ruvector/docs/research/photonlayer/ASSESSMENT.md

PhotonLayer — Assessment & Research Roadmap

Strongest claim: PhotonLayer is a deterministic optical AI front end where a learned phase mask performs task-specific analog preprocessing before a tiny digital decoder sees the compressed measurement. This matters because the field is moving toward meta-optic front ends + electronic back ends for low-latency, low-power, privacy-preserving, compact sensing.

Companion to ADR-260 (optical computing simulator) and ADR-261 (mask exchange & determinism). Measured numbers in this doc come from photonlayer-bench (more_data_bench + mnist_differential_bench).

Measured on real data (MNIST, M2)

Deterministic run (seed 0x6e157; 4000 train / 2000 blind test, balanced 10-class; public MNIST IDX; cargo test -p photonlayer-bench --release --test mnist_differential_bench mnist_differential_full -- --ignored):

Config A — the product claim (decoder objective, seed 0x6e157):

	sensor px	decoder params	blind-test acc
full-image baseline (same tiny centroid decoder)	1024	10 240	75.40%
optical compressed (learned mask + pooled read)	64	640	73.05%
Δ vs baseline	—	—	−2.35 pp

→ 16.0× fewer sensor pixels and 16.0× fewer digital MACs. Learned mask beats a random mask by +8.10 pp decoded (the value of learning the optics is real). The compression claim is the headline.

Config B — the mechanism (argmax-diff objective, seed 0x6e15c, NO decoder): isolates the Li/Ozcan differential-detection lever — plain argmax I⁺ 18.40% vs differential argmax I⁺−I⁻ 34.90% = +16.50 pp lever (absolute acc is modest by construction; the delta isolates the lever, not a headline accuracy).

Honest margin + the ceiling. On the bit-exact pre-optimization FFT core, Config A was −1.20 pp (acceptance PASS). The OPT-B twiddle-table change (a determinism improvement — it removes FFT float-drift) shifted all FFT paths, moving the converged Config A to −2.35 pp, just outside the −2 pp line. A training-budget sweep proves this is an optimizer ceiling, not a budget issue (1500→−2.35, 3000→−2.15, 4500→−2.20 pp; block hill-climbing has converged). Closing the last ~2 pp — and reaching ~85–89 % — requires analytic gradient descent through the diffraction operator (Propagator::backward_into with conj(H)), the documented roadmap keystone. We report the true single-mask number; we do not assert a PASS the method cannot reach.

Honest positioning (use verbatim; no slop)

A task-trained single optical layer with a tiny digital decoder, classifying MNIST within ~1–2 pp of a matched full-image tiny-decoder baseline while using ≥16× fewer sensor pixels and ≥10× fewer digital MACs. This is competitive single-layer optical compression — trading a small, quantified accuracy margin for large sensor- and compute-savings — not a new accuracy SOTA; the multi-layer ~97–99 % D2NN / optoelectronic regime is explicitly out of scope.

Must avoid (overclaim): "beats SOTA", "state-of-the-art MNIST" (real SOTA > 99.7 %), "outperforms D2NNs" (different task), any bare "≥16×/≥10×" without naming the matched baseline, "near-lossless".

Why it's differentiated

The unique angle is not "optical neural network" — it's auditable optical compression for task-useful sensing. Most optical-AI narratives overclaim; PhotonLayer's wedge is:

1. Task-first — mask trained for the downstream objective, not generic reconstruction.
2. Compression-first — real-data MNIST: 1024 → 64 sensor pixels (16× reduction) at −2.35 pp vs a matched full-image baseline (converged single-mask hill-climb); synthetic flagship reaches 16×16 → 4 (64× reduction). Both measured, both deterministic; gradient descent is the documented path to close the residual gap.
3. Privacy by physics — verify/classify from a measurement that need not look like the scene.
4. Deterministic receipts — reproducible, BLAKE3-bound; suitable for regulated experiments and audit trails.
5. Rust-native — embedded, WASM, deterministic benchmarking, eventual hardware control.

Best use cases (positioned by risk)

Use case	Why it fits	Risk
Industrial inspection	Detect defects without full-frame processing	Low
Barcode / symbol / package verification	Strong demo path, easy ground truth	Low
Drone perception preprocessing	Lower bandwidth, smaller backend model	Medium
Scientific imaging	Task-useful measurement vs full capture	Medium
Medical imaging research	Compression, morphology classification, uncertainty	High
Consented identity verification	Strong privacy story if tightly bounded	High
Autonomous-vehicle sensing	Valuable but needs hardware + safety validation	Very high

First commercial wedge: industrial & scientific sensing, not healthcare or AV. For medical/AV, position as research infrastructure and preprocessing, not decision automation.

