Keio University Team Proposes Quantum Feature Selection Method for Enhanced Machine Learning

Researchers from Keio University and JSR Corporation have developed a feature selection method specifically for quantum machine learning. Traditional feature selection methods, which remove redundant features to improve predictive performance, are not fully applicable to quantum tasks.

The new method treats lightcones structures as quantum features, selecting beneficial ones through kernel parameter training. This allows for feature selection of classical inputs and quantum data tasks. The method has shown potential in four applications, including circuit architecture search for data embedding and compression of quantum machine learning models. This development could broaden the applicability of quantum machine learning.

What is the Role of Feature Selection in Quantum Machine Learning?

Feature selection is a critical aspect of machine learning, both classical and quantum. It plays a pivotal role in enhancing the predictive performance and interpretability of trained models. In classical machine learning, feature selection techniques help in removing redundant features, thereby improving the predictive performance. Moreover, the interpretability of the models can be enhanced due to the reduced features.

However, the usability of conventional feature selection could be limited for quantum machine learning tasks. While classical feature selection methods still play a critical role in embedding data on quantum circuits, it does not help understand why some data embedding quantum circuits work better than others. More importantly, when dealing with quantum data tasks such as quantum phase recognition, classical input features do not explicitly appear in input quantum data and thus one cannot employ the classical techniques straightforwardly.

In this context, a team of researchers from the Department of Mechanical Engineering at Keio University, Materials Informatics Initiative RD Technology & Digital Transformation Center at JSR Corporation, and Quantum Computing Center at Keio University, have proposed a feature selection method with a specific focus on quantum machine learning problems.

How Does the Proposed Feature Selection Method Work?

The proposed feature selection method treats the lightcones structures, i.e., subspaces of local quantum kernels, as quantum features and then selects a subset of beneficial ones through the training of parameters in the kernels. By local quantum kernels, the researchers mean a family of quantum kernels that measure local similarity between a pair of data examples, such as projected quantum kernels (PQKs) and quantum Fisher kernels (QFKs).

This extension makes the proposed framework applicable not only to the feature selection of classical inputs but also to the quantum data tasks that conventional feature selection methods cannot handle. This is the first work to introduce a quantum machine learning-oriented feature selection method.

What are the Applications of the Proposed Feature Selection Method?

The researchers have demonstrated the effectiveness of their proposed feature selection method through numerical simulations using toy tasks. The method has shown versatility for four different applications:

1. Feature selection of classical inputs: The method can be used to select relevant features from classical data inputs, thereby improving the performance of quantum machine learning models.
2. Circuit architecture search for data embedding: The method can help in searching for the most effective circuit architecture for embedding data in quantum circuits.
3. Compression of quantum machine learning models: The method can be used to compress quantum machine learning models, thereby making them more efficient and manageable.
4. Subspace selection for quantum data tasks: The method can be used to select relevant subspaces for quantum data tasks, thereby improving the performance of quantum machine learning models on these tasks.

What is the Significance of the Proposed Feature Selection Method?

The proposed feature selection method paves the way towards applications of quantum machine learning to practical tasks. It could be used to practically test if the quantum machine learning tasks really need quantumness. These results indicate that the proposal will encourage practitioners to use quantum machine learning models for real-world applications.

In conclusion, the proposed feature selection method for quantum machine learning tasks is a significant step forward in the field. It not only enhances the performance and interpretability of quantum machine learning models but also broadens their applicability to a wider range of tasks. This work is expected to have a profound impact on the future development and application of quantum machine learning.