pith. sign in

arxiv: 2604.14476 · v1 · submitted 2026-04-15 · 📡 eess.SP

ProtoAoA: Few-Shot Angle-of-Arrival Estimation using Prototypical Networks

Elsayed Mohammed , Omar Mashaal , Alec Digby , Pasquale Leone , Lorne Swersky , Ashkan Eshaghbeigi , Hatem Abou-Zeid This is my paper

Pith reviewed 2026-05-10 12:05 UTC · model grok-4.3

classification 📡 eess.SP
keywords prototypical networksfew-shot learningangle of arrivalwireless signal processingIQ samplessoftware defined radiodeep learning
0
0 comments X

The pith

Prototypical networks trained on complex IQ samples can estimate unseen angles of arrival to within a few degrees using only four to thirty-two additional examples after exposure to just 23 percent of possible directions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes using prototypical networks for angle-of-arrival estimation to overcome the data hunger of standard deep learning methods in wireless communications. The approach extracts embeddings from IQ samples to create prototypes for known angles, then adapts to new angles with few shots. Results from a real SDR testbed show low error on unseen angles, making the technique practical where collecting large labeled datasets is difficult. A reader would care if this means reliable localization and beamforming become feasible with far less upfront data collection.

Core claim

The paper establishes that a prototypical network architecture called ProtoAoA, when trained on complex IQ samples from only 23% of the angle classes in a dataset, can achieve a mean absolute error of 3 degrees on unseen angles with 4-shot training and 2 degrees with 32-shot training, as validated on a software-defined radio testbed.

What carries the argument

Prototypical networks that compute class prototypes from few-shot embeddings of complex in-phase and quadrature samples to classify or estimate angle-of-arrival.

If this is right

  • Models require far less training data than conventional deep learning for AoA tasks.
  • Adaptation to new angles happens with minimal additional samples.
  • The method works on real-world collected data from SDR hardware.
  • Similar techniques may apply to other wireless functions limited by data availability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Deployment in environments with changing conditions could become more feasible without full retraining.
  • Integration with existing wireless protocols might allow on-the-fly angle estimation for beamforming.
  • Extensions to multi-antenna systems or higher frequency bands could be tested next.

Load-bearing premise

The embeddings from the network form stable prototypes for angles that continue to work when the wireless channel changes or when new angles appear.

What would settle it

Collect new IQ samples for unseen angles in a noisy multipath environment different from the training testbed and check whether the mean absolute error rises above 5 degrees for 4-shot adaptation.

Figures

Figures reproduced from arXiv: 2604.14476 by Alec Digby, Ashkan Eshaghbeigi, Elsayed Mohammed, Hatem Abou-Zeid, Lorne Swersky, Omar Mashaal, Pasquale Leone.

Figure 1
Figure 1. Figure 1: Prototypical Network Architecture and Training [6]. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Proposed "Skip-3" few shot learning AoA data set-up: [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The prototypical network encoder architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy and MAE performance on few-shot testing with training and test data across different data setups. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of Top-1 and Top-2 classification accu [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
read the original abstract

Angle-of-arrival (AoA) estimation is a crucial function in wireless communications used for localization, beam-forming, interference management, and other applications. Deep learning (DL) solutions have been proposed for AoA to mitigate limitations of traditional AoA estimation techniques such as sensitivity to noise and the inability to generalize across different array characteristics. A challenge, however, of DL-based approaches is their reliance on large data collection campaigns and model training. This paper proposes the application of Prototypical Networks (PN) to address this challenge and utilizes a real-world dataset collected on a software defined radio (SDR) testbed to validate the effectiveness of the proposed solution. Prototypical Networks excel in extracting representative embeddings from unstructured input data, establishing class prototypes during training that can be few-shot trained on unseen classes. We demonstrate the efficacy of PNs for AoA classification using complex IQ samples, focusing on its ability to correctly classify new, unseen angles that the model was not trained on previously. Our results show that training our proposed ProtoAoA on only 23% of the AoA dataset classes can attain a mean absolute error (MAE) of 3 degrees with only 4-shots of training on the unseen angles - and an MAE of 2 degrees with 32-shots of training data. These results demonstrate that the developed prototypical network architecture requires remarkably few data samples to achieve reliable AoA estimation - and highlights its potential for other wireless applications where data availability is limited.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces ProtoAoA, a prototypical network for few-shot angle-of-arrival (AoA) estimation that operates directly on complex IQ samples. Meta-trained on 23% of discrete angle classes from a real SDR testbed dataset, the model is claimed to achieve 3° mean absolute error (MAE) after 4-shot adaptation and 2° MAE after 32-shot adaptation on held-out angles, demonstrating reduced data requirements compared with conventional deep-learning AoA approaches.

