Radar, sonar and seismic arrays locate sources by estimating how signals correlate across sensors, a quantity called the covariance matrix. When the scene changes quickly, only a few time samples ("snapshots") are available, often fewer than there are sensors. In that regime, called sample starvation, the empirical covariance is singular and classical direction finding fails. Deep networks trained offline can produce a usable substitute covariance. Their weakness is that they are tied to the array they were trained on. A deployed array with unmodeled sensor coupling, displaced elements or unfamiliar noise needs full retraining on tens of thousands of labeled examples.
Few-shot learning with large language models points to a different fix. A few concrete examples placed in the input often do not teach the model anything new. They calibrate knowledge the model already has to the task at hand. We propose to carry that mechanism over to arrays. A model is trained across thousands of simulated arrays and mismatch conditions. At deployment it receives, alongside the starved test window, one to five calibration snapshots from sources at known directions (pilot emitters, catalogued seismic events). From these it outputs a corrected covariance for the test window, with no weight update. The corrected covariance then feeds standard, interpretable subspace direction finding.
Core Claim
The claim is that a few labeled snapshots, used as calibration context, recover most of the accuracy lost to array mismatch when sample support is extreme, meaning fewer test snapshots than sensors. We further claim this holds specifically when the mismatch does not fit a simple parametric error model, such as compound mutual coupling, position error and colored noise. Under such mismatch, neither classical least-squares calibration nor generic covariance regularization can capture the distortion.
The claim includes a mechanism that can be checked. If the model is truly calibrating rather than just receiving more data, then scrambling the direction labels on the context snapshots should destroy most of the gain. Supplying the same snapshots without labels should help only about as much as pooling them into the covariance estimate.
What exists and what does not
Two hybrid methods supply the method template and motivate the gap. Deep Augmented MUSIC does three things: it replaces the sample covariance with a recurrent network's estimate, keeps a differentiable eigendecomposition, and learns both source counting and peak finding. DCD-MUSIC cascades two networks that produce surrogate covariances for near-field angle and range. It also handles array miscalibration, but only the miscalibration present in its training distribution.
Both methods are trained once, offline, on a matched array. Neither has any route by which a few target-domain examples can change its output. We keep their key design choice, a learned covariance followed by analytic eigendecomposition and a spectrum. What we change is the input: the network conditions on a labeled calibration context, and it is trained across a distribution of arrays rather than on a single one.
AutoMPDNN shows that single-snapshot direction finding can be learned as classification over 1° bins. It is restricted to one source and fixed simulated arrays, and it used 181,000 training samples per noise level. That figure is the retraining cost our approach aims to avoid.
The closest conceptual neighbor is Meta-AF, which meta-learns the update rule of audio adaptive filters with a small recurrent network shared across frequency bins. It shows that learned adaptation can beat hand-designed filter updates. However, Meta-AF adapts gradually from a continuous stream of frames. It does not condition on labeled exemplars and does not estimate covariance in the regime of fewer snapshots than sensors. Our setting has no stream to adapt on, only a handful of labeled snapshots.
On the few-shot side, the work we draw on motivates the calibration hypothesis rather than supplying a method. Yancey et al. found that one example per score level raised GPT-4 essay-scoring agreement from below a length baseline to QWK 0.81. Hu et al. found that a few annotated samples brought GPT-4 close to a fine-tuned clinical tagger. Spokoyny et al. found that concrete exemplars beat abstract persona instructions. That last result maps onto our central tension: are concrete calibration exemplars worth more than rigid structural priors under starvation?
Two papers represent those structural priors, and they constrain the design: persymmetry (Pascal et al.) and Riemannian shrinkage (Kalunga et al.). Both work well when their assumptions hold. Our estimator must therefore be shown to beat them where those assumptions break, and to at least match them where they hold.
