When 90 % of a signal is absent, reconstruction becomes a problem of structural inference. Short tokenization, 4D positional encoding and edge inference make the approach viable.
Some sensors deliver only a fraction of the expected signal. Sparse measurements, gapped windows, rare samples. Reconstructing a signal with 90 % missing is not denoising. It is inferring the underlying structure from very few observed points.
Tokenize short, encode space
We split the signal into short tokens. Brief segments limit drift and keep fine granularity over the rare regions where each sample counts. The model learns reusable local patterns rather than one long, fragile sequence.
Positional encoding is four-dimensional. A sample's position is not just a time index. It carries its place in space and in time. This 4D frame lets the foundation model reason about the real geometry of the acquisition, not a mere rank.
The edge argument
The model has to run close to the sensor. Quantization reduces the weights to short integers with no notable loss on this task. The memory footprint drops and CPU inference becomes possible, with no dedicated accelerator.
This changes the economics of deployment. No embedded GPU, no network dependency for each inference. The principle to keep. A well-tokenized, well-quantized foundation model reconstructs a heavily incomplete signal right next to the measurement.