Neural networks can learn to utilize correlated auxiliary noise

Aida Ahmadzadegan, Petar Simidzija, Ming Li, Achim KempfScientific 2021 · DOI

Abstract

We demonstrate that neural networks that process noisy data can learn to exploit, when available, access to auxiliary noise that is correlated with the noise on the data. In effect, the network learns to use the correlated auxiliary noise as an approximate key to decipher its noisy input data. An example of naturally occurring correlated auxiliary noise is the noise due to decoherence. Our results could, therefore, also be of interest, for example, for machine-learned quantum error correction.

Introduction

In signal processing, noise is typically viewed as a detrimental factor to be filtered or suppressed. However, recent research published in Scientific Reports introduces a counterintuitive paradigm: utilizing "correlated auxiliary noise" (UCAN) to enhance data recovery. By providing a neural network with access to external noise sources that share patterns with the primary signal's interference, the model can learn to "subtract" the noise, effectively using it as a decryption key.

The Challenge of Correlated Noise

Traditional denoising methods often struggle when noise is not purely random but structured by the environment. This is particularly prevalent in quantum computing, where "decoherence" causes information to leak from qubits into their surroundings. This leakage creates environmental noise that is intrinsically correlated with errors in the quantum processor. The UCAN framework explores whether a neural network can leverage this environmental information to correct errors in the primary data.

Methodology and Experimental Setup

To evaluate the UCAN approach, researchers utilized the Fashion-MNIST dataset, consisting of 70,000 grayscale images across 10 categories. The images were modified to include a 6-pixel wide "bezel" or border, creating a 40x40 pixel input.

Visualization of the Fashion-MNIST dataset with varying noise levels (30% to 95%) and the bezel setup used for UCAN training.

The experimental design compared three scenarios:

  1. Set A: Clean images with a neutral bezel.
  2. Set B: Noisy images with a neutral bezel (no auxiliary information).
  3. Set C (UCAN): Noisy images where the bezel contains noise mathematically correlated with the noise obscuring the central image.

The team varied the noise-to-signal ratio and the dimensionality of the noise—defined by the number of independent spatial modes—to test the limits of the network's learning capacity.

Key Insights: Noise as a Reference

The core finding is that Convolutional Neural Networks (CNNs) do not simply learn to ignore the noisy bezel. Instead, they treat the bezel noise as a reference pattern. By identifying the specific noise structure in the bezel, the network implicitly "subtracts" that same pattern from the main image, maintaining high classification accuracy even in extreme noise regimes.

Dimensionality and Complexity

The effectiveness of UCAN is influenced by the "dimensionality" of the noise. As the number of independent modes increases, the noise becomes more diverse and harder to characterize.

Noise distributions for different numbers of modes (5x5, 15x15, 22x22), illustrating how increasing dimensionality creates more complex interference patterns.

While CNNs excel at utilizing structured noise (e.g., lower dimensionality), they face a "capacity limit" when correlations become algorithmically complex. This suggests that while classical networks are powerful, they may eventually be surpassed by quantum neural networks, which are naturally better suited for processing complex probability distributions.

Results and Performance

The UCAN-enabled models demonstrated significant resilience:

  • Accuracy Recovery: In scenarios where standard CNN accuracy dropped to 10% (equivalent to random guessing), UCAN models maintained accuracies between 40% and 90%, depending on noise dimensionality.
  • Efficiency Scaling: The performance gap between UCAN and standard models actually widens as noise levels increase, showing that the method is most valuable in the most challenging environments.

Conclusion and Future Directions

The UCAN framework represents a shift from noise suppression to noise utilization. While classical computational capacity currently limits the complexity of noise patterns that can be learned, the implications for quantum error correction are profound. Future work may involve hybrid quantum-classical networks to perform "ghost imaging" style error correction, where monitoring environmental noise allows for the real-time reconstruction of quantum information lost to decoherence.

Summary updated

Abstract

We demonstrate that neural networks that process noisy data can learn to exploit, when available, access to auxiliary noise that is correlated with the noise on the data. In effect, the network learns to use the correlated auxiliary noise as an approximate key to decipher its noisy input data. An example of naturally occurring correlated auxiliary noise is the noise due to decoherence. Our results could, therefore, also be of interest, for example, for machine-learned quantum error correction.
BibTeX
@inproceedings{ahmadzadegan2021neural,
  author = {Aida Ahmadzadegan and Petar Simidzija and Ming Li and Achim Kempf},
  title = {{Neural networks can learn to utilize correlated auxiliary noise}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-00502-4},
}
Figure