DeepMind reduces quantum computing errors with AlphaQubit decoder
Trained on Google's 49-qubit Sycamore processor, the Transformer-based decoder cut errors 6% versus the most accurate prior method and 30% versus the fastest, but remains too slow for real-time use.
- Models & capabilities
- Minor
Google DeepMind and Google Quantum AI published AlphaQubit, a neural-network decoder that reads the consistency checks run on a quantum processor’s logical qubits and predicts where errors have occurred, so they can be corrected before they corrupt a computation. The work appeared as a peer-reviewed paper in Nature alongside the announcement.
The decoder, built on the Transformer architecture that underpins large language models, was trained on data from Google’s 49-qubit Sycamore processor and on simulated systems scaled up to 241 qubits. On new Sycamore data it made 6% fewer errors than tensor-network decoding, the most accurate method previously available, and 30% fewer than correlated matching, the faster method used where speed matters more than accuracy. It also held its accuracy on error-correction runs far longer than any it had trained on, up to 100,000 rounds.
DeepMind and Google Quantum AI were explicit about the gap between this result and a working machine: AlphaQubit is still too slow to decode errors in real time on a live superconducting processor, and scaling it to the millions of qubits a fault-tolerant computer would eventually need requires more data-efficient training than was used here. The result was presented as evidence that machine learning could substitute for the hand-engineered decoding algorithms quantum error correction has relied on, one of several instances of DeepMind turning its model-building expertise directly on a problem in physical science rather than on language or games.