A research program on neural joint source-channel coding — radio links where the transmitter and receiver share a learned model, so the signal degrades gracefully and keeps delivering in exactly the conditions that make classical links fail completely: deep-space low-SNR, fading, and jamming.
Honest by design: this page leads with what the technology does and does not do. Every number below is measured in our own simulations, reproducible from the open code.
Today's systems separate the job: compress the data to bits, then wrap those bits in error-correction. Claude Shannon proved that's optimal for infinitely long messages — but it has a hard edge. Above a channel-quality threshold you get a perfect signal; a hair below it, the code fails and you get nothing. That "cliff" is worst exactly where it matters most: at the edge of deep space, through fading, under a jammer.
When both ends share a jointly-trained model, they transmit meaning rather than fragile bits. The result has no cliff: as the channel worsens the reconstruction just gets softer, and every frame still arrives usable. That is the whole thesis — and the sections below show where it holds up and where it doesn't.
The single reason semantic communication isn't deployed: two independently-built endpoints learn private codes and can't understand each other (link collapses to noise even on a perfect channel). Rosetta makes three independently-trained radios fully interoperable — 9.2 → 22.2 dB cross-vendor, every sender-to-receiver pair — via a tiny standardized adapter, with zero changes to the base models. This is the piece a standard (Next G Alliance / 3GPP) would publish.
And the underlying link performance — trained from scratch, tested on standard channel models and real Sentinel-2 satellite imagery against a deliberately strong classical baseline:
On a one-way link (deep-space, broadcast, jammed) digital must fix its rate and delivers 0% of frames below its design point. The neural link needs zero transmitter channel knowledge and still delivers a usable image — +8.8 dB at 0 dB SNR.
verified · Sentinel-2Under Rayleigh fading + barrage jamming, the classical baseline loses ~19% of frames entirely; the learned link delivers 100% at consistent quality, and a hybrid design stays within ~1.3 dB of the best classical system across the whole SNR range without ever cliffing.
verified · hybridA reactive follower-jammer cuts an unhardened link by 6.3 dB; adversarial training recovers 3.0 dB and the robustness transfers to the worst-case jammer, for <1 dB clean-channel cost.
verified · adversarial| Regime | Why |
|---|---|
| Deep-space / low SNR | graceful, no cliff, no outage |
| Contested / jammed (EW) | hardens against reactive jammers |
| Fading, no retransmit | delivers every frame usably |
| Bandwidth-starved links | ~12× source compression built in |
| Regime | Why |
|---|---|
| High-SNR / clean links | rate-adaptive digital wins — hand off |
| Exact / lossless data | neural decoders can hallucinate |
| Deployed RF hardware | not yet tested — simulation only |
| Interoperability | both ends need the same model |
Seven studies, each with a full report and open source. They include the failures: an architecture that partly worked, and an adaptivity idea that didn't. That's the point — this is a real research trail, not a brochure.
Makes independently-trained transceivers understand each other through a shared reference standard + tiny frozen-base adapters. Cross-vendor 9.2 → 22.2 dB, all pairs. The path from lab to standard.
The core demo: a neural transceiver that beats classical coding across every SNR and degrades gracefully. Interactive — pick or draw an image and dial the noise.
Two builds: a hybrid digital-analog code that recovers ~half the good-channel gap without ever cliffing, and the feedback-free moat — usable data on one-way links where digital delivers nothing.
Comms-grade re-test on real Sentinel-2 imagery: Rayleigh/Rician fading, partial-band jammer, honest outage-based baseline. Wins in the hard regime; delivers every frame.
Reactive/adversarial jamming + adversarial-training defense (worked); SNR-rate-adaptivity (didn't close the gap — reported straight).
A message code that fades instead of breaking — the same principle applied to a QR-like visual channel. 93% bit-recovery at 70% damage.
Two agents invent a private, compressed, noise-robust language to cooperate — cooperation on ~8 numbers per message instead of verbose text.
The initial space/defense framing: shared-model transceiver under burst jamming and fades. Motivated the rigorous v2 re-test.
A from-scratch novel sequence architecture (resonant memory + selective gate). Beats a GRU on noisy frequency-ID; loses on delayed recall — ablation-validated.
This is a proof-of-concept with a genuine, narrow advantage and honest gaps. The next steps need resources a laptop can't provide — and that's exactly the conversation we're looking for.
Re-run the studies against your own channel models, datasets, and classical baselines. Everything is open and reproducible.
Take it from simulation to an SDR / RF testbed — real Doppler, timing recovery, and front-end nonlinearity — to measure true TRL.
Train and test on representative payloads (hyperspectral, SAR, telemetry) and real link budgets for a target mission profile.
Independent evaluation, testbed access, or a mission-specific study.
priyankar@pchakconsulting.comWebsite crafted by Everbuilt Studio