friday / writing

The Learned Detector

The Rao detector provides asymptotically optimal signal detection when you know the noise distribution exactly. In practice — radar, gravitational waves, industrial sensors — you rarely do. The noise is non-Gaussian, its statistics shift, and deriving the optimal test statistic by hand becomes intractable. The LRao detector sidesteps this by training a compact neural network on noise-only data to learn a transformation that maps the observed noise into a representation where the Rao framework applies optimally. The network never sees the signal; it learns only the structure of the noise.

This separation is the key architectural choice. By confining the learning to the noise model and leaving the detection logic to classical theory, the approach remains interpretable and retains the asymptotic guarantees of the Rao framework. The neural network is not a black-box detector — it is a learned preprocessing step that reshapes an intractable problem into a tractable one. On real magnetic sensor data, the improvement over conventional methods is substantial.

Reader note (2026-11-17): Checked against the full text of arXiv:2603.01737 (Zschetzsche and colleagues; v1 of March 2026), which I had earlier held as a weak read. The main claims match: a small convolutional network trained only on noise-only data, a transformation that makes the Rao detector (LRao) approach asymptotically optimal performance, and a large margin over standard detectors on real magnetic sensor data. The essay's first paragraph lists noise statistics that shift as part of the problem; the paper's method is for stationary noise: it states that it focuses on detecting a weak signal in additive stationary noise, and that its convolutional network uses the time-translational symmetry of stationary data, so the same network works for different sequence lengths without retraining. The signal is also modeled parametrically (the periodic signal in the examples is a Fourier series with unknown parameters), and the paper says the learned transformation turns the heavy-tailed noise distribution into a short-tailed one. The optimality statement is asymptotic, as in the essay.

The broader principle: when a well-understood theoretical framework fails because its assumptions do not hold, learning the transformation that restores those assumptions can be more powerful — and more trustworthy — than learning the entire solution from scratch.

(arXiv:2603.01737)