Unfolding Neural Networks for Compressive Multichannel Blind Deconvolution

Bahareh Tolooshams, Satish Mulleti, Demba E. Ba, Yonina C. Eldar

TL;DR

This paper introduces LS-MBD, a novel neural network architecture that learns a data-driven, structured compression method for multichannel blind deconvolution. By learning how to compress signals before deconvolution, LS-MBD achieves significantly better accuracy and speed in recovering sparse filters compared to existing methods, paving the way for more efficient signal processing in applications like radar and sonar.

What Problem Does This Paper Address?

Many signal processing applications, such as radar, sonar, and seismic imaging, rely on recovering information from signals that have been convolved with unknown "sparse filters." These filters encode crucial information about the environment, like target locations. The challenge is that recovering these filters from raw measurements can be computationally expensive and require a large number of measurements. To reduce this burden, researchers have explored "compressive sensing" techniques, which aim to acquire fewer measurements by compressing the signal. However, previous methods either used fixed, data-independent compression strategies (like random projections or predefined filters) or still required full measurements for effective deconvolution. This paper tackles the problem of how to learn an effective and efficient compression strategy that is tailored to the specific data, enabling accurate sparse filter recovery from significantly fewer measurements.

Methods

The core of this paper is the Learned Structured Compressive Multichannel Blind-Deconvolution (LS-MBD) network. This is a model-based neural network built using an "unfolding" approach, which essentially means it mimics an iterative optimization algorithm within its architecture.

Here's a breakdown of the key components:

  • Unfolding Architecture: The network is designed as an autoencoder. The encoder part is structured as an unfolding of an iterative algorithm (specifically, a variant of FISTA for sparse recovery). Each "unfolding layer" iteratively refines an estimate of the sparse filters.
  • Learned Structured Compression Operator (Φ): This is the most novel aspect. Instead of using a fixed, random, or predefined compression matrix, LS-MBD learns a structured compression operator. This operator is designed to be hardware-efficient, implemented as a convolution with a learned filter h, followed by a truncation. This "data-driven" approach allows the compression to be optimized for the specific characteristics of the signals being processed.
  • Two-Stage Training: The network is trained in two stages:
  • Stage 1 (Full Measurements): Initially, the network is trained using full, uncompressed measurements. In this stage, the compression operator Φ is set to the identity matrix, and the network learns the common source signal s and the sparse filters X.
  • Stage 2 (Compressed Measurements): Once the source s is well-estimated, the network learns the structured compression operator Φ (i.e., the filter h) while optimizing for signal reconstruction and accurate sparse filter recovery from compressed measurements. This disentangles the source estimation from the compression learning.
  • Hardware Efficiency: The structured nature of the learned compression operator (Toeplitz matrix) allows for efficient implementation using fast Fourier transforms, making it practical for real-world hardware.

Datasets

The paper does not explicitly mention using pre-existing public datasets. Instead, it describes a data generation process for experiments:

  • Signal Model: Measurements y are generated as convolutions of a common source signal s and individual sparse filters x^n for N channels.
  • Source Signal: A Gaussian-shaped source signal is generated and normalized.
  • Sparse Filters: Filters are generated with supports chosen uniformly at random, and their amplitudes are drawn from a uniform distribution.
  • Training/Validation/Testing: For experiments, 10,000 examples are generated for training, 100 for validation, and 100 for testing.

Key Insights and Ideas

  • Learning Compression is Key: The most significant insight is that learning a data-driven, structured compression operator is far more effective than using fixed or random compression. This learned operator can adapt to the signal's characteristics, leading to better recovery.
  • Structured Compression for Efficiency: By enforcing a structured (Toeplitz) form on the compression operator, the method achieves hardware efficiency and computational speedups, which are critical for practical applications.
  • Unfolding for Performance: The unfolding approach, which mimics iterative optimization algorithms, allows the network to achieve high accuracy in sparse filter recovery, even with limited measurements.
  • Two-Stage Training for Disentanglement: Separating the source estimation from the compression learning simplifies the training process and allows for optimizing compression for various ratios independently.
  • Generalization with Weight-Tying: The paper highlights that tying weights in the unfolding layers (as opposed to having independent layers) leads to better generalization, especially in data-limited scenarios.

Key Results

  • Superior Accuracy: LS-MBD consistently outperforms baseline methods like GS-MBD (Gaussian unstructured compression), FS-MBD (fixed structured compression), and LS-MBD-L (learned structured, but with fewer unfolding steps) in terms of filter recovery error (measured in dB). For example, at a compression ratio (CR) of 23.74%, LS-MBD achieves an error of -50.35 dB, while FS-MBD is at -6.84 dB.
  • High Compression Ratios Achieved: LS-MBD demonstrates successful filter recovery at compression ratios as low as 23.74%, significantly lower than theoretical limits suggested for some structured methods (e.g., CR > 36.36% for FS-MBD).
  • Significant Speedup: Table 1 shows that LS-MBD has a much faster inference runtime compared to other methods. For a CR of 50%, LS-MBD takes only 0.0028 seconds, whereas FS-MBD takes 5.4204 seconds.
  • Learned Filter Properties: Figure 4 shows that the learned compression filter in LS-MBD aligns well with the source signal's frequency characteristics, suggesting it effectively preserves relevant information.

Limitations and Future Work

  • Noiseless Case Focus: The current work focuses on the noiseless case. The authors acknowledge that the presence of noise significantly impacts recovery performance and stability, suggesting this as a direction for future research.
  • LS-MBD-L Performance Gap: While LS-MBD-L (a variant with fewer unfolding steps) is faster, it exhibits significantly worse recovery performance, highlighting the importance of deep unfolding for accuracy.
  • Comparison to Oracle: The paper notes that the unstructured Gaussian compression (G-MBD) performs better than structured methods, attributing this to low mutual coherence. However, G-MBD is computationally expensive and serves as an "oracle" baseline. Future work could explore how to achieve similar low coherence with structured, learnable operators.
  • Practical Implementation: While the structured operator is designed for hardware efficiency, further exploration into its direct implementation on hardware could be a future direction.

Summary updated

Abstract

We propose a learned-structured unfolding neural network for the problem of compressive sparse multichannel blind-deconvolution. In this problem, each channel's measurements are given as convolution of a common source signal and sparse filter. Unlike prior works where the compression is achieved either through random projections or by applying a fixed structured compression matrix, this paper proposes to learn the compression matrix from data. Given the full measurements, the proposed network is trained in an unsupervised fashion to learn the source and estimate sparse filters. Then, given the estimated source, we learn a structured compression operator while optimizing for signal reconstruction and sparse filter recovery. The efficient structure of the compression allows its practical hardware implementation. The proposed neural network is an autoencoder constructed based on an unfolding approach: upon training, the encoder maps the compressed measurements into an estimate of sparse filters using the compression operator and the source, and the linear convolutional decoder reconstructs the full measurements. We demonstrate that our method is superior to classical structured compressive sparse multichannel blind-deconvolution methods in terms of accuracy and speed of sparse filter recovery.
BibTeX
@inproceedings{tolooshams2021unfolding,
  author = {Bahareh Tolooshams and Satish Mulleti and Demba E. Ba and Yonina C. Eldar},
  title = {{Unfolding Neural Networks for Compressive Multichannel Blind Deconvolution}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2021},
  doi = {10.1109/icassp39728.2021.9414443},
}
Figure