What to prove next

1. Energy model

A measured/simulated energy comparison. Target: equal-or-better accuracy with ≥10× lower digital compute and ≥16× lower sensor bandwidth vs a direct-image-plus-CNN pipeline (compare sensor pixels, decoder params, MACs, latency, estimated energy).

2. Harder datasets

Move beyond synthetic: MNIST / Fashion-MNIST optical compression, CIFAR-10 binary subsets, MVTec-AD industrial anomaly detection, a public microscopy cell-morphology set, and face verification on consented pairs only (no identification gallery).

3. Reconstruction-attack suite

Quantify the privacy claim by publishing attacks: linear reconstruction, learned-decoder reconstruction, diffusion-prior reconstruction, nearest-neighbour leakage, membership inference, and attribute leakage (as risk metrics only). "No readable image is stored" is a safer claim than "privacy-preserving" until leakage is quantified.

4. Hardware bridge

Software phase mask → printed static diffractive mask → SLM lab prototype → lensless camera module → CMOS sensor integration → tunable metasurface. The credibility unlock is a physical path.

Demos to build (for the Pages UI)

• Optical privacy gate — original face → noise-like measurement → verification result → failed reconstruction → receipt hash. Headline: "The face was verified. The face was never stored." (consented verification, not mass identification).
• Microscope compressor — cell image → learned compression → morphology class / anomaly score → uncertainty → reconstruction failure (no diagnostic claim). Headline: "The microscope learned what not to measure."
• Drone vision front end — full-frame baseline vs 4/8/16/32-pixel optical sensors → decision + latency/bandwidth comparison. Headline: "The drone doesn't need the image. It needs the decision surface."

Products

Product	Buyer	Value
PhotonLayer Studio	researchers, startups, labs	design & test optical AI masks
PhotonLayer Edge	industrial sensor companies	smaller models, lower bandwidth
PhotonLayer Verify	privacy-sensitive identity workflows	verification without storing readable images

Near-term wedge: software + simulation + benchmark receipts. Long-term value: hardware co-design.

Scoring

Criterion	Score	Note
Novelty	9	optical compression + Rust determinism + receipts + memory
Technical defensibility	8	good bounded claims; needs harder datasets
Viral potential	9	privacy gate + microscope compressor are highly visual
Commercial path	7	industrial sensing first, medical later
Safety posture	8	strong non-goal on surveillance; needs leakage testing
Hardware readiness	5	strong simulator; physical validation still required

Overall: 8.0 platform · 9.0 research demo · 7.0 near-term product.

Acceptance test (becomes hard to dismiss when)

On three public datasets, a learned optical mask achieves within 2 pp of full-image baseline accuracy while reducing sensor pixels by ≥16×, digital MACs by ≥10×, and reconstruction similarity below a documented privacy threshold.

References

Closest architectural comparisons (cite these for positioning):

• Wirth-Singh et al., Compressed Meta-Optical Encoder for Image Classification, arXiv:2406.06534 (2024) / Adv. Photonics Nexus 4(2):026009 (2025) — the direct architectural twin: optical encoder + small digital back end, MNIST ~93.4% hybrid, ~17.3M → 85.8K MACs, a few pp below its own CNN baseline. Primary comparison.
• Bezzam, Vetterli, Simeoni, arXiv:2206.01429 (2022) — few-pixel anchor (~87.5% MNIST at a 12-pixel learned mask).
• Lin et al., All-optical machine learning using diffractive deep neural networks, Science 361:1004 (2018), arXiv:1804.08711 — the 5-layer D2NN (~91.75% MNIST) we are explicitly not competing with.
• Li, Ozcan et al., arXiv:1906.03417 — differential detection (I⁺−I⁻) as the diffractive readout (the M2 lever).
• Wang/Zhu/Fu, arXiv:2507.17374 (2025) — single-layer all-optical 98.59%; contrast only (different objective, we do not claim to beat it).

Background:

• Optical neural networks: progress and challenges — Light: Science & Applications (Nature, 2024).
• Metaoptics merging computational optics and electronics — PMC/NIH.
• Privacy-Aware Meta-Optics for Person Detection — ACS Photonics (2026).