TLDR: DUSE (Dynamic Uncertainty–driven Sample Expansion) is a new framework designed to address data scarcity in Automatic Modulation Recognition (AMR). It uses an uncertainty scoring function to identify the most informative samples from large auxiliary datasets and an active learning strategy to iteratively refine the selection process. Experiments show DUSE consistently outperforms existing data selection methods, improves model generalization across different architectures, and maintains better class balance in expanded datasets, significantly boosting AMR performance with limited initial data.
Deep learning models have achieved remarkable success in Automatic Modulation Recognition (AMR), a critical task in wireless communication. However, these powerful models typically demand vast amounts of labeled data for effective training. In many real-world scenarios, obtaining such extensive datasets is a significant challenge, often involving high costs in terms of time and labor for manual collection and expert annotation. While data augmentation techniques can enrich existing samples, they don’t introduce genuinely new data, thus failing to fundamentally address the problem of data scarcity.
Introducing DUSE: A Novel Data Expansion Framework
To tackle these limitations, a new framework called Dynamic Uncertainty–driven Sample Expansion (DUSE) has been introduced. DUSE aims to intelligently expand a small, initial dataset by selectively incorporating the most informative samples from a larger, readily available auxiliary dataset. This approach not only increases the quantity and diversity of training data but also strives to maintain the original distribution characteristics of the target dataset, which is crucial for model generalization.
The DUSE framework operates on two core components:
-
Uncertainty Scoring Function: This function quantifies how ‘informative’ each sample in the auxiliary dataset is. It works by feeding a sample into a deep learning model and then analyzing the model’s predicted probabilities for different classes. Specifically, it calculates the difference between the highest and second-highest predicted probabilities. A smaller difference indicates higher uncertainty from the model, suggesting that the sample is near a decision boundary and thus highly valuable for refining the model’s understanding.
-
Active Learning Strategy: DUSE employs an active learning approach to continuously refine the selection process. It starts by training a model on the initial small target dataset. Then, it uses this trained model to score all samples in the auxiliary dataset based on their uncertainty. The most uncertain (and thus most informative) samples are then selected and added to the target dataset, while being removed from the auxiliary set. This process is iterative: the model is retrained on the newly expanded target dataset, and the scoring and selection process repeats. This dynamic feedback loop ensures that the model’s uncertainty estimates are continuously refined, leading to the selection of increasingly valuable samples.
Also Read:
- Smarter Data Pruning: Accounting for Class-Specific Learning Challenges
- Optimizing AI Energy Use in 5G Networks: A DeepRx Study on Efficiency and Knowledge Distillation
Performance and Generalization
Extensive experiments were conducted using DUSE across three widely-used AMR benchmark datasets: RML2016.10a, Sig2019-12, and RML2018.01a. DUSE was compared against eight existing coreset selection baselines, which are methods designed to select a compact yet informative subset of data. The results demonstrated that DUSE consistently outperformed these baselines in both class-balanced and class-imbalanced settings, showcasing its robustness and effectiveness in diverse data scenarios.
A significant finding was DUSE’s strong cross-architecture generalization. Even though the samples were selected using a 2D-CNN model, the expanded dataset proved effective when used to train other deep learning architectures like 1D-CNN, SigNet, AlexNet, GRU, and MCLDNN. This indicates that the data selected by DUSE is not overly specialized to a particular model, making it broadly useful.
Further analysis revealed the underlying reasons for DUSE’s superior performance. Visualizations of the expanded datasets showed that DUSE maintains a more faithful class balance compared to other methods, which often over-select certain categories. Additionally, t-SNE visualizations confirmed that DUSE ensures samples of the same category form tight clusters, highlighting the method’s excellent ability to discriminate between different modulation types.
An ablation study specifically highlighted the critical role of the active learning strategy within DUSE. Removing the iterative active learning loop significantly reduced the accuracy, underscoring that the continuous refinement of uncertainty estimates is key to selecting high-quality samples and achieving substantial performance gains.
In conclusion, DUSE offers a promising solution for data scarcity in AMR by intelligently expanding datasets with informative samples. Its dynamic, uncertainty-driven approach, coupled with active learning, makes it a powerful tool for improving the performance and generalization of deep learning models in low-resource environments. For more details, you can refer to the research paper.
