Neural networks learn hidden parameters from recovered signals, not labels

Some of the most stubborn problems in modern engineering share a quiet, frustrating structure. A wireless receiver trying to decode a signal through a moving city, or an MRI machine attempting to reconstruct an image from an undersampled scan, must first estimate physical quantities that nobody can directly measure. Channel responses drift, carrier frequencies slip, and the sensitivity of a coil is invisible on its own. Standard practice is to train a neural network to guess these hidden parameters, but that approach demands labels for quantities that are often unreliable, expensive, or simply unavailable. A team of researchers from the National University of Defense Technology and the Academy of Military Science has now demonstrated a different route, one in which a neural network learns to estimate unobservable parameters without ever being shown a single label for them.

The method, called physics-embedded inverse learning, or PEIL, is described in a study published in National Science Review. Its central idea is deceptively simple: rather than supervising the network on the parameters themselves, supervise it on the quality of the final signal that those parameters help produce. PEIL keeps the familiar estimate-then-solve architecture of physical inverse problems. A neural network first predicts the parameters required by a physical solver, such as the carrier-frequency offset and channel state information in a wireless link. The solver, which is fixed and never trained, then uses those estimates to recover the target signal. Because the solver is differentiable, the error between the recovered signal and a reference signal can be propagated backward through the entire physics-based pipeline, updating only the parameter-estimation network.

This arrangement changes what the network is actually learning to do. In conventional supervised training, the network is asked to reproduce the true value of a parameter as closely as possible. In PEIL, the network is asked a subtler question: which parameter estimates, when fed into the fixed solver, produce the best possible signal recovery? The distinction matters because the two objectives are not always identical, and the experiments revealed exactly why. When the parameters themselves are the quantities of scientific interest, their physical accuracy remains essential, and nothing in the new method argues otherwise. But when parameters are merely intermediate machinery in service of a final task, the task itself can provide the training signal.

The team put this principle to a demanding test in a high-mobility wireless communication scenario. From pilot observations alone, the network predicted the carrier-frequency offset and the channel state information, while a fixed receiver used those predictions to recover the transmitted symbols. Training relied exclusively on symbol-recovery error, with no carrier-frequency offset or channel labels provided at any point. Even under this label-free regime, PEIL learned to produce physically meaningful estimates of both quantities and achieved strong symbol recovery. At signal-to-noise ratios above 30 decibels, the system actually achieved a lower symbol error rate than a least-squares baseline that had been handed the exact carrier-frequency offset and ground-truth channel responses at the pilot positions, with the same fixed interpolation applied in both cases.

The generalization results were perhaps the most striking part of the wireless experiments. After training on a single channel profile at 120 kilometers per hour, PEIL outperformed parameter-supervised baselines on channel profiles it had never seen and at speeds reaching 360 kilometers per hour, all without any retraining. The sample efficiency told a similar story: PEIL reached comparable recovery performance using only one twentieth as many training samples as the parameter-supervised baseline. For anyone who has watched a communication system degrade the moment real-world conditions drift away from the training distribution, this robustness across unseen profiles and speeds is the kind of result that draws attention.

The explanation lies in how the network allocates its effort. The experiments showed that more accurate estimates of individual parameters do not always translate into better signal recovery. Within a limited range, a small deviation in the carrier-frequency offset can be offset by a corresponding adjustment in the phase of the channel state information, so that different parameter combinations recover nearly identical symbols. PEIL learned coordinated estimates of both quantities that jointly supported recovery by the fixed receiver. The researchers describe these as task-optimal parameters, selected according to final recovery performance rather than fidelity to some ground truth. It is a pragmatic reframing: the network is not chasing an abstract ideal of parameter accuracy but is instead tuning itself to the demands of the downstream physics.

To establish that the principle was not an artifact of wireless systems, the team applied PEIL to parallel magnetic resonance imaging, a domain with entirely different physics but the same underlying structure. In parallel MRI, reconstructing an image from undersampled multi-coil data requires knowledge of coil sensitivity maps, spatial patterns that describe how each coil responds to the signal and that cannot be observed directly without calibration. The network predicted these sensitivity maps from the undersampled data, and a fixed physical solver used them to reconstruct images. Remarkably, without sensitivity-map labels and without any separate calibration step, PEIL estimated stable coil sensitivity maps and produced anatomically faithful reconstructions under both fourfold and eightfold undersampling.

The MRI implementation also introduced a refinement: the network provided an image-domain reference to the fixed solver, and the final physics-based reconstruction was assessed relative to this reference. The result was a substantial improvement, with peak signal-to-noise ratio gains of 5.5 to 6.1 decibels over the reference. That a label-free, physics-anchored procedure could deliver gains of that magnitude in a clinical imaging context suggests the approach is more than a communications curiosity. It indicates a genuinely transferable design pattern for inverse problems across sensing technologies.

What unites the wireless and imaging applications is a shift in the basis of supervision. Learning is guided by final task performance rather than by ground-truth labels for intermediate parameters. Whenever the hidden parameters feed a differentiable physical solver, and the solver’s output can be evaluated against some reference, the question of whether the result is correct can itself provide the training signal for parameter estimation. This observation opens a path for the many inverse problems in sensing, communications, and computational imaging where labels for the true parameters are scarce, noisy, or nonexistent, while reliable references for the final signal are comparatively easy to obtain.

The implications extend beyond the two demonstrated applications. Systems that recover signals from incomplete or indirect observations are pervasive, from radar and acoustic sensing to astrophysical imaging, and each depends on physical parameters that resist direct measurement. By making the physics differentiable and the supervision task-based, PEIL turns a long-standing bottleneck into an optimization opportunity. The study, published in National Science Review with the identifier 10.1093/nsr/nwag508, was based on computational simulation and modeling, and its authors argue that the principle may extend broadly across domains where physical solvers and learnable estimators can be coupled. If the wireless and MRI results are any indication, the era of needing perfect labels for invisible quantities may be drawing to a close, replaced by a simpler demand: show the network what a good final answer looks like, and let the physics do the teaching.

Subject of Research: Label-free estimation of unobservable physical parameters via differentiable signal recovery in wireless communications and parallel MRI

Article Title: Signal recovery teaches neural networks to estimate unobservable parameters without their labels

Article References: Signal recovery teaches neural networks to estimate unobservable parameters without their labels. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: physics-embedded inverse learning, PEIL, neural networks, inverse problems, signal recovery, carrier-frequency offset, channel state information, parallel MRI, coil sensitivity maps, wireless communications, label-free learning, differentiable physics