Sparse Multi-Family Deep Scattering Network

Romain Cosentino, Randall BalestrieroarXiv 2020 · arXiv

Introduction

Deep Neural Networks (DNNs) have revolutionized signal processing but often lack mathematical interpretability and stability against adversarial perturbations. Deep Scattering Networks (DSNs) address these issues by replacing learned convolutional weights with fixed, predefined wavelet transforms. However, traditional DSNs face two significant challenges: the "Filter Selection Problem," where choosing the optimal wavelet family requires domain expertise, and a lack of adaptive non-linearity to handle non-stationary noise.

This paper introduces the Sparse Multi-Family Deep Scattering Network (SMF-DSN), which overcomes these hurdles through a multi-family architectural design and a risk-based thresholding strategy.

The Multi-Family Architecture

The proposed Multi-Family Deep Scattering Network (MF-DSN) avoids the need for manual filter selection by utilizing a bank of complementary wavelet families (e.g., Morlet, Gammatone, and Paul). In the first layer, the input signal is transformed by all families simultaneously. The core innovation lies in the second layer, where the network computes "cross-terms" by applying every wavelet family to the outputs of every other family from the preceding layer.

Architecture of the Multi-Family Deep Scattering Network showing the generation of cross-family scattering coefficients.

By crossing wavelets with different time-frequency properties—such as applying a transient-focused Paul wavelet to the output of a harmonic-focused Morlet wavelet—the network captures complex acoustic patterns that a single-family DSN would miss. This approach allows the network to naturally leverage the strengths of different wavelets, such as the Paul family's ability to isolate sharp transients and the Morlet family's resolution for continuous harmonics.

Universal Frame Thresholding

To introduce sparsity and enhance noise robustness, the authors implement Universal Frame Thresholding. Unlike the standard complex modulus used in DSNs, this operator is an analytical, input-dependent threshold derived from empirical risk minimization. It acts as a locally adaptive mask, effectively distinguishing between salient signal activity and background noise without requiring learned scalar biases like ReLU. This ensures that the input-output mapping remains stable while aggressively filtering out non-stationary noise.

Experimental Evaluation

The SMF-DSN was evaluated on the Freefield1010 dataset for the task of Bird Activity Detection. The scattering coefficients were used as features for a standard Random Forest classifier to evaluate their representational power.

Performance comparison (AUC) between the Sparse Multi-Family DSN and the standard Multi-Family DSN.

As shown in the results, the SMF-DSN (incorporating both multi-family transforms and thresholding) achieved a mean Area Under the Curve (AUC) of 73.52, outperforming the MF-DSN (71.17). This improvement demonstrates that the risk-based thresholding operator significantly enhances the model's generalization and robustness in noisy field recordings.

Conclusion

The Sparse Multi-Family Deep Scattering Network provides a robust, interpretable framework for audio feature extraction. By leveraging the strengths of multiple wavelet families and a mathematically grounded sparsity operator, the SMF-DSN achieves superior performance in complex acoustic environments. While the combinatorial explosion of cross-terms currently limits the network's depth, future research may explore mitigating this complexity or applying the framework to multi-dimensional data like imagery and seismic signals.

Summary updated

Abstract

In this work, we propose the Sparse Multi-Family Deep Scattering Network (SMF-DSN), a novel architecture exploiting the interpretability of the Deep Scattering Network (DSN) and improving its expressive power. The DSN extracts salient and interpretable features in signals by cascading wavelet transforms, complex modulus and extract the representation of the data via a translation-invariant operator. First, leveraging the development of highly specialized wavelet filters over the last decades, we propose a multi-family approach to DSN. In particular, we propose to cross multiple wavelet transforms at each layer of the network, thus increasing the feature diversity and removing the need for an expert to select the appropriate filter. Secondly, we develop an optimal thresholding strategy adequate for the DSN that regularizes the network and controls possible instabilities induced by the signals, such as non-stationary noise. Our systematic and principled solution sparsifies the network's latent representation by acting as a local mask distinguishing between activity and noise. The SMF-DSN enhances the DSN by (i) increasing the diversity of the scattering coefficients and (ii) improves its robustness with respect to non-stationary noise.
BibTeX
@misc{cosentino2020sparse,
  author = {Romain Cosentino and Randall Balestriero},
  title = {{Sparse Multi-Family Deep Scattering Network}},
  year = {2020},
  eprint = {2012.07662},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.07662},
}
Figure