A working demo of AI-native (semantic) communication: two neural models —
a sender's encoder and a receiver's decoder — trained end-to-end to push images through a noisy
radio channel. No pretrained weights, no hand-designed codec. Trained on Fashion-MNIST, MPS, this session.
The idea. Classical digital comms separate the problem: compress to bits (source
coding), then protect those bits (channel coding). Shannon proved that's optimal — but only for infinitely long
messages. For real images over a real channel, a jointly-learned analog code that both
ends share can do better, and crucially it degrades gracefully instead of falling off the
digital "cliff". The encoder and decoder ARE the shared model; what they transmit is closer to meaning than to bits.
Result — graceful vs. the cliff
+4.7 dBLatent Radio advantage at -6 dB (bad channel)
·
+2.6 dBadvantage at 14 dB (good channel)
Latent Radio beats the Shannon-optimal digital baseline at every channel quality, and the gap
widens as the channel worsens — at -6 dB the digital system has collapsed to the blank-image floor
(10.7 dB) while Latent Radio still delivers a recognizable image.
Latent Radio (AI-native)Classical digital (Shannon-optimal)
See it — same images, both systems, three channel conditions
Row 1 — original. Rows 2–4 — Latent Radio at -4, 2, 10 dB.
Rows 5–7 — classical digital at the same SNRs. At -4 dB the digital reconstructions are
unrecognizable smear; Latent Radio's are blurred but clearly the right garment.
Honest failure modes
AI-native communication buys graceful degradation, but it has a hard requirement and a real fragility.
Both are shown here rather than hidden.
Latent Radio (AI-native)Failure: mismatched modelsFailure: out-of-distribution
1 · Both ends must share the exact same model
Pair a sender's encoder with a receiver's decoder from a separately trained model — each excellent on its
own — and the link produces garbage at every SNR, even a perfect channel (red, flat ~9 dB, below the
blank-image floor). There is no interoperability standard here the way there is for, say, Wi-Fi: the "protocol"
is the shared weights.
Row 1 — original. Row 2 — matched encoder+decoder (same model). Row 3 — encoder from model A,
decoder from model B, on a clean channel. Both models are individually good; together they are meaningless.
2 · It leans on knowing the source distribution
Train on clothing, then transmit handwritten digits (green, dashed): quality drops well below the
in-distribution curve (to 18.6 dB vs 22.9 dB at the best SNR). The shared
model encodes priors about what it expects to see; off-distribution inputs cost you. A production system would need
a model trained on the true source, or an explicit fallback.
What this is and isn't
Is: a real, reproducible demonstration that a shared learned model can transmit images over a noisy channel
more robustly than the classical separation architecture, with the signature graceful-degradation behavior — and an
honest accounting of the two failure modes that come with it.
Isn't: a new physical layer or new spectrum (physics is fixed). It's a new abstraction layer — the same
kind of move radio and the internet were. This is one bandwidth (k=128 real channel uses), one dataset, small
models. A real system would need larger models, real channel effects (fading, interference), and a shared-model
distribution/versioning story. The point is that the core mechanism works and behaves as theory predicts.
