This paper investigates how to efficiently compress and represent the complex "Transmission Matrices" (TMs) of multimode optical fibers. By benchmarking a simulation-pretrained autoencoder against standard Principal Component Analysis (PCA), the authors find that while PCA is a superior compressor when data is abundant, the autoencoder acts as a powerful "prior" that excels when real-world measurements are scarce and provides significantly better robustness to experimental drift.
TL;DR
Researchers compared deep learning autoencoders and PCA for compressing the massive data needed to image through flexible multimode fibers. While PCA is more accurate given enough data, the autoencoder can reconstruct fiber states from just two measurements and remains accurate even when the camera shifts or zooms—conditions that completely break traditional linear models.
What Problem Does This Paper Address?
Multimode fibers (MMFs) are the "holy grail" for ultra-thin endoscopes because they can guide thousands of light modes through a core thinner than a human hair. However, they have a major flaw: they are incredibly sensitive. If the fiber bends, moves, or even changes temperature slightly, the light scrambles into a random-looking speckle pattern.
To "unscramble" this light and see through the fiber, researchers measure a Transmission Matrix (TM)—essentially a giant lookup table that describes how every input pixel maps to every output speckle. These matrices are huge (often ~50 MB each) and must be re-measured every time the fiber moves. This creates a massive data bottleneck for storage and real-time imaging. The authors ask: how compactly can we represent these matrices, and can we use physics-based simulations to help us estimate them when we only have a few real-world measurements?
Methods
The authors benchmarked two primary approaches for 62× compression:
- Principal Component Analysis (PCA): A standard linear baseline that finds a shared subspace for the speckle patterns.
- Permutation-Equivariant Autoencoder (AE): A custom deep learning architecture designed to process each column of the TM (each speckle pattern) independently. This ensures the model doesn't care about the order of input patterns.
The autoencoder uses a "Linformer-style" channel self-attention mechanism to capture non-local correlations in the speckle patterns. Crucially, the AE is pretrained on simulated data generated via the Transfer Matrix Method (TMM). To save on the high computational cost of simulating long fibers, the authors used "composite" simulations—multiplying five 1-meter simulated segments to approximate a 5-meter experimental fiber.
Datasets
- Simulated Data: 5,000 1m TMs and 2,000 "composite" 5m TMs.
- Primary Experimental Data: 204 TMs collected from a 5m MMF with a 50 $\mu$m core.
- Transfer Dataset: 164 TMs from a larger 105 $\mu$m core fiber to test how well a model trained on one fiber generalizes to another.
Key Insights and Ideas
- The "Prior" Power: The autoencoder’s real value isn't as a better compressor, but as a better prior. In a "zero-shot" scenario (no real data), the AE achieves 29% error compared to PCA’s 91%. This suggests the AE learns the underlying "physics" of speckle patterns better than a linear model.
- Robustness to Drift: This is the "aha" moment. Linear bases like PCA are extremely fragile; if the camera shifts by just a few pixels or zooms slightly, the PCA reconstruction fails. The autoencoder, being non-linear and spatially aware, maintains high accuracy under these common experimental perturbations.
- The Data Crossover: There is a clear tipping point. If you have more than 10 measured TMs, PCA’s simplicity and linear accuracy win. If you have fewer than 5, the simulation-pretrained autoencoder is the only viable option.
- Simulation-to-Experiment Gap: Even with "composite" simulations, there is a significant gap between simulated and real data. Pretraining helps bridge this gap, but fine-tuning on at least a few real TMs is essential for high-fidelity imaging.
Key Results
- Compression: At 62× compression, the AE adapted with just two real TMs could focus light at 98% of the ideal enhancement.
- Accuracy: With a full dataset (163 TMs), PCA achieved a tiny 1.1% error, while the AE stayed at 11.9%.
- Geometric Robustness: When TMs were shifted by 25 pixels, PCA error spiked to nearly 30%, while the AE error barely moved, staying around 16%.
- Transferability: Transferring a model from a 50 $\mu$m fiber to a 105 $\mu$m fiber was only partially successful, suggesting that core size and mode count significantly change the "latent space" of the fiber.
Limitations and Future Work
The authors acknowledge that they haven't yet tested the most exciting application: recovering TMs from partial measurements. Currently, they compress a full measured TM. The next step is to use the autoencoder as a prior to reconstruct a TM from a fraction of the usual measurements, which would directly translate to faster imaging speeds. Additionally, the partial success in transferring between different fiber types suggests that a "universal" fiber model may require training on a much wider variety of simulated fiber geometries.
Summary updated