Optical Random Projections through Disordered Scattering Media

75 concepts48 papers1 proposal

In modern machine learning and numerical linear algebra, multiplying a high-dimensional vector by a massive random matrix is a ubiquitous but expensive operation. As datasets grow, digital hardware faces a severe bottleneck: storing a matrix with millions of rows and columns requires vast amounts of memory, and computing the product incurs a quadratic computational cost.

To bypass these constraints, a growing body of research explores a radical alternative: offloading matrix multiplication to the physical propagation of light through disordered scattering media. By replacing digital logic gates with the natural, deterministic random walk of photons, this paradigm achieves high-dimensional random projections in near-constant time with minimal energy consumption.

The Intuition: Scattering as a Matrix Multiplication

To understand why this physical process computes a random projection, consider how light behaves when it hits a disordered medium, such as a layer of white paint or a slab of nanoparticles. When a coherent laser beam is spatially modulated to encode an input vector $x$, the light waves enter the medium and undergo multiple scattering events.

As demonstrated by Saade et al. in their work on Random Projections through multiple optical scattering: Approximating kernels at the speed of light, this scattering is not purely chaotic noise; it is a deterministic, linear physical process. The complex interference pattern—or speckle pattern—that emerges on the other side of the medium is a superposition of the scattered waves. Mathematically, the relation between the input optical field and the output field can be modeled as a multiplication by a fixed, high-dimensional random matrix whose entries are determined by the microscopic structure of the medium.

Crucially, because the weights of the transformation are implicitly stored in the physical arrangement of the scattering particles, there is no need to generate or store a digital matrix in memory. As Krzakala et al. note in Artificial intelligence: From electronics to optics, this approach exploits the inherent physical complexity of light propagation to achieve massive computational throughput, effectively performing the core linear algebra operation at the speed of light.

Bypassing the Memory Wall in Constant Time

In traditional digital architectures, scaling up a neural network or a matrix operation means facing the "memory wall"—the energy and time cost of moving data between the processor and memory. For a matrix of size $M \times N$, digital computation scales as $O(MN)$.

Optical scattering systems fundamentally upend this scaling law. The principle of constant-time physical projection, detailed by Cappelli et al. in Photonic co-processors in HPC: using LightOn OPUs for Randomized Numerical Linear Algebra, leverages the fact that the time required for light to traverse a physical medium is fixed. Whether the input vector has ten dimensions or one million dimensions, the physical propagation time remains identical.

This property allows hardware accelerators, such as the Optical Processing Units (OPUs) described by Cappelli et al. in LightOn Optical Processing Unit : Scaling-up AI and HPC with a Non von Neumann co-processor, to scale to massive dimensionalities without running out of VRAM or encountering memory bandwidth bottlenecks. The dense all-to-all connectivity comes as a natural physical default, turning what is a worst-case scenario for digital chips into a zero-cost operation for optics.

Scaling Up Randomized Algorithms

The ability to perform fast, passive random projections has direct utility across several domains of artificial intelligence and high-performance computing:

Reservoir Computing and Echo-State Networks

In reservoir computing, an input signal is mapped into a high-dimensional dynamical space (the reservoir) via a fixed random matrix, and only a simple linear readout layer is trained. Dong et al. demonstrated in Scaling up Echo-State Networks with multiple light scattering that physical scattering can scale these networks to 100,000 neurons, achieving orders-of-magnitude speedups over standard digital servers. To handle spatiotemporal chaotic systems, Rafayelyan et al. and Dong et al. showed in Large-Scale Optical Reservoir Computing for Spatiotemporal Chaotic Systems Prediction that combining spatial light modulators with strongly scattering media allows for the processing of massive reservoir states with minimal power overhead.

Randomized Numerical Linear Algebra (RandNLA)

RandNLA algorithms rely on random matrices to compress large datasets while preserving their geometric properties (e.g., via the Johnson-Lindenstrauss lemma). The authors of Linear Optical Random Projections Without Holography, Ohana et al., proved that physical scattering media naturally produce high-dimensional Gaussian random matrices that perfectly match theoretical benchmarks like the Marchenko-Pastur law. By recovering the linear projections without complex holographic phase measurements, they enabled robust dimensionality reduction on million-dimensional vectors.

Real-Time Streaming and Data Compression

Because the projection happens instantaneously, this mechanism is highly suited for edge computing and streaming applications. For instance, Keriven et al. utilized optical processing units in NEWMA: a new method for scalable model-free online change-point detection to compute high-dimensional random features in near-constant time, allowing for real-time anomaly detection in massive data streams. Similarly, Redding et al. developed an Integrated photonic encoder for low power and high-speed image processing that compresses high-speed image data in the optical domain before it even reaches a digital sensor, drastically reducing the downstream energy footprint.

Introducing Nonlinearity

While linear random projections are powerful, deep learning requires non-linear transformations. In standard optical setups, a square-law nonlinearity is naturally introduced when a camera or photodiode measures the intensity of the light (which corresponds to the squared magnitude of the complex field).

However, researchers are also embedding nonlinearity directly into the scattering medium itself. As shown by Wang et al. in Large-scale photonic computing with nonlinear disordered media, using a disordered slab of lithium niobate nanocrystals allows for the combination of multiple light scattering with second-harmonic generation. This architecture provides a built-in, physical optical nonlinearity during propagation, mapping inputs into a complex latent space that significantly enhances performance on downstream classification and graph analysis tasks.

By turning material disorder from a hindrance into a computational asset, optical random projection offers a compelling path forward: a hardware paradigm where the complexity of the physical world directly solves the complexity of high-dimensional data.

Key Foundational Work

An Explicit Universal Gate-Set for Exchange-Only Quantum Computation (2003) — contributed 2 concept elements to this direction.
Coded aperture computed tomography (2009) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_optical_random_projection_a1dededb,
  title = {optical random projection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optical-random-projection/rdr_a1dededbec36428e8fbfd8628b20b4df},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_optical_random_projection_a1dededb_v1,
  title = {optical random projection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optical-random-projection/rdr_a1dededbec36428e8fbfd8628b20b4df/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake