Yan and Colleagues Develop Neural Transfer Unification for Efficient Quantum Decoding

A new framework, neural transfer unification (NTU), accelerates the training of foundation decoders by aligning decoding tasks across different code sizes. Ge Yan and colleagues from College of Computing and Data Science and Singapore University of Technology and Design and Nanyang Technological University and Tokyo University of Agriculture and Technology and Shanghai Jiao Tong University, have created this set of tools to address a key challenge in fault-tolerant quantum computing: the efficient decoding of quantum information at increasing code distances. Their implementation, NTU-Transformer, shows improved performance on planar surface codes and bivariate bicycle codes, surpassing existing methods such as correlation-aware matching and Relay-BP, and providing a scalable path towards amortized cross-distance training for future quantum processors.

Neural transfer unification enables efficient scaling of decoding performance

NTU-Transformer achieved a computational scaling exponent of 1.51, a marked improvement over the 1.64 previously required for training foundation decoders. This reduction unlocks the possibility of scaling to code distances previously unattainable due to prohibitive computational cost. Foundation decoders, a promising architecture for fault-tolerant quantum computation, rely on neural networks to infer the most likely original quantum state given a noisy, error-ridden measurement. However, the complexity of training these networks increases dramatically with the code distance, a measure of the code’s ability to protect quantum information. A lower scaling exponent signifies a more efficient training process, allowing for larger code distances to be reached with a given computational budget. The previous scaling exponent of 1.64 indicated a rapid increase in computational resources needed as code distance increased. NTU’s reduction to 1.51 represents a substantial improvement in scalability. This is crucial because achieving fault tolerance necessitates codes with large code distances to effectively suppress errors.

Neural transfer unification, or NTU, aligns decoding tasks across different code sizes and enables knowledge transfer from smaller to larger codes, circumventing the need to retrain decoders from scratch for each new code distance. Traditionally, training a decoder for a code distance of, for example, 15, would require a completely new training process, even if a decoder for distance 7 already existed. NTU addresses this by formulating a unified training objective that considers multiple code distances simultaneously. The decoder learns to extract shared features and patterns applicable across different code sizes, effectively leveraging prior knowledge. This transfer learning approach significantly reduces the training time and computational resources required for larger codes. The underlying principle is analogous to how humans learn, applying knowledge gained from simpler tasks to more complex ones.

Gate fidelity increased five-fold, and NTU-Transformer outperformed correlation-aware matching on the $[..]$ code and exceeded standard matching on the $[..]$ code, demonstrating its ability to generalise and adapt. Gate fidelity is a critical metric in quantum computing, representing the accuracy of individual quantum operations. A five-fold increase indicates a substantial improvement in the reliability of the decoding process. The $[..]$ and $[..]$ codes represent specific instances of quantum error correcting codes, and outperforming existing methods on these benchmarks demonstrates the decoder’s effectiveness. Correlation-aware matching is a classical decoding algorithm that considers the correlations between errors, while standard matching is a simpler, less sophisticated approach. The ability to surpass these established methods highlights the advantages of the neural network-based approach and the effectiveness of the NTU framework. The decoder successfully processed a planar surface code of size $[..]$, outperforming correlation-aware matching, and then extended this performance to a $[..]$ code, exceeding the accuracy of standard matching techniques through the application of transferred knowledge. This sequential demonstration of performance improvement across increasing code sizes further validates the knowledge transfer capability of NTU. On the bivariate bicycle code with $[..]$, NTU-Transformer surpassed the performance of the Relay-BP decoder in scenarios with low physical error rates, indicating its strong durability. Relay-BP is another classical decoding algorithm, and its performance being exceeded by NTU-Transformer, particularly at low error rates, suggests the robustness of the framework.

While these results highlight a scalable approach to training, demonstrating practical quantum error correction requires achieving comparable performance with even more substantial code distances and realistic hardware noise models. Current quantum computers are prone to various types of noise, and accurately modelling these noise characteristics is crucial for developing effective decoders. The current work primarily focuses on idealised noise models; incorporating more realistic noise models will be essential for validating the decoder’s performance in real-world scenarios. Reduced computational demands mean less processing power is required for larger, more complex codes. This allows researchers to explore larger code distances and more sophisticated error correction schemes. Further research will focus on extending the framework’s capabilities to even larger code distances and incorporating more realistic hardware noise models to demonstrate practical quantum error correction.

Transfer learning accelerates decoder training for scalable quantum error correction

Researchers at Nanyang Technological University are building increasingly sophisticated neural decoders to correct errors in quantum computers, but scaling these systems to handle larger, more complex codes remains a formidable challenge. Quantum computers utilise qubits, which are susceptible to errors due to environmental noise and imperfections in quantum operations. Quantum error correction aims to protect quantum information by encoding it into multiple physical qubits, allowing for the detection and correction of errors. However, the decoding process, inferring the original quantum state from noisy measurements, is computationally intensive, particularly for large codes. The team’s new framework offers a potential solution by allowing knowledge gained from smaller codes to be applied to larger ones, reducing the computational demands of training. Current work concentrates on planar surface codes and bivariate bicycle codes, however, leaving open whether this approach will generalise to other, fundamentally different quantum error correction schemes. Planar surface codes are a leading candidate for fault-tolerant quantum computation due to their relatively simple structure and high threshold for error correction. Bivariate bicycle codes offer an alternative approach with potentially improved performance in certain scenarios.

These initial successes establish an important principle: knowledge gained from smaller, simpler quantum error correction systems can indeed accelerate the development of decoders for larger, more complex ones. This approach circumvents the typical computational bottlenecks associated with training neural decoders from scratch for each new code size, establishing a scalable path for training these essential components for correcting errors in future quantum computers. The traditional approach to training decoders involves treating each code size as an independent problem, requiring a separate training process for each. NTU’s unified framework allows the decoder to learn a more general representation of quantum errors, enabling it to adapt to different code sizes with minimal retraining. Aligning decoding tasks across different code sizes enabled knowledge transfer, reducing the computational demands of development. As a result, this work opens questions regarding the extent to which these transferable learning techniques can be broadened to encompass diverse quantum error correction schemes and hardware platforms. Investigating the applicability of NTU to other error correction codes, such as colour codes or topological codes, will be crucial for establishing its generality. Furthermore, exploring its performance on different hardware platforms, including superconducting qubits, trapped ions, and photonic qubits, will be essential for demonstrating its practical relevance.

The researchers demonstrated a new framework, neural transfer unification, which accelerates the training of neural decoders for quantum computers. This matters because training these decoders typically becomes computationally expensive as code distances, a measure of a code’s ability to correct errors, increase. By aligning decoding tasks across codes of varying sizes, the NTU-Transformer decoder outperformed existing methods on planar surface codes of sizes $[![361,1,19]!]$ and $[![625,1,25]!]$, and the $[![72,12,6]!]$ bivariate bicycle code. The authors intend to investigate whether this approach can be extended to other quantum error correction schemes and hardware platforms.