Quantum Machine Learning's Potential Impact on Science

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Jan 4 2024

An article recently published in the journal Nature discussed the potential of quantum machine learning (QML) in revolutionizing the field of science. 

Growing attention on QML

QML, which can be realized by combining two advanced technologies of quantum computing and machine learning (ML), has recently gained significant attention in the scientific community and among tech companies. Quantum computers built at sufficiently large scales can potentially solve several problems more effectively compared to conventional digital electronics by harnessing the subatomic world’s unique properties.

One of those problems can be ML, an artificial intelligence (AI) form in which computers detect patterns in data and learn rules to make inferences in new/unfamiliar situations. The introduction of Chat generative pre-trained transformer (GPT), a highly advanced AI system that depends on ML to drive its human-like conversations, and the rapid growth in the power and size of quantum computers have increased the focus on QML.

For instance, tech companies, such as IBM and Google, and start-ups, such as Rigetti and IonQ, are studying the prospects of QML. The scientific community has also shown strong interest in QML. For instance, scientists from CERN are primarily focusing on using quantum computers to improve or speed up classical ML models. However, the application scenarios in which QML is advantageous over classical ML are yet to be conclusively identified.

Although quantum computers can theoretically accelerate calculations for specialized computing tasks, like simulating molecules, researchers currently lack enough evidence to support this case. QML has also been proposed for more effectively detecting patterns that classical computers miss, even when it is not faster than classical counterparts.

However, researchers are increasingly becoming convinced about the lack of prospects of this approach for short-term applications, with several of them shifting their focus on applying QML algorithms to inherently quantum phenomena, as QML has a clear advantage in this application area.

Challenges in QML

In the last two decades, several quantum algorithms have been developed that can theoretically increase the efficiency of ML. For instance, a quantum algorithm was invented in 2008 that is exceptionally faster compared to classical computers at solving large linear equation sets, one of the core challenges of ML. However, in some cases, quantum algorithms have been ineffective.

For instance, an algorithm that can run on an ordinary computer was developed in 2018. The performance demonstrated by this algorithm was similar to the performance of a QML algorithm devised in 2016 to provide recommendations that Internet shopping services/companies give to customers based on their previous choices at an exponentially faster rate compared to classical algorithms.

This development indicated that the claims of significant quantum speed-up in practical ML problems must be thoroughly evaluated. Another significant challenge is efficiently fusing quantum computation and classical data in all instances. A quantum-computing application primarily has three major steps, including the initialization of the quantum computer, followed by the computer performing a sequence of operations, and finally, the computer performing a read-out.

Although the quantum algorithms developed can speed up the quantum operations/second step, the first and last steps can be very slow in several applications and negate the gains achieved by quantum algorithms. Specifically, the loading of classical data in the first step/initialization step onto the quantum computer and translating the data into a quantum state is an inefficient process.

Additionally, the read-out can have an element of randomness as quantum physics is intrinsically probabilistic, requiring the computer to repeat all stages several times and average the obtained results to get the final answer. Thus, no conclusive evidence has been obtained until now that displays the feasibility of using quantum computers on classical data/need for quantum effects on classical data.

The way forward

The issues associated with using classical data can be avoided using QML algorithms on quantum data. The quantum sensing technique enables the measurement of a system’s quantum properties using only quantum instrumentation. This approach involves loading quantum states directly onto a quantum computer’s qubits and then using QML to detect patterns without interfacing with a classical system. Thus, the technique can offer significant advantages over systems where quantum measurements are collected as classical data points in ML, as the world is inherently quantum-mechanical.

A proof-of-principle experiment performed on one of Google’s Sycamore quantum computers demonstrated that the technique is exponentially faster compared to classical data analysis and measurement. Physicists can effectively tackle questions indirectly answered by classical measurements by analyzing and collecting data entirely in the quantum domain. For instance, particle physicists can use quantum sensing to tackle data generated by future particle colliders.

Similarly, astronomical observatories in faraway locations can utilize quantum sensors for collecting data and transmitting them to a central lab, where the data will be processed using a quantum computer to obtain images with exceptional sharpness. In conclusion, experimentation can decide whether quantum computers will benefit ML in place of mathematical proofs. Thus, more research is required on QML without excessively focusing on the “quantum speed-up” aspect.