerous tech enterprises, including major players like Google and IBM, along with startups such as Rigetti in Berkeley, California, and IonQ in College Park, Maryland, are exploring the potential of quantum machine learning. Academic scientists are also actively engaged in this pursuit.
CERN, situated outside Geneva, Switzerland, a prominent European particle-physics laboratory, employs machine learning to scrutinize data from the Large Hadron Collider for indications of specific subatomic particle production. Scientists at CERN are among those in academia experimenting with quantum machine learning.
Physicist Sofia Vallecorsa, leading a research group at CERN focusing on quantum computing and machine learning, articulates the goal: "Our idea is to use quantum computers to expedite or enhance classical machine-learning models."
A pivotal, unresolved query revolves around whether quantum machine learning holds advantages over its classical counterpart. Theoretical insights suggest quantum computers excel in specialized tasks, like simulating molecules or deciphering large prime numbers' factors. However, tangible evidence supporting these advantages in machine learning applications is still insufficient. Some argue that quantum machine learning may discern patterns overlooked by classical computers, even if speed isn't the defining factor.
Maria Schuld, a physicist based in Durban, South Africa, notes a spectrum in researchers' attitudes towards quantum machine learning. While interest persists, there's a growing sense of resignation regarding short-term applications, according to Schuld, associated with quantum-computing firm Xanadu in Toronto, Canada.
A shift in focus is evident among some researchers, directing attention towards applying quantum machine-learning algorithms to inherently quantum phenomena. Physicist Aram Harrow at MIT highlights this as a promising domain showcasing a distinct quantum advantage.
Do quantum algorithms help?
Over the last two decades, researchers in the field of quantum computing have formulated an array of quantum algorithms with the potential to enhance the efficiency of machine learning. A significant breakthrough occurred in 2008 when MIT physicists Harrow, Lloyd, and Hassidim developed a quantum algorithm that exhibited exponential speed-up compared to classical computers in solving large sets of linear equations—a core challenge in machine learning.
However, not all quantum algorithms have lived up to their promises. In a notable case from 2018, computer scientist Ewin Tang devised a classical algorithm that outperformed a quantum machine-learning algorithm designed in 2016. This quantum algorithm aimed to provide recommendations similar to those seen in online shopping and services like Netflix, demonstrating an exponential speed advantage over any known classical algorithm.
Tang, an undergraduate at the University of Texas at Austin at the time, crafted an algorithm nearly as fast but executable on a conventional computer. This incident cast doubt on the pursuit of achieving exponential quantum speed-ups in practical machine-learning problems, according to Scott Aaronson, a quantum-computing researcher at UT who advised Tang. Tang, now at the University of California, Berkeley, remains skeptical of claims asserting significant quantum speed-ups in machine learning.
A more significant challenge lies in the compatibility of classical data with quantum computation. A typical quantum-computing application involves three primary steps: initialization, where qubits are placed in an entangled quantum state; a sequence of quantum operations analogous to classical logical operations; and a read-out, often achieved by measuring the state of a single qubit conveying information about the quantum operation's result. The integration of classical data into this process poses a complex hurdle, potentially limiting the seamless fusion of classical and quantum computing.
The thinnest of straws
Quantum algorithms, exemplified by the work of Harrow, Hassidim, and Lloyd, pledge to expedite the second phase of quantum operations. However, the potential gains are often nullified by the sluggishness of the first and third steps in many applications. The initialization process involves transferring 'classical' data onto the quantum computer and converting it into a quantum state—a frequently inefficient procedure. Moreover, the inherently probabilistic nature of quantum physics introduces randomness into the read-out, necessitating multiple repetitions of all three stages to average results and derive a conclusive answer.
Even after processing classical data into a final quantum state, extracting an answer could be time-consuming, notes Nathan Wiebe, a quantum-computing researcher at the University of Washington. The challenge lies in obtaining information from the quantum state, akin to drawing from the thinnest of straws, Wiebe remarked during a quantum machine-learning workshop in October.
