Quantum processors now access qubit states quickly enough for real-time error correction. Spiking neural networks classify qubit states as measurement data arrives, circumventing the need to wait until an entire trace is acquired. This delivers a continuously updated estimate of the qubit’s condition; surpassing matched filtering techniques and equalling the accuracy of conventional artificial neural networks. A new system utilising ‘spiking’ neural networks rapidly determines the condition of qubits, the fundamental building blocks of quantum computers.

Unlike existing methods which require complete data collection before analysis, this technique processes incoming signals in stages allowing for continuous updates regarding the qubit’s status. Researchers at Ulster University and collaborating institutions have pioneered a new approach to reading information from qubits using artificial intelligence inspired by the human brain. The system employs ‘spiking’ neural networks that process incoming signals in stages offering continuous updates regarding the qubit’s status, unlike traditional methods requiring complete data collection before analysis.

Frequency-multiplexed readout is like listening to multiple radio stations simultaneously; separating each signal demands careful processing. The innovation surpasses matched filtering techniques, akin to comparing an unknown signal against pre-defined templates, while achieving comparable accuracy to conventional artificial neural networks. This allows for real-time assessment of a qubit’s condition and opens up possibilities for faster error correction but can these spiking networks be efficiently implemented on existing quantum hardware.

Spiking neural networks enable real-time superconducting qubit state estimation

Latency in assigning states to superconducting qubits has now been reduced, achieving inference updates within 50ns where previously reliance on full measurement records proved restrictive. A system utilising spiking neural networks (SNNs), developed at Ulster University with collaborators from Technical University of Munich and the Walther-Meißner-Institut, matches the accuracy of conventional artificial neural networks while simultaneously providing continuous, time-resolved estimations of qubit states during signal capture. Sequential processing analyses incoming signals in segments rather than awaiting completion; this circumvents limitations inherent in matched filtering techniques which struggle with complex multivariate data arising from frequency-multiplexed readout.

Inference updates within just 50ns were achieved through utilisation of field programmable gate arrays (FPGAs), specialised integrated circuits configurable after manufacturing. Quantisation-aware training, a technique reducing computational precision without significantly impacting accuracy, enabled more efficient FPGA processing.

Successful implementation using HTML synthesis confirms practical feasibility beyond simulation as the automated tool converts machine learning models into hardware descriptions suitable for deployment on FPGAs, while conventional matched filtering demands complete data acquisition before analysis. Processing incoming signals in segments offers continuous estimations of qubit states and circumvents limitations when dealing with complex signal variations inherent in frequency-multiplexed readout systems where multiple qubits share measurement lines.

Advancing responsive error correction through biologically inspired quantum control

The demonstration of spiking neural networks provides a pathway towards truly responsive quantum systems; rapid feedback loops are vital to accelerate error correction routines as processors scale in complexity. However, the current implementation focuses heavily on proving feasibility within a limited five-qubit setup, raising questions about maintaining these performance gains when applied to larger architectures. Scaling this approach beyond FPGA capacity presents significant engineering challenges given that multiplexed readout inherently increases data volume and signal intricacy.

Acknowledging limitations regarding scalability beyond five qubits is important for future development. Successful deployment via quantisation-aware training and hls4ml synthesis has demonstrated practical feasibility beyond simulation, paving the way for low-latency real-time processing which contrasts with conventional techniques reliant upon full measurement records. This limits speed and introduces latency into time-critical applications like error correction protocols or feedback control loops. Furthermore, this implementation of spiking neural networks offers a new method for analysing signals from superconducting qubits as these artificial intelligence systems process data incrementally as it arrives rather than requiring complete signal acquisition.

The research successfully demonstrates that spiking neural networks can accurately determine qubit states in superconducting quantum processors. These networks processed incoming readout signals segmentally, offering continuous estimations of state and reducing delays compared to methods needing complete data collection. Using quantisation-aware training and hls4ml synthesis, the team showed inference could be completed on FPGA hardware before subsequent data arrived. This approach provides low-latency processing which is important for time-critical applications such as error correction within a five-qubit system.

👉 More information
🗞 Spiking neural networks for streaming qubit readout
✍️ Barry M. Dillon, Aqib Javed and Jim Harkin (Ulster University); Benjamin Lienhard (Affiliation: Technical University of Munich); Patryk Dabkowski
🧠 ArXiv: https://arxiv.org/abs/2610.02129

More like this