Research Program
The model treats a test window and its calibration context as one set of tokens. Each sensor is a token carrying its nominal coordinates. Attention alternates between the sensor axis and the snapshot axis. Context snapshots additionally carry an embedding of their known direction, and test snapshots carry an "unknown" marker. Because sensors are tokens with position embeddings, one network can serve arrays of different geometry and size.
The network outputs a factor $L$ (a $P \times P$ complex lower-triangular matrix, where $P$ is the number of sensors). From it we form the corrected covariance
$$\hat{R} = L L^{H} + \epsilon I,$$
which is positive definite by construction for a small $\epsilon > 0$. Here $L^{H}$ is the conjugate transpose and $I$ the identity. We then run a differentiable eigendecomposition of $\hat{R}$, take the noise subspace $\hat{U}_n$, and compute the standard MUSIC spectrum
$$S(\theta) = \frac{1}{\| \hat{U}_n^{H} a(\theta) \|^2},$$
where $a(\theta)$ is the nominal steering vector, the array's response to a source at angle $\theta$. Using the nominal rather than the true steering vector is deliberate. It forces $\hat{R}$ to absorb the mismatch, so that the downstream estimator stays the classical, interpretable one.
Training uses episodes. Each episode samples an environment, consisting of:
- an array geometry with random position errors, gain and phase errors, and a random mutual-coupling matrix;
- a noise color and an SNR;
- a number of sources between one and three.
From that environment we draw $K \in \{0,\dots,5\}$ single-source calibration snapshots with their angles, plus $N$ test snapshots with $N$ between 1 and $P/2$. The loss combines direction error from a soft peak finder, following DA-MUSIC, with a subspace-distance term against the true noise subspace.
Mismatch families are split into three groups. Gain and phase errors alone are the parametric regime. Coupling, position error and colored noise together are the compound regime. A third family is held out from training entirely to test generalization.
Baselines
The naive reference point is classical MUSIC with diagonal loading on the starved sample covariance, using the nominal steering vectors. It shows how badly mismatch and starvation together hurt with no correction at all.
The structural-prior baseline is shrinkage MUSIC. It uses the Schäfer-style shrinkage covariance used by Kalunga et al. plus forward–backward (persymmetric) averaging following Pascal et al.. On real radar data Pascal et al. reached 3 dB SINR loss with $r$ rather than $2r$ secondary samples, where $r$ is the clutter rank. This baseline also gets the context snapshots pooled into its covariance. That way it sees exactly the data our model sees, except for the labels.
The parametric calibration baseline is the classical one. It estimates per-sensor gain and phase, plus a coupling matrix, by least squares from the same $K$ labeled calibration snapshots. It then runs shrinkage MUSIC with the corrected steering vectors.
The strongest learned baseline is DA-MUSIC (paper). We run it in three variants:
- Trained on the nominal array.
- Trained with the same randomized mismatch distribution as our model, a domain-randomization variant that sees no context at test time.
- Variant 2 fine-tuned by gradient descent on the $K$ calibration snapshots of each test environment. Learning rate and number of steps are tuned on validation environments under a fixed per-environment step budget.
DA-MUSIC reports 98% accuracy against 89% for classical MUSIC on synthetic 8-element uniform linear arrays with non-coherent sources. That result is in-distribution, so we rerun it rather than quote it under mismatch. For single-snapshot data, AutoMPDNN (paper) reports 98.38% of estimates within 1° on its simulated underwater set. We include it on the single-source slices only.
All learned models get the same simulator, parameter count within 20%, training compute and hyperparameter search budget.
Two ablations isolate the key ingredient. The first gives our model the context snapshots with labels removed. The second gives the context with labels randomly permuted.
Beating the field means two things on the compound and held-out mismatch families at $N < P$. First, lower median direction error and higher resolution probability than fine-tuned DA-MUSIC and parametric calibration, outside seed-level confidence intervals. Second, permuted labels must lose most of that margin.
Why not just…?
