New methods for processing large classical datasets represent a rapid advance in quantum computing. Google Quantum AI’s Haimeng Zhao and colleagues have demonstrated that relatively small quantum computers can classify and reduce dimensionality of large amounts of classical data much more efficiently than classical machines. Machine learning on classical data is now an important area where quantum computers outperform classical computers, using fewer than 60 logical qubits to achieve size reductions of four to six orders of magnitude. This benefit comes from a new technique called quantum oracle sketching. This gives quantum computers unique access to classical information, bypassing the limitations that hinder classical machine learning approaches. The impact of this research has the potential to extend to numerous fields that rely on large-scale data analysis, such as genomics, materials science, and financial modeling, to accelerate discovery and innovation.
Reducing traditional data input bottlenecks with quantum sampling
The quantum oracle sketch at the heart of this advance works by allowing quantum computers to “peek” at classical data without having to copy the entire data set. By processing samples on-site, it accomplishes this, similar to quickly surveying a field to estimate crop yield, without having to count each plant individually. Traditionally, loading classical data into a quantum computer requires converting the data into a quantum state, but this process does not scale well with data size, creating a significant bottleneck. This conversion typically involves measuring classical bits and using the results to prepare qubits, a resource-intensive operation. Quantum oracle sketches avoid this complete data loading by adopting a probabilistic approach. Quantum computers query classical data using specially designed “oracles” that provide information about the data without revealing the entire data. This oracle is built to allow quantum algorithms to extract relevant features from data with a limited number of queries. The number of required queries grows polylogarithmically with data size. This means that even for very large datasets, the computational cost increases relatively slowly.
This method relies on a polylogarithmic number of qubits and handles noisy data, but the increase in sample requirements is proportional to the repetition of the data. Quantum advantages in classical data processing were observed for fewer than 60 logical qubits. This simplified data representation captures important features without requiring a complete exact copy. The impact of this technique extends to scenarios with inherently repetitive or incomplete data, as increased sampling can accommodate imperfections without sacrificing performance. Robustness to noise is critical because current quantum hardware is sensitive to errors. Although the researchers have adopted techniques to reduce these errors, further improvements in quantum error correction are essential to extend the method to larger datasets. Polylogarithmic scaling of qubit requirements is particularly important because it suggests that even modestly sized quantum computers can tackle problems that are currently unsolvable with classical machines. This is in contrast to many other quantum algorithms, which require many qubits that scale linearly or polynomially depending on the problem size, making them impractical for short-term implementation.
Quantum oracle sketch enables efficient analysis of large datasets
Single-cell RNA sequencing and sentiment analysis of movie reviews achieved a six-order reduction in machine size and demonstrated the clear advantages of quantum computers in processing large datasets. Previously, classical machines required exponentially larger sizes to achieve comparable performance on classification and dimensionality reduction tasks. Such scale was impractical and limited the analysis of large datasets common in modern science and industry. Polylogarithm-sized quantum computers can now perform these tasks efficiently. Single-cell RNA-sequencing allows researchers to use this technology to analyze gene expression patterns in thousands of individual cells and identify subtle differences that may be missed by traditional methods. Sentiment analysis of movie reviews can process large numbers of reviews to accurately measure public opinion. These applications highlight the potential of quantum machine learning to derive insights from complex data that is currently hidden.
Datasets from single-cell RNA-seq and movie review sentiment analysis validated these quantum advantages, achieving a 4- to 6-order reduction in machine size with less than 60 logical qubit equivalents. Even a slightly smaller classical machine would require a hyperpolynomial increase in both the number of data samples and processing time if attempting the same task. This advantage persists whether or not classical computers are given unlimited computing time, or even if the theoretical equivalence between quantum and classical computation known as BPP=BQP holds. Careful calibration of the quantum system minimized noise and ensured accurate sampling during the experimental setup. Experimental validation includes a rigorous comparison of the quantum algorithm’s performance with state-of-the-art classical machine learning algorithms. The results consistently showed that the quantum algorithm achieves significantly better performance in terms of both accuracy and computational cost. To obtain reliable results, the use of logical qubits that are protected from errors by quantum error correction is essential. However, the overhead associated with quantum error correction remains a major challenge in building practical quantum computers.
Scalable error correction is needed to accelerate quantum computation in machine learning
Researchers and Google Quantum AI are unlocking the potential of quantum computers to tackle problems that are beyond the reach of even the most powerful supercomputers. Efficient processing of vast amounts of classical data underpins this new capability and serves as the foundation for modern machine learning and scientific discovery. However, fully characterizing the hardware requirements, especially error correction, remains an open problem. Although the current demonstration uses fewer than 60 logical qubits, scaling the method to larger datasets will require significant advances in quantum hardware and error correction techniques. The development of fault-tolerant quantum computers that can correct errors without destroying quantum information is a major focus of ongoing research.
Demonstrating significant computational savings of four to six orders of magnitude on real-world datasets such as genome sequencing and movie reviews is a breakthrough. Machine learning using classical data is a promising area where quantum computers can outperform classical computers despite current limitations, highlighting the need for continued investment in quantum hardware development. This demonstration of quantum advantage establishes a new benchmark for practical quantum computing, going beyond theoretical possibilities and yielding tangible benefits in classical information processing. By bypassing the limitations inherent in loading large datasets into quantum states, polylog-sized quantum computers can classify and reduce the dimensionality of data much more efficiently than classical quantum computers, validating machine learning on classical data as a key area where quantum computers clearly excel. Future research will focus on optimizing the quantum oracle sketch algorithm and exploring its applicability to a broader range of machine learning tasks. The ultimate goal is to develop quantum machine learning algorithms that can solve real-world problems that classical computers cannot currently solve.
An important discovery is the demonstration of significant computational savings of four to six orders of magnitude for tasks such as single-cell RNA sequencing and sentiment analysis of movie reviews. This means that quantum computers using fewer than 60 logical qubits can perform these tasks with significantly less computational power than classical machines require. This research demonstrated that such quantum computers can efficiently process large amounts of classical data on the fly. This is a feat impossible on classical machines, which do not grow exponentially in size. The researchers plan to use quantum oracle sketches and classical shadows to validate this benefit and optimize the algorithm for broader applications.