According to Schuld, many researchers emphasize that speed isn't the sole criterion for evaluating a quantum algorithm. There are indications that a quantum AI system, driven by machine learning, might excel in recognizing patterns missed by its classical counterparts. This potential advantage could stem from quantum entanglement creating correlations among quantum bits and, consequently, among data points, as explained by Karl Jansen, a physicist at the DESY particle-physics lab in Zeuthen, Germany. Jansen expresses hope that quantum algorithms could unveil correlations challenging for classical algorithms to detect.
Quantum machine learning could help to make sense of particle collisions at CERN, the European particle-physics laboratory near Geneva, Switzerland.Credit: CERN/CMS Collaboration; Thomas McCauley, Lucas Taylor
Aaronson holds a differing viewpoint. Quantum computers adhere to established laws of physics, making their operations and the results of quantum algorithms entirely foreseeable by a regular computer, with ample time. According to Aaronson, the primary concern is whether the quantum computer outpaces a flawless classical simulation of its performance.
In essence, the only pertinent query revolves around the quantum computer's comparative speed against an impeccably simulated classical counterpart, as outlined by Aaronson.
Fundamental quantum change
An alternative approach is to bypass the challenge of translating classical data altogether, opting instead for quantum machine-learning algorithms applied to data that inherently possess quantum properties.
Traditionally, measuring a quantum phenomenon involved using instruments rooted in the macroscopic, classical world. A novel concept, quantum sensing, is emerging, enabling the measurement of quantum properties using exclusively quantum instrumentation. By directly loading these quantum states onto the qubits of a quantum computer, quantum machine learning could discern patterns without any reliance on a classical system.
In the realm of machine learning, this approach presents significant advantages over systems that gather quantum measurements as classical data points, asserts Hsin-Yuan Huang, a physicist at MIT and a Google researcher. Huang contends that our world fundamentally operates on quantum principles, making a quantum machine capable of learning inherently more potent.
Huang and his collaborators conducted a proof-of-principle experiment using one of Google's Sycamore quantum computers. They allocated some qubits to simulate the behavior of an abstract material, with another section of the processor analyzing the information through quantum machine learning. The results revealed an exponential speed-up compared to classical measurement and data analysis.
Is it a superconductor?
Engaging in the full spectrum of data collection and analysis within the quantum realm could empower physicists to confront inquiries that classical measurements only address indirectly, as suggested by Huang. A prime example is determining if a specific material resides in a particular quantum state indicative of superconductivity—possessing the ability to conduct electricity with virtually zero resistance. Traditional experiments necessitate indirect proofs of superconductivity, such as examining the material's response to magnetic fields.
Exploration within the realm of particle physics extends to the utilization of quantum sensing for managing data generated by upcoming particle colliders, exemplified by the LUXE experiment at DESY, where electrons and photons will collide, notes Jensen. Despite the concept being a decade away from realization, distant astronomical observatories might harness quantum sensors to gather and transmit data via a prospective 'quantum internet' to a central lab for processing on a quantum computer. The aspiration is to capture images with unparalleled sharpness.
Successful applications of quantum sensing could pave the way for quantum machine learning to play a role in consolidating measurements from these experiments and analyzing the resultant quantum data.
Ultimately, the determination of whether quantum computers bring advantages to machine learning lies in experimentation, rather than relying on mathematical proofs of their superiority or lack thereof. Harrow underscores the impracticality of expecting everything to be proven in the manner typical of theoretical computer science.
Aaronson asserts the ongoing value of studying quantum machine learning, irrespective of whether an observable efficiency boost materializes. Schuld concurs, advocating for research unencumbered by the immediate need to prove a speed-up, at least for the time being.
References
1. Harrow, A. W., Hassidim, A. & Lloyd, S. Phys. Rev. Lett. 103, 150502 (2009).
2. Tang, E. in Proc. 51st Annu. ACM SIGACT Symp. Theory Comput. (eds Charikar, M. & Cohen, E.) 217–228 (ACM, 2019).
3. Biamonte, J. et al. Nature 549, 195–202 (2017).
4. Huang, H.-Y. et al. Science 376, 1182–1186 (2022).
-Google Scholars