Why not just estimate gain, phase and coupling from the calibration sources by least squares? This is a real threat, and in the gain-and-phase-only regime we expect it to tie with or beat our model, since the error model is then exactly correct. The proposal's claim is restricted to compound and held-out mismatch, where a few snapshots cannot pin down a full coupling matrix plus position errors and the fit becomes ill-posed. The experiment separates the two regimes explicitly. If least squares also wins in the compound regime, the core claim is falsified, and we would report that.
Why not just train DA-MUSIC with heavy domain randomization? Randomization teaches robustness on average but gives the model no way to tell which of the many possible arrays it is facing. Our $K=0$ condition is essentially this baseline in our own architecture. The size of the jump from $K=0$ to $K=3$ directly measures what calibration adds beyond randomization.
Why not just fine-tune the pretrained network on the few calibration snapshots? Gradient fine-tuning on three to five snapshots is prone to overfitting, and on an embedded processor it costs optimization time per environment. We give it a tuned step budget and report both accuracy and wall-clock latency. If it matches in-context calibration on accuracy, the remaining argument is only about latency, and we will say so plainly.
Why not just use shrinkage or persymmetric structure, which needs no training? These priors fix the conditioning of the covariance, but they cannot repair steering-vector errors. Shrinking toward a well-conditioned matrix still leaves the subspace mismatched to $a(\theta)$. The pooled-context shrinkage baseline tests exactly this, because it receives the same snapshots without labels.
Proposed Experiments
The primary testbed is a mismatch-augmented version of the synthetic narrowband 8-element uniform linear array setting from DA-MUSIC, extended to 16 and 32 elements. Within it we sweep the ratio $N/P$ from $1/P$ to $1/2$, the context size $K$ from 0 to 5, and SNR from 0 to 30 dB. Metrics are:
- median and 90th-percentile direction error;
- probability of resolving two sources closer than a beamwidth;
- source-count accuracy.
The single-snapshot underwater configuration of AutoMPDNN serves as a second simulated domain. It uses horizontal arrays of 8 to 32 elements at 10, 20 and 30 dB, and is extended to two sources.
The near-field suite of DCD-MUSIC is a stretch test of transfer to range estimation, scored by root mean square position error. No reported RMSPE number was found on Lacuna for DCD-MUSIC, so we rerun it.
The real-data test is the 25-sensor GERES seismic array used by DA-MUSIC. Its sensors have genuine, unmodeled imperfections, and past catalogued events with known azimuth are natural calibration exemplars. We train only on simulation. At test time we use $K$ catalogued events nearby in time as context and score azimuth error on held-out events. No reported numerical result was found on Lacuna for DA-MUSIC on GERES, so all methods are rerun under the same protocol.
Three checks decide the outcome.
- Label dependence. On compound mismatch at $N<P$, the context-label ablation must show that correct labels matter, with permuted labels collapsing toward the unlabeled result. Otherwise the model is merely pooling extra data.
- Saturation. The gain as a function of $K$ should rise steeply and level off by three to five examples, the pattern seen in LLM calibration. A gain that grows without leveling off would suggest the model is learning the array rather than calibrating to it.
- Real data. On GERES, in-context calibration from simulation-only training must beat both simulation-trained DA-MUSIC and least-squares calibration from the same catalogued events. Matching them only on simulated data would not be enough.
References
- Deep Augmented Music Algorithm for Data-Driven Doa Estimation
- DCD-MUSIC: Deep-Learning-Aided Cascaded Differentiable MUSIC Algorithm for Near-Field Localization of Multiple Sources
- An automated multi-layer perceptron discriminative neural network based on Bayesian optimization achieves high-precision one-source single-snapshot direction-of-arrival estimation
- Meta-AF: Meta-Learning for Adaptive Filters
- Rating Short L2 Essays on the CEFR Scale with GPT-4
- Improving large language models for clinical named entity recognition via prompt engineering
- Aligning Unstructured Paris Agreement Climate Plans with Sustainable Development Goals
- Exploiting persymmetry for low-rank Space Time Adaptive Processing
- Online SSVEP-based BCI using Riemannian geometry