Don't take it lightly: Phasing optical random projections with unknown operators

Sidharth Gupta, Rémi Gribonval, Laurent Daudet, Ivan Dokmanic

Random projections are a staple of modern machine learning, used for everything from dimensionality reduction to kernel approximation. While traditional CPUs and GPUs struggle with the memory and computational overhead of projecting massive datasets, Optical Processing Units (OPUs) offer a compelling alternative. By shining lasers through scattering media (like white paint), OPUs can perform these complex matrix multiplications at the speed of light with minimal energy consumption.

However, this NeurIPS 2019 paper by Gupta et al. highlights a fundamental limitation in optical computing: the loss of phase information. By reframing a physics problem as a geometry problem, the authors unlock the ability to perform complex-valued linear algebra on high-speed optical hardware without ever needing to calibrate the system.

The Challenge of Optical Computing

Despite their speed, OPUs face two major hurdles that limit their application in complex machine learning tasks:

  1. Phase Loss: Camera sensors only measure light intensity (the squared magnitude). This means the phase of the complex-valued projection is entirely lost during measurement.
  2. The "Black Box" Operator: The transmission matrix ($A$) representing the scattering medium is essentially unknown. While it could theoretically be calibrated, the process is slow, highly sensitive to environmental changes, and impractical for high-resolution applications.

The core challenge the authors tackle is: How do you recover the phase of a projection $y = A\xi$ if you only have its magnitude and don't even know what the operator $A$ is?

Methodology: Phase Retrieval as a Geometry Problem

The researchers' breakthrough relies on a simple observation: while they don't know the transmission matrix $A$, they do control the input signal $\xi$. They cast the "Measurement Phase Retrieval" (MPR) problem as a Euclidean distance geometry problem.

Optical setup and geometric interpretation of measurements.
  • Distance as Interference: By interfering the signal of interest $\xi$ with known reference signals (anchors, denoted as $r$), the OPU measures the squared magnitude of their difference: $|A(\xi - r)|^2$. In the complex plane, this physical measurement is mathematically equivalent to the squared Euclidean distance between the points $A\xi$ and $Ar$.
  • Multidimensional Scaling (MDS): If you collect enough pairwise distances between the signal and several different anchors, you can reconstruct the relative positions of these points in the complex plane. The authors use Multidimensional Scaling (MDS)—a classic manifold learning technique—to achieve this.
  • Procrustes Analysis: Because MDS reconstructions inherently have arbitrary rotations and reflections, the authors apply Procrustes analysis to align different "frames" of data to a consistent global orientation.
  • Refinement: The initial MDS estimate is further polished using a gradient descent approach (MDS-GD) to minimize "squared stress," effectively denoising the measurements.

Key Insights

  • Distributional Invariance: The authors exploit the fact that for many randomized ML tasks, the exact values of $A$ don't matter as long as $A$ follows a known distribution (like a Gaussian iid matrix). Because changing the phase of a row in a Gaussian matrix doesn't change its overall distribution, the system only needs to recover the relative phase between frames, not the absolute true phase.
  • Phase Retrieval as Localization: Instead of relying on complex, expensive wave-front sensing hardware, the authors turned phase retrieval into a "GPS" problem. The known anchors act like satellites, localizing the unknown signal in the complex plane.
  • The "Aha!" Moment on Denoising: Surprisingly, the algorithm doesn't just recover the missing phase; it actually improves the magnitude measurements. By taking multiple "views" of the signal relative to different anchors, the system effectively averages out quantization noise and overcomes the sensitivity thresholds of the 8-bit camera sensor.

Results and Applications

The paper validates this approach using numerical simulations, real hardware experiments on a physical OPU via a cloud interface (LightOn), and practical ML tasks.

Improvement in measurement precision (average good bits) as the number of anchors increases.
  • Magnitude Denoising: In hardware tests, the proposed method gained approximately 1 additional "good bit" of precision compared to raw sensor measurements. This held true even when many raw measurements fell below the camera's sensitivity threshold.
  • Linearity Recovery: The "linearity error" (a measure of how well the phase was recovered) dropped significantly as more anchors were added, proving the system successfully mimics a true linear operator.
  • Optical RSVD: To demonstrate a practical application, the authors successfully performed a Randomized Singular Value Decomposition (RSVD) on 500 images from the MNIST dataset using the OPU. The resulting singular vectors were virtually indistinguishable from those calculated using standard digital methods (like Python/NumPy), proving the recovered phase is accurate enough for downstream ML tasks.

Limitations and Future Work

While groundbreaking, the method does come with a throughput trade-off. To recover the phase, multiple reference signals (anchors) must be sent for every signal of interest, which reduces the effective data processing speed of the OPU. Additionally, the phase is recovered up to a global row-wise phase and conjugation; while fine for randomized algorithms, this global ambiguity might limit applications requiring absolute phase.

Future work will likely focus on optimizing the "duty cycle"—finding the absolute minimum number of anchors needed to maintain accuracy while maximizing the raw speed of the optical processor.

Summary updated

Abstract

In this paper we tackle the problem of recovering the phase of complex linear measurements when only magnitude information is available and we control the input. We are motivated by the recent development of dedicated optics-based hardware for rapid random projections which leverages the propagation of light in random media. A signal of interest $\mathbf{\xi} \in \mathbb{R}^N$ is mixed by a random scattering medium to compute the projection $\mathbf{y} = \mathbf{A} \mathbf{\xi}$, with $\mathbf{A} \in \mathbb{C}^{M \times N}$ being a realization of a standard complex Gaussian iid random matrix. Such optics-based matrix multiplications can be much faster and energy-efficient than their CPU or GPU counterparts, yet two difficulties must be resolved: only the intensity ${|\mathbf{y}|}^2$ can be recorded by the camera, and the transmission matrix $\mathbf{A}$ is unknown. We show that even without knowing $\mathbf{A}$, we can recover the unknown phase of $\mathbf{y}$ for some equivalent transmission matrix with the same distribution as $\mathbf{A}$. Our method is based on two observations: first, conjugating or changing the phase of any row of $\mathbf{A}$ does not change its distribution; and second, since we control the input we can interfere $\mathbf{\xi}$ with arbitrary reference signals. We show how to leverage these observations to cast the measurement phase retrieval problem as a Euclidean distance geometry problem. We demonstrate appealing properties of the proposed algorithm in both numerical simulations and real hardware experiments. Not only does our algorithm accurately recover the missing phase, but it mitigates the effects of quantization and the sensitivity threshold, thus improving the measured magnitudes.
BibTeX
@inproceedings{gupta2019don,
  author = {Sidharth Gupta and Rémi Gribonval and Laurent Daudet and Ivan Dokmanic},
  title = {{Don't take it lightly: Phasing optical random projections with unknown operators}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=Vv6GMK4qaDQ},
}
Figure