⚡ electrical engineering

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

This paper proposes ProtoAoA, a few-shot learning framework utilizing Prototypical Networks to estimate Angle-of-Arrival from complex IQ samples, demonstrating that it can achieve high accuracy (2–3 degrees MAE) with minimal training data (as few as 4 shots) on unseen angles using a real-world SDR dataset.

Original authors: Elsayed Mohammed, Omar Mashaal, Alec Digby, Pasquale Leone, Lorne Swersky, Ashkan Eshaghbeigi, Hatem Abou-Zeid

Published 2026-04-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Elsayed Mohammed, Omar Mashaal, Alec Digby, Pasquale Leone, Lorne Swersky, Ashkan Eshaghbeigi, Hatem Abou-Zeid

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a friend in a crowded, noisy concert hall. You can't see them, but you can hear their voice coming from a specific direction. Your brain naturally uses the difference in sound reaching your left and right ears to pinpoint where they are. This is essentially what Angle-of-Arrival (AoA) estimation does for wireless signals: it figures out exactly where a signal is coming from so your phone or radar can focus its attention there.

For a long time, engineers used complex math to solve this. More recently, they tried using Artificial Intelligence (AI). But here's the catch: traditional AI is like a student who needs to read every single book in the library to pass a test. It needs massive amounts of data to learn every possible angle a signal could come from. If the environment changes or the hardware is slightly different, that "student" often fails because it hasn't seen that specific scenario before.

The Problem: The "Data Starvation" Dilemma

The authors of this paper asked: What if we can't collect millions of data samples? What if we only have a tiny bit of information about new, unseen angles?

In the real world, collecting data is expensive, slow, and sometimes impossible. You can't always set up a testbed to measure signals from every single degree of the compass.

The Solution: ProtoAoA (The "Smart Guide" Analogy)

The paper introduces a new method called ProtoAoA, which uses something called Prototypical Networks.

Think of this like a travel guide instead of a dictionary.

Traditional AI (The Dictionary): Tries to memorize every single word and definition. If you ask it about a word it hasn't seen, it's stuck.
ProtoAoA (The Travel Guide): Instead of memorizing every single street, it learns the essence or the "prototype" of different neighborhoods.

Here is how it works in simple terms:

Learning the "Vibe" (Prototypes): During training, the AI looks at a few examples of signals coming from specific directions. It doesn't memorize every single wave; instead, it creates a "mental average" or a prototype for that direction. Imagine it creates a "fingerprint" for a signal coming from the North, another for the East, etc.
The "Few-Shot" Magic: Now, imagine you walk into a new city (a new environment) and see a street sign you've never seen before. Instead of needing a whole library of books about that city, you just look at one or two examples of that new street.
Matching: The AI takes your new signal, creates a quick "fingerprint" for it, and asks: "Does this new fingerprint look more like the 'North' prototype or the 'East' prototype?" It picks the closest match.

The Experiment: The "Skip" Game

To test this, the researchers built a real-world test setup using radio equipment (SDRs). They created a grid of 255 possible angles (like a giant map of directions).

They played a game called "Skip-3":

They trained the AI on only 23% of the angles (the "Blue" dots on their map).
They left huge gaps between the training angles (skipping 3 angles in between).
Then, they tested the AI on the unseen angles (the "Red" dots) it had never seen before, giving it only a tiny handful of examples (called "shots") to learn from.

The Results: Doing More with Less

The results were impressive, almost like a magic trick:

With just 4 examples (4-shots) of a new angle, the AI could guess the direction with an error of only 3 degrees.
With 32 examples, the error dropped to just 2 degrees.

To put that in perspective: If you are trying to point a flashlight at a target 100 meters away, an error of 2 degrees means you are off by only about 3.5 meters. That is incredibly precise, especially considering the AI had never seen that specific angle before and only had a few seconds of data to learn it.

Why This Matters

This paper proves that we don't need to feed AI a "data buffet" to make it smart. By using Prototypical Networks, we can teach machines to be adaptable.

Real-world impact: This is huge for things like 5G/6G networks, self-driving cars, and radar systems. These technologies often face new environments (new cities, new weather, new hardware) where they can't wait to collect years of data.
The Takeaway: Just like a human can learn a new language by picking up a few key phrases and understanding the grammar, this AI can learn a new signal direction by understanding the "prototype" of the signal. It's efficient, smart, and ready for a world where data is scarce.

In short, ProtoAoA is the difference between a robot that needs to read the entire encyclopedia to find a friend, and a smart guide who can find that friend just by recognizing their general style and a few key details.

1. Problem Statement

Angle-of-Arrival (AoA) estimation is a critical signal processing task in wireless communications used for localization, beamforming, and interference management. While traditional methods (e.g., MUSIC algorithm) and deep learning (DL) solutions exist, they face significant limitations:

Traditional Methods: Sensitive to noise, multipath conditions, and hardware imperfections.
Standard Deep Learning: Requires massive, comprehensive datasets covering all possible angles and environmental conditions to generalize. Collecting such data is expensive, time-consuming, and often impractical in real-world scenarios where antenna configurations or environments change.

