The spectrum is an image
A spectrogram is a picture: time on one axis, frequency on the other, power as tone. Convolutional networks are machines that learned to read pictures. Point one at a spectrogram and it can learn a fingerprint the way it learns a glyph — including fingerprints that a hand-built library never named.
That is the classification half of the Mind & Signal vision. Super-resolution direction finding, including MUSIC-class angle-of-arrival on a coherent array, produces the geometry. The network points that array at the threat in a crowded band. This team has built both halves before — beamforming, software-defined radio, and direction finding — and is putting them on one sensor.
Train and retrain as the waveform moves. Keep the model small enough to live next to the radio. MobileNet-class networks get it onto the handheld and the vehicle; an FPGA or edge engine is the home as the node multiplies. Compute stays about the size and cost of the radio, which is how the sensor leaves the tripod.