Significance. If the reported generalization holds under broader conditions, the work would offer a practical route to low-data AoA estimation in wireless systems, with direct relevance to localization, beamforming, and interference management. The use of real SDR-collected IQ data rather than simulated channels is a positive feature; however, the absence of baselines and robustness checks limits the immediate impact.

major comments (3)
  1. Abstract: the central performance claims (MAE of 3° at 4 shots and 2° at 32 shots after training on 23% of classes) are stated without any baseline comparison (e.g., MUSIC, ESPRIT, or standard supervised CNNs), without reported standard deviations or number of trials, and without explicit description of how the angle classes were partitioned into meta-train and meta-test sets. These omissions make it impossible to judge whether the numbers reflect genuine few-shot generalization or dataset-specific artifacts.
  2. Evaluation section (inferred from abstract and results description): the manuscript reports results on a single SDR testbed collection without cross-environment, cross-time, or cross-SNR splits. This directly bears on the weakest assumption that the learned embedding space remains stable under unmodeled variations in multipath, phase noise, and array calibration that differ between training and test collections.
  3. Abstract and methods: no details are provided on the precise architecture of the embedding network (number of layers, input representation of complex IQ samples, distance metric used for prototypes), the total number of angle classes in the dataset, or the exact procedure for forming prototypes from the support shots. These omissions prevent reproduction and assessment of whether the reported MAE values are robust.
minor comments (2)
  1. Abstract: the phrase 'training our proposed ProtoAoA on only 23% of the AoA dataset classes' is ambiguous; clarify whether this refers to the fraction of discrete angle classes or of total samples.
  2. The manuscript would benefit from a table summarizing the dataset (number of angles, samples per angle, SNR range, array size) and from explicit statements of the loss function and optimization hyperparameters used during meta-training.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive feedback on our manuscript. We have revised the paper to address the major concerns regarding the presentation of results, evaluation robustness, and methodological details. Our point-by-point responses are as follows.

read point-by-point responses

  1. Referee: Abstract: the central performance claims (MAE of 3° at 4 shots and 2° at 32 shots after training on 23% of classes) are stated without any baseline comparison (e.g., MUSIC, ESPRIT, or standard supervised CNNs), without reported standard deviations or number of trials, and without explicit description of how the angle classes were partitioned into meta-train and meta-test sets. These omissions make it impossible to judge whether the numbers reflect genuine few-shot generalization or dataset-specific artifacts.

    Authors: We agree that these details are necessary for proper evaluation of the results. In the revised manuscript, we have added baseline comparisons to MUSIC, ESPRIT, and a standard supervised CNN. We now include standard deviations for the reported MAE values, computed over multiple independent trials. We have also added an explicit description of the angle class partitioning into meta-train and meta-test sets to clarify the few-shot generalization setup. revision: yes

  2. Referee: Evaluation section (inferred from abstract and results description): the manuscript reports results on a single SDR testbed collection without cross-environment, cross-time, or cross-SNR splits. This directly bears on the weakest assumption that the learned embedding space remains stable under unmodeled variations in multipath, phase noise, and array calibration that differ between training and test collections.

