Academic

The Radio-Frequency Transformer for Signal Separation

arXiv:2603.09201v1 Announce Type: new Abstract: We study a problem of signal separation: estimating a signal of interest (SOI) contaminated by an unknown non-Gaussian background/interference. Given the training data consisting of examples of SOI and interference, we show how to build a fully data-driven signal separator. To that end we learn a good discrete tokenizer for SOI and then train an end-to-end transformer on a cross-entropy loss. Training with a cross-entropy shows substantial improvements over the conventional mean-squared error (MSE). Our tokenizer is a modification of Google's SoundStream, which incorporates additional transformer layers and switches from VQVAE to finite-scalar quantization (FSQ). Across real and synthetic mixtures from the MIT RF Challenge dataset, our method achieves competitive performance, including a 122x reduction in bit-error rate (BER) over prior state-of-the-art techniques for separating a QPSK signal from 5G interference. The learned representat

E
Egor Lifar, Semyon Savkin, Rachana Madhukara, Tejas Jayashankar, Yury Polyanskiy, Gregory W. Wornell · · 1 min read · 15 views
#cs.LG

arXiv:2603.09201v1 Announce Type: new Abstract: We study a problem of signal separation: estimating a signal of interest (SOI) contaminated by an unknown non-Gaussian background/interference. Given the training data consisting of examples of SOI and interference, we show how to build a fully data-driven signal separator. To that end we learn a good discrete tokenizer for SOI and then train an end-to-end transformer on a cross-entropy loss. Training with a cross-entropy shows substantial improvements over the conventional mean-squared error (MSE). Our tokenizer is a modification of Google's SoundStream, which incorporates additional transformer layers and switches from VQVAE to finite-scalar quantization (FSQ). Across real and synthetic mixtures from the MIT RF Challenge dataset, our method achieves competitive performance, including a 122x reduction in bit-error rate (BER) over prior state-of-the-art techniques for separating a QPSK signal from 5G interference. The learned representation adapts to the interference type without side information and shows zero-shot generalization to unseen mixtures at inference time, underscoring its potential beyond RF. Although we instantiate our approach on radio-frequency mixtures, we expect the same architecture to apply to gravitational-wave data (e.g., LIGO strain) and other scientific sensing problems that require data-driven modeling of background and noise.

Executive Summary

This article presents a novel approach to signal separation using a radio-frequency transformer for estimating a signal of interest (SOI) contaminated by an unknown non-Gaussian background/interference. Building upon Google's SoundStream, the authors develop a discrete tokenizer and a fully data-driven signal separator that achieves competitive performance in separating a QPSK signal from 5G interference. The proposed method demonstrates substantial improvements over conventional mean-squared error (MSE) and showcases zero-shot generalization to unseen mixtures at inference time. The authors anticipate the architecture's applicability to gravitational-wave data and other scientific sensing problems requiring data-driven modeling of background and noise.

Key Points

  • ▸ The authors propose a fully data-driven signal separator using a radio-frequency transformer.
  • ▸ The method achieves competitive performance in separating a QPSK signal from 5G interference.
  • ▸ The proposed architecture demonstrates zero-shot generalization to unseen mixtures at inference time.

Merits

Improved Performance

The method demonstrates substantial improvements over conventional mean-squared error (MSE) and achieves competitive performance in separating a QPSK signal from 5G interference.

Zero-Shot Generalization

The proposed architecture showcases zero-shot generalization to unseen mixtures at inference time, underscoring its potential beyond RF.

Applicability to Other Domains

The authors anticipate the architecture's applicability to gravitational-wave data and other scientific sensing problems requiring data-driven modeling of background and noise.

Demerits

Limited Scope

The method's development and testing are primarily focused on radio-frequency mixtures, limiting its generalizability to other domains.

Dependence on Training Data

The method's performance relies heavily on the quality and quantity of training data, which may not be readily available in certain applications.

Expert Commentary

While the article presents an innovative approach to signal separation, its scope and applicability are limited to radio-frequency mixtures. However, the proposed architecture's ability to adapt to different interference types and demonstrate zero-shot generalization is a significant advancement. The article highlights the potential of data-driven modeling in scientific sensing applications, underscoring the need for further research in this area. Future studies should focus on exploring the method's generalizability to other domains and investigating its robustness to different types of interference.

Recommendations

  • ✓ Further research should focus on exploring the method's generalizability to other domains, such as gravitational-wave data and other scientific sensing problems.
  • ✓ Investigating the method's robustness to different types of interference is essential for its practical application in various fields.

Sources

Related Articles