Quantum Machine Learning: Munich Team Tackles Data Encoding Challenge

Researchers from the Technical University of Munich, Munich Center for Quantum Science and Technology, BMW Group Central Invention, and Leiden University have made strides in solving the data encoding problem in quantum machine learning (QML). They improved a variational algorithm to prepare encoded data using shallow circuits that fit the native gate set and topology of current quantum computers. The team applied the improved algorithm to the FashionMNIST dataset, a standard machine learning dataset, and achieved moderate accuracies on a quantum computer, providing a proof of concept for the near-term usability of their data encoding method.

What is the Potential Impact of Quantum Machine Learning Algorithms on Industrial Applications?

Quantum machine learning algorithms have the potential to revolutionize industrial applications, but their practical implementation remains an open question. Conventional methods for encoding classical data into quantum computers are costly and limit the scale of feasible experiments on current hardware. Despite claims of the near-term suitability of their algorithms, recent works do not provide experimental benchmarking on standard machine learning datasets.

A team of researchers from the Technical University of Munich, Munich Center for Quantum Science and Technology, BMW Group Central Invention, and Leiden University have attempted to solve the data encoding problem by improving a recently proposed variational algorithm. This algorithm prepares the encoded data using asymptotically shallow circuits that fit the native gate set and topology of currently available quantum computers.

The team applied the improved algorithm to encode the FashionMNIST dataset, a standard machine learning dataset, which they have made openly available for future empirical studies of quantum machine learning algorithms. They deployed simple quantum variational classifiers trained on the encoded dataset on a current quantum computer, ibmqkolkata, and achieved moderate accuracies. This provides a proof of concept for the near-term usability of their data encoding method.

How Does Supervised Machine Learning Fit into Quantum Computing?

Supervised machine learning has been viewed as a potentially promising application for quantum computers. Early supervised quantum machine learning (QML) algorithms suggested potential speedups against the best-known classical algorithms. This fostered the development of competing quantum-inspired classical algorithms.

With the development of noisy intermediate-scale quantum (NISQ) computers, supervised QML algorithms with parameterized quantum circuits (PQCs) have raised increasing attention. These algorithms have been rigorously shown to solve certain classically intractable learning problems efficiently. However, whether supervised QML algorithms will also achieve a practical advantage in any industry-relevant application remains an exciting and open question.

The practical application of supervised QML algorithms faces several challenges, including the data encoding problem. This is the preparation of quantum states that represent the training and testing data, which is a prerequisite for any supervised learning task. Conventional methods for data encoding need to compromise between the number of qubits and circuit depth.

What is the Data Encoding Problem in Quantum Machine Learning?

The data encoding problem is a significant challenge in the practical application of supervised QML algorithms. It involves the preparation of quantum states that represent the training and testing data, which is a prerequisite for any supervised learning task. Conventional methods for data encoding need to compromise between the number of qubits and circuit depth.

At one extreme, product encoding uses a tensor product circuit with just one single-qubit rotation per qubit but requires one qubit per dimension of data. At the other extreme, amplitude encoding utilizes the superposition property of quantum states to encode the data in the amplitudes of a logarithmic number of qubits, yet demanding a circuit of depth exponential in the number of qubits.

The data encoding problem not only acts as a bottleneck on the potential speedup in many supervised QML algorithms but also limits the scale of benchmarking problems that can be currently experimentally implemented. Most research either considered simple binary datasets or applied dimension reduction techniques such as principal component analysis or coarse-graining to preprocess the datasets.

How Can the Data Encoding Problem be Solved?

To solve the data encoding problem, the team of researchers improved a recently proposed variational algorithm. This algorithm prepares the encoded data using asymptotically shallow circuits that fit the native gate set and topology of currently available quantum computers.

The team applied the improved algorithm to encode the FashionMNIST dataset, a standard machine learning dataset. This dataset was then used for training a variational quantum classifier on classical computers. The team provided results from deploying the trained circuits on a quantum computer, ibmqkolkata.

This approach holds the broad potential to facilitate empirical studies of supervised QML algorithms on quantum computers in the near future. The team’s work is a significant contribution to the field, improving the resource efficiency of the PQCs in the variational algorithm and using only the typical native gates of current quantum computers in the ansatz.

What are the Implications of this Research?

The research conducted by the team from the Technical University of Munich, Munich Center for Quantum Science and Technology, BMW Group Central Invention, and Leiden University has significant implications for the field of quantum computing.

By improving the resource efficiency of the PQCs in the variational algorithm and using only the typical native gates of current quantum computers in the ansatz, the team has made a significant contribution to solving the data encoding problem. This work provides a proof of concept for the near-term usability of their data encoding method.

The team’s work also opens up the possibility for future empirical studies of quantum machine learning algorithms. By making the encoded FashionMNIST dataset openly available, they have provided a valuable resource for other researchers in the field. This research represents a significant step forward in the practical application of quantum machine learning algorithms in industrial applications.