    Authors: We acknowledge the limitation of using a single SDR testbed collection. While we cannot perform cross-environment splits without additional data, we have included cross-SNR evaluations within the existing dataset and added a discussion on the potential impact of unmodeled variations such as multipath and phase noise on the embedding space. This limitation is now explicitly stated in the manuscript. revision: partial

  3. Referee: Abstract and methods: no details are provided on the precise architecture of the embedding network (number of layers, input representation of complex IQ samples, distance metric used for prototypes), the total number of angle classes in the dataset, or the exact procedure for forming prototypes from the support shots. These omissions prevent reproduction and assessment of whether the reported MAE values are robust.

    Authors: We have expanded the methods section to include the precise details of the embedding network architecture, including the number of layers and input representation of complex IQ samples as real and imaginary channels. We specify the Euclidean distance metric for prototype computation and describe the procedure for forming prototypes by averaging the embeddings of the support shots. The total number of angle classes in the dataset is now stated, along with the partitioning details. revision: yes

standing simulated objections not resolved
  • The evaluation relies on a single SDR testbed dataset, and we do not have additional collections to enable cross-environment or cross-time validation.

Circularity Check

0 steps flagged

No circularity: results are direct empirical measurements on held-out classes

full rationale

The paper applies the established prototypical networks framework to AoA estimation on complex IQ samples from a single SDR-collected dataset. The reported MAE values (3° with 4 shots, 2° with 32 shots) after training on 23% of angle classes are presented as observed performance on explicitly unseen angle classes. No equations, self-citations, or ansatzes reduce these numbers to fitted parameters, self-definitions, or prior author results by construction. The method relies on standard PN prototype computation and nearest-neighbor classification without importing uniqueness theorems or renaming known patterns as new derivations. The central claim therefore remains an independent empirical finding rather than a tautology.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The method rests on the standard assumptions of prototypical networks plus typical deep-learning hyperparameters; no new physical entities or ad-hoc constants are introduced in the abstract.

free parameters (2)
  • shot count for adaptation
    4-shot and 32-shot regimes are selected for evaluation to demonstrate few-shot behavior.
  • initial training class fraction
    23% of classes used for base training; the split is chosen to test generalization.
axioms (1)
  • domain assumption Prototypical networks produce class prototypes from embeddings of complex IQ samples that support accurate classification of unseen angles.
    This is the core inductive bias transferred from few-shot learning literature to the AoA task.

pith-pipeline@v0.9.0 · 5597 in / 1272 out tokens · 57573 ms · 2026-05-10T12:05:51.496733+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [1]

    A scalable fingerprint-based angle-of- arrival machine learning approach for cellular mobile radio localization,

    R. D. Timoteo and D. C. Cunha, “A scalable fingerprint-based angle-of- arrival machine learning approach for cellular mobile radio localization,” Computer Communications, vol. 157, pp. 92–101, 2020

    work page 2020

  2. [2]

    Rethinking beam management: Generalization limits under hardware heterogeneity,

    N. Zeulin, O. Galinina, I. Kilinc, S. Andreev, and R. W. Heath Jr, “Rethinking beam management: Generalization limits under hardware heterogeneity,”arXiv preprint arXiv:2602.18151, 2026

    work page arXiv 2026

  3. [3]

    Protobeam: Generalizing deep beam prediction to unseen antennas using prototypical networks,

    O. Mashaal, E. Mohammed, A. Digby, L. Swersky, A. Eshaghbeigi, and H. Abou-Zeid, “Protobeam: Generalizing deep beam prediction to unseen antennas using prototypical networks,” inGLOBECOM 2024 - 2024 IEEE Global Communications Conference, pp. 133–138, 2024

    work page 2024

  4. [4]

    Angle-of-arrival estimation using an adaptive machine learning framework,

    A. Khan, S. Wang, and Z. Zhu, “Angle-of-arrival estimation using an adaptive machine learning framework,”IEEE Communications Letters, vol. 23, no. 2, pp. 294–297, 2019

    work page 2019

  5. [5]

    Aoa-net: Estimating angle-of-arrival using wi-fi channel state information based on deep neural networks with subcarrier selection,