The Core Challenge: How to achieve accurate AoA estimation with limited training data, specifically enabling the model to generalize to unseen angles (classes) without retraining the entire network.

2. Methodology: ProtoAoA

The authors propose ProtoAoA, a framework utilizing Prototypical Networks (PN), a meta-learning approach designed for few-shot learning.

A. Data Collection & Setup

Hardware: A Software Defined Radio (SDR) testbed using USRP X300 transceivers and an Ettus Octoclock.
Dataset: Real-world IQ (In-phase/Quadrature) signals captured outdoors with 4 receiver antennas.
Scope: 225 unique angle combinations (Azimuth/Elevation) ranging from -70° to 70°.
Variability: Includes diverse modulation schemes (BPSK, 16-QAM, 64-QAM) and sampling rates.

B. The "Skip" Few-Shot Paradigm

To simulate real-world data scarcity, the authors defined a "Skip" training strategy:

Training Set (Blue Angles): Only a subset of angles (e.g., 23% of the total dataset) is used for training the network encoder.
Testing Set (Red Angles): The model is tested on "unseen" angles that were not present during training.
Skip-N Configurations:
- Skip-1: 1 angle gap between training and testing classes.
- Skip-2: 2 angle gaps.
- Skip-3: 3 angle gaps (most challenging).
Goal: Train on a sparse set of angles, then use only a few "shots" (samples) of a new, unseen angle to create a prototype and classify it.

C. Network Architecture & Algorithm

Encoder: A lightweight Multi-Layer Convolutional Neural Network (CNN) processes complex 4-channel IQ data (shape: $4 \times n$ ). It uses 1D convolutions, max-pooling, and fully connected layers to map inputs to a 10-dimensional embedding space.
Prototypical Mechanism:
- Support Set: For a new class (unseen angle), $K$ samples are selected. The encoder extracts features, and the prototype ( $P_y$ ) is calculated as the mean of these feature embeddings.
- Query Set: New samples are passed through the encoder.
- Classification: The distance (Euclidean) between the query feature and the class prototypes is calculated. The class with the nearest prototype is selected.
- Loss Function: Cross-entropy loss is minimized during the training episodes to ensure the embedding space clusters same-class samples tightly and separates different classes.
Inference: No retraining is required for new angles. The system simply computes the prototype for the new angle using the few available shots and performs nearest-neighbor classification.

3. Key Contributions

First Application of PNs to AoA: This is the first study to apply Prototypical Networks specifically for efficient AoA classification using raw IQ signals.
Real-World Validation: Unlike many simulation-based papers, this work validates the approach on a physical SDR testbed with diverse modulations and sampling rates.
Data Efficiency Demonstration: The study proves that high-accuracy AoA estimation is possible with extremely limited data, addressing the "data scarcity" bottleneck in wireless AI.

4. Experimental Results

The model was evaluated using Classification Accuracy and Mean Absolute Error (MAE) across Skip-1, Skip-2, and Skip-3 scenarios.

Performance on Unseen Angles:
- Skip-3 (Most Challenging):
  - With 4 shots (4 samples per unseen angle): Achieved 3° MAE and ~90% accuracy.
  - With 32 shots: Achieved 2° MAE.
- Skip-1: Achieved even lower errors (approaching 1° MAE with 32 shots).
Accuracy Trends:
- Training data accuracy reached up to 99% with sufficient shots.
- Unseen test data accuracy dropped slightly (10-20%) compared to training data but remained robust (up to 95% in Skip-1).
Top-2 Accuracy: The model showed high confidence even when the top-1 prediction was wrong. Top-2 accuracy exceeded 85% even with only 2 shots, suggesting the network correctly identifies the neighborhood of the true angle.
Data Efficiency: The model required only 23% of the total angle classes for training to generalize effectively to the remaining 77%.

5. Significance and Future Work

Significance: ProtoAoA demonstrates that meta-learning can overcome the reliance on massive labeled datasets in wireless communications. It enables rapid adaptation to new antenna configurations, environments, or user locations with minimal data collection.
Implications: This approach is vital for dynamic 5G/6G networks where beam management and localization must adapt in real-time without retraining massive models.
Future Directions:
- Refining the model to better distinguish between neighboring angles (reducing Top-1 vs. Top-2 gaps).
- Investigating Raw-IQ Foundation Models (e.g., IQFM) to provide pre-trained representations that further enhance label efficiency before few-shot adaptation.

In conclusion, the paper successfully validates that Prototypical Networks offer a robust, data-efficient solution for AoA estimation, capable of achieving high precision with minimal training data in real-world hardware environments.