Skip to content

Quantum Machine Learning in Your Business

Technology

Nov 3, 2022 - 4 minute read

2827 Quantum Machine Learning 416X300
Michał Bączyk Quantum Computing Specialist

He is a Cambridge and ETH Zurich graduate. Over the course of his career, Michał completed internships in Quantum AI groups at CERN and Los Alamos National Laboratory in the United States. In his work, he aims to connect the worlds of quantum computing and business.

See all Michał's posts

2988 HC Digital Transformation 476X381

Share

AI is Already Here

When you unlock your phone and check the travel time to the restaurant recommended to you for dinner, you have already used AI several times. By employing Graph Neural Networks, Google Maps improved the accuracy of ETAs predictions by 21% in New York, 16% in London, 21% in Berlin and 43% in Sydney. Apple enhanced Face ID security with machine learning, overall reducing the probability of a stranger unlocking your device to one in a million (with a single appearance). There’s no doubt AI has become integral to our everyday life. In addition, it influences business decisions and the shape of companies, for example by suggesting suitable candidates to recruiters through LinkedIn tools. Now a new era of quantum machine learning (QML) begins.

What is QML?

Quantum machine learning differs from machine learning in the same fundamental way as quantum computing from classical computing. Whenever we deal with quantum, the steps of algorithms are sent for evaluation to quantum processing units (QPUs) instead of the classical central processing units (CPUs) that perform calculations in our phones and laptops. To get ahead of CPUs, QPUs utilise laws of quantum physics to their advantage. Although it requires adjustments how you encode information and algorithm steps to QPUs, the high-level idea of ML algorithms is directly transferable to the quantum domain. Thus, QML algorithms fit into the categories of classical learning theory, and we distinguish accordingly:

  • Supervised QML – able to solve classification (“What type of product is this particular customer most likely to buy?”) and regression (“What is going to be the oil price tomorrow?”) problems.
  • Unsupervised QML – used to uncover insights about data through exploring, analysing, and simplifying its structure. (“What audience groups can we identify among our customers?”)
  • Reinforcement QML – prepared to make decisions following a predetermined measure of what constitutes success. (“What should be my real-time trading and pricing strategy for portfolio optimisation?”)

Case Studies

With machine learning impacting every person and every business, it is worth paying attention to the potential advantage that QML theoretically has over classical ML. In the following, we discuss two business cases from 2022 that present the capabilities and limitations of current QML approaches.

The first example of an advantage achieved by bringing QML into play comes in the automotive industry. As part of their quantum computing challenge, the BMW Group posted an anomaly detection classification problem – the task was to assess whether the currently operating, image-trained ML system for uncovering cracks in the manufactured components might be upgraded. High-resolution images of the produced components were necessary to effectively train the model. Moreover, thecompany needed a large number of them, since flaws like that occur very rarely. Obtaining and storing these images takes time and memory, which presents room for improvement.

A QML solution achieved 97% accuracy while the benchmark models reached 80%. The trick was not to adapt the whole problem to implement it in quantum hardware but to enhance the classical algorithm by running QPU calculations only in the most crucial parts of the analysis. Such a solution is a hybrid approach. Creating services in this paradigm makes QML applications more tangible and closer in reach.

Kudos for the second demonstration go to Barclays for implementing quantum neural networks (QNNs) in the financial context of time series forecasting, a technique indispensable in pricing as well as asset and risk management. They managed to show that QNNs as a stock pricing tool perform equally well as classical neural networks on all data while offering an advantage for signals with high fluctuations. This analysis was performed on both synthetically generated data as well as historical Apple stock and Bitcoin prices. While the solution is still limited in terms of the volume of data QML architecture can digest, Barclays plans to continue to evaluate the potential use cases focusing more on customer-oriented tasks such as fraud detection.

QML’s Secret Advantage

For the BMW case, QML solution not only achieved higher accuracy but did so using less information. The quantum model was trained on 40% of the whole dataset while the benchmark model was on 70%. This shows that the QML algorithm generalises more efficiently than the classical approach since it still manages to provide accurate predictions for new data with less input. The ability to generalise even from limited data is an inherent feature of QML. It comes in handy when training data is scarce or costly to obtain.

In addition, the Barclays proof of concept shows that QNNs for particular use cases are less prone to start treating the inevitable noise as data and learn it (overfitting). The reason is that QML architectures can be much, much smaller. For the scenario discussed, the quantum neural network used 1830 times fewer parameters than the classical one. Therefore, the ability-to-size ratio might also be a plus for QML approaches.

When Is the Question, Now Is the Answer

In this article, we discussed two real-life QML use cases from different industries. The growing awareness of quantum architectures' intrinsic features and capabilities allows for tackling business problems more effectively. Situations with low volumes of data can already be approached with added value via fully quantum algorithms even today. While machine learning implemented purely on today's quantum hardware is still limited, using QML in the right place as part of a hybrid quantum-classical architecture can be highly effective. As Objectivity, we continue to evaluate all cases where QML has the potential to revolutionise business. We follow trends and implement solutions ourselves, so we are ready to provide guidance to our clients. If you’re interested in learning more about solving business problems with this cutting-edge technology, feel free to visit our quantum computing services page.

2988 HC Digital Transformation 476X381
Michał Bączyk Quantum Computing Specialist

He is a Cambridge and ETH Zurich graduate. Over the course of his career, Michał completed internships in Quantum AI groups at CERN and Los Alamos National Laboratory in the United States. In his work, he aims to connect the worlds of quantum computing and business.

See all Michał's posts

Related posts

You might be also interested in

Contact

Start your project with Objectivity

CTA Pattern - Contact - Middle