    T. Kumrai, Z. Cai, T. Maekawa, T. Hara, K. Ohara, T. Murakami, and H. Abeysekera, “Aoa-net: Estimating angle-of-arrival using wi-fi channel state information based on deep neural networks with subcarrier selection,”Journal of Information Processing, vol. 32, pp. 863–872, 2024

    work page 2024

  6. [6]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,”Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  7. [7]

    Autocorrelation analysis and near-field localization of the radiating sources with cyclostationary properties,

    Y . Kuznetsov, A. Baev, M. Konovalyuk, A. Gorbunova, and J. A. Russer, “Autocorrelation analysis and near-field localization of the radiating sources with cyclostationary properties,”IEEE Transactions on Electromagnetic Compatibility, vol. 62, no. 5, pp. 2186–2195, 2020

    work page 2020

  8. [8]

    Learning and data-driven beam selection for mmwave communications: An angle of arrival-based approach,

    C. Antón-Haro and X. Mestre, “Learning and data-driven beam selection for mmwave communications: An angle of arrival-based approach,” IEEE Access, vol. 7, pp. 20404–20415, 2019

    work page 2019

  9. [9]

    Multiple emitter location and signal parameter estimation,

    R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Transactions on Antennas and Propagation, vol. 34, no. 3, pp. 276–280, 1986

    work page 1986

  10. [10]

    Deepaoanet: Learning angle of arrival from software defined radios with deep neural networks,

    Z. Dai, Y . He, V . Tran, N. Trigoni, and A. Markham, “Deepaoanet: Learning angle of arrival from software defined radios with deep neural networks,”IEEE Access, vol. 10, pp. 3164–3176, 2022

    work page 2022

  11. [11]

    Prototypical network for few-shot signal recognition,

    H. Wang, L. Qi, Y . Han, and Y . Lin, “Prototypical network for few-shot signal recognition,” in2022 9th International Conference on Dependable Systems and Their Applications (DSA), pp. 980–985, 2022

    work page 2022

  12. [12]

    Prototypical network with residual attention for modulation classification of wireless communication signals,

    B. Zang, X. Gou, Z. Zhu, L. Long, and H. Zhang, “Prototypical network with residual attention for modulation classification of wireless communication signals,”Electronics, vol. 12, no. 24, p. 5005, 2023

    work page 2023

  13. [13]

    Fine-grained transduc- tive prototypical network-based few-shot signal modulation classification using coarse labels,

    S. Feng, Y . Wang, Z. Wen, L. Xu, and M. Yan, “Fine-grained transduc- tive prototypical network-based few-shot signal modulation classification using coarse labels,”IEEE Transactions on Cognitive Communications and Networking, vol. 12, pp. 2189–2204, 2026

    work page 2026

  14. [14]

    Hybrid feature fused few-shot learning for csi-based adaptive beam prediction,

    C. Thiangkate, S. Nishiyama, D. Chakraborty, M. Okada, K. Thonglek, P. Leelaprute, A. Rungsawang, B. Manaskasemsak, and K. Cham- nongthai, “Hybrid feature fused few-shot learning for csi-based adaptive beam prediction,” in2025 24th International Symposium on Communi- cations and Information Technologies (ISCIT), pp. 13–18, 2025

    work page 2025

  15. [15]

    Lightweight and generalizable aoa estimation for iot: A novel few-shot learning approach,

    O. Mashaal, E. Mohammed, A. Digby, P. Leone, L. Swersky, A. Eshagh- beigi, and H. Abou-Zeid, “Lightweight and generalizable aoa estimation for iot: A novel few-shot learning approach,” inICC 2025 - IEEE International Conference on Communications, pp. 686–691, 2025

    work page 2025

  16. [16]

    Iqfm—a wireless foundation model for i/q streams in ai-native 6g,

    O. Mashaal and H. Abou-Zeid, “Iqfm—a wireless foundation model for i/q streams in ai-native 6g,”IEEE Open Journal of the Communications Society, vol. 7, pp. 1426–1441, 2026

    work page 2026