Researchers at UC San Diego’s Picasso Lab have released technical results evaluating the NVIDIA Ising neural pre-decoder. The study explores how lightweight neural networks can accelerate and improve Quantum Error Correction (QEC) by preprocessing syndromes before they reach a primary decoder like PyMatching. While the system demonstrated significant performance gains on traditional surface codes, the research also investigated the transition to quantum Low-Density Parity-Check (qLDPC) codes, specifically bivariate bicycle (BB) codes, which offer a higher encoding rate.

The Role of the Pre-Decoder

A pre-decoder acts as a “filter” that identifies and corrects high-confidence, local errors. By outputting a “sparser residual” syndrome, it simplifies the task for the primary decoder. In surface codes—where qubits are arranged on a regular 2D grid—the researchers utilized a 3D Convolutional Neural Network (CNN). Because the CNN’s receptive field matches the code’s regular spatial structure, it effectively captures complex spatiotemporal noise correlations that standard decoders often approximate or discard.

Performance on Surface Codes

The evaluation of the NVIDIA Ising CNN on surface codes (distance d=9) revealed two primary technical advantages:

  • Accuracy: The model achieved a 1.66x reduction in Logical Error Rate (LER). By learning hardware-specific noise patterns (such as CNOT propagation), the CNN identifies errors that conventional Minimum-Weight Perfect Matching (MWPM) might misinterpret.
  • Latency: The “sparsification” of syndromes allowed PyMatching to run up to 2.12x faster. As code distance increases, this classical speedup becomes more pronounced, addressing a critical bottleneck in real-time fault tolerance.

Transition to Bivariate Bicycle Codes

The study contrasted these results with an MLP (Multi-Layer Perceptron) pre-decoder designed for bivariate bicycle codes. BB codes are highly efficient; for example, a [[144, 12, 12]] code stores 12 logical qubits in 144 physical qubits, a 12x improvement over surface codes. However, these codes feature non-local connectivity, which does not align with a CNN’s local filters.

The results showed that while the MLP provided a 14x LER reduction at very low physical error rates (p=0.01), the benefit vanished at higher error rates. This highlights a core principle identified by the Picasso Lab: architecture-code alignment. A neural pre-decoder is most robust when its internal structure (like a CNN’s grid) reflects the physical connectivity of the quantum code’s Tanner graph. Future work will focus on using Graph Neural Networks (GNNs) to better match the non-planar connectivity of qLDPC codes.

For the full technical blog and benchmarking code, visit the Picasso Lab at UC San Diego here.

April 15, 2026

Print Friendly, PDF & Email