ProtoAoA : Few-Shot Angle-of-Arrival Estimation using Prototypical Networks This research was funded by Alberta Innovates and the Natural Sciences and Engineering Research Council of Canada (NSERC) through the NSERC Alliance-Alberta Innovates Program and the NSERC Discovery Grant RGPIN-2021-04050.

Elsayed Mohammed 3 , Omar Mashaal 3 , Alec Digby 1 , Pasquale Leone 1 , Lorne Swersky 1 , Ashkan Eshaghbeigi 1 , and Hatem Abou-Zeid 3 · 2026

Read ProtoAoA : Few-Shot Angle-of-Arrival Estimation using Prototypical Networks This research was funded by Alberta Innovates and the Natural Sciences and Engineering Research Council of Canada (NSERC) through the NSERC Alliance-Alberta Innovates Program and the NSERC Discovery Grant RGPIN-2021-04050. by Elsayed Mohammed 3 , Omar Mashaal 3 , Alec Digby 1 , Pasquale Leone 1 , Lorne Swersky 1 , Ashkan Eshaghbeigi 1 , and Hatem Abou-Zeid 3 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-14476.txt.

Introduction

Mohammed and six colleagues apply prototypical networks to angle-of-arrival estimation, a wireless function used for localization, beamforming, and interference control. Deep-learning methods beat traditional noise sensitivity and array-generalization limits but demand large data campaigns, so the team trains class prototypes from complex IQ samples on a software-defined-radio testbed. Prototypes learned from a few examples then classify new, unseen angles. Training on only 23 percent of angle classes reaches a mean absolute error of 3 degrees with 4 shots on unseen angles, and 2 degrees with 32 shots.

Catalog record

CREATORElsayed Mohammed 3 , Omar Mashaal 3 , Alec Digby 1 , Pasquale Leone 1 , Lorne Swersky 1 , Ashkan Eshaghbeigi 1 , and Hatem Abou-Zeid 3
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27618
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.14476

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ProtoAoA : Few-Shot Angle-of-Arrival Estimation using Prototypical Networks This research was funded by Alberta Innovates and the Natural Sciences and Engineering Research Council of Canada (NSERC) through the NSERC Alliance-Alberta Innovates Program and the NSERC Discovery Grant RGPIN-2021-04050. Elsayed Mohammed 3 , Omar Mashaal 3 , Alec Digby 1 , Pasquale Leone 1 , Lorne Swersky 1 , Ashkan Eshaghbeigi 1 , and Hatem Abou-Zeid 3 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 ∘ with only 4-shots of training on the unseen angles − - and an MAE of 2 ∘ 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. I Introduction Angle-of-arrival (AoA) estimation is a signal processing technique used in wireless communications to determine the direction from which a signal reaches an antenna array. This technique helps identify the precise location or direction of signal source. Therefore, it is crucial for applications such as beam-forming, localization, radar systems, and interference management. In cellular systems [ 13 ] , AoA estimation plays a significant role in enhancing signal quality, and improving overall communication efficiency. As a result, AoA estimation remains a key area of research, with focus in developing accurate, adaptable, and efficient methods suitable for various dynamic environments. More broadly, recent work has highlighted that hardware heterogeneity, codebook mismatch, and environment shift remain major obstacles to the practical generalization of ML-based wireless learning methods, including beam management systems [ 16 ] . Related efforts such as ProtoBeam further illustrate the need for adaptable wireless learning frameworks that can generalize across unseen antenna configurations [ 9 ] . Due to their ability to extract meaningful features from unstructured data, data-driven machine learning (ML) solutions have been proposed in different communication networks and signal processing tasks including AoA estimation. The works in [ 4 , 13 , 5 ] provide various data and artificial intelligence (AI) driven frameworks for AoA estimation. Such AI models can achieve super-resolution with lower complexity and signal snapshots that traditional AoA estimation techniques, and can be resilient to antenna perturbations, phase and amplitude inconsistencies, and can be antenna-agnostic. However, a key challenge in building these AI-driven solutions is the need for comprehensive large data collection and model training campaigns to achieve this robust performance. Fig. 1: Prototypical Network Architecture and Training [ 11 ] . Recently, meta-learning approaches such as prototypical networks (PN), have gained significant attention for their effectiveness in data-limited data and domain adaptation scenarios. PN [ 11 ] excel at generalizing across classes with minimal labeled data, making them useful in environments where acquiring comprehensive labeled datasets is challenging. PN work by learning how to create an embedded representation ( prototypes ) for each class based on a small set of labeled examples, a paradigm that is referred to as few-shot learning. Based on these prototypes, it classifies a new sample by finding the closest prototype to this sample’s embedded feature. As shown in Figure 1 , during training, the network learns an embedding space that groups samples of the same class closer together while separating those from different classes. Given that PNs are well-regarded for few-shot learning, we explore their potential for generalizing to unseen classes in the AoA estimation task. This investigation aims to answer two main questions: how effectively can a PN learn distinctive feature representations for AoA classification, and to what extent can this learned feature space generalize to novel angles not encountered during training? We hypothesize that the prototypical network’s ability to generate discriminative embeddings will support accurate AoA classification, even under few-shot conditions of unseen angles. The main contributions of this paper are: • We propose ProtoAoA , a few-shot learning approach that uses real-world IQ signals to efficiently perform AoA estimation with a very limited training dataset. To the best of our knowledge, this study is the first to adopt prototypical networks for efficient AoA classification. The proposed approach is extensible to other tasks offering a novel perspective on addressing data scarcity in AI for wireless communications. • Our work demonstrates the robustness and effectiveness of prototypical networks across various data-limited conditions. We evaluate the network’s performance in progressively challenging data setups, where it encounters an increasing number of unseen classes, evaluating the trade-off between data scarcity and accuracy. Our findings show that training on only 23% of the AoA dataset classes can attain a mean absolute error (MAE) of 3 ∘ with only 4-shots of training on the unseen angles and an MAE of 2 ∘ with 32-shots of training data. • We provide the aforementioned analysis with a real-world dataset collected on a software defined radio (SDR) testbed under diverse sample rates and modulations. The rest of this paper is organized as follows. Section II summarizes of related work. In Section III, we describe our testbed, dataset, and proposed ProtoAoA methodology. Section IV discusses our results and findings, and we conclude the paper in Section V. II Background & Related Work Angle-of-Arrival Estimation. Angle-of-Arrival (AoA) estimation is the task of deducing the direction from which a signal arrives at a receiver. It involves determining the angle between the source (e.g., user equipment) and the receiver (e.g., base station). AoA estimation is critical for accurately pinpointing the source’s direction and is often a step in positioning and localization or in refining the beam’s focus [ 6 , 1 ] . The MUSIC algorithm [ 10 ] is commonly used to estimate the AoA by analyzing the phase differences in channel state information (CSI) across multiple antennas. However, MUSIC faces several challenges, particularly under multipath conditions, where the assumption that the signal and noise subspaces are mutually orthogonal may be violated. In this context, methods such as AoA-net [ 5 ] and DeepAoANet [ 2 ] have been proposed for AoA estimation. In particular, DeepAoANet [ 2 ] , a data-driven method, was proposed to estimate the number of signals received by an antenna array and their respective AoAs. Its two developed DL architectures outperform the standard MUSIC algorithm. However,

Read the complete text