The Pseudo Projection Operator: Applications of Deep Learning to Projection Based Filtering in Non-Trivial Frequency Regimes

Matthew L. Weiss, Nathan C. Frey, Siddharth Samsi, Randy C. Paffenroth, Vijay GadepallyarXiv 2021 · arXiv

Abstract

Traditional frequency-based projection operators (PO) are limited by the requirement for prior knowledge of noise frequencies and an inability to handle overlapping signal-noise spectra. This paper introduces the Pseudo Projection Operator (PPO), a hybrid architecture that integrates a neural network into the Hilbert space projection framework. By learning continuous weighting coefficients rather than binary masks, the PPO adapts to complex noise regimes. Evaluated on the URMP music dataset, the PPO demonstrates superior performance over standard POs and Denoising Autoencoders (DAEs), particularly in scenarios with correlated noise and significant frequency overlap.

Introduction

Traditional frequency-based filters, or Projection Operators (PO), function by transforming a signal into the frequency domain and applying a binary mask: signal frequencies are weighted as 1, while noise frequencies are weighted as 0. However, this "all-or-nothing" approach suffers from two primary flaws. First, it requires precise prior knowledge of noise frequencies. Second, it fails in "non-trivial" regimes where signal and noise share the same frequencies; discarding a noisy frequency in such cases inevitably results in the loss of vital signal information.

Methods

The PPO is a neural-network hybrid model built on the theoretical foundation of Hilbert Spaces. It utilizes an analysis operator (to move from the time domain to the frequency domain) and a synthesis operator (to move back), typically implemented using trigonometric polynomials.

Block diagram of the Pseudo Projection Operator (PPO) architecture, showing the integration of a neural network (N) to predict continuous weights (M) between analysis and synthesis stages.

In a standard PO, the weighting matrix is binary. The PPO replaces this with a diagonal matrix $\tilde{M}$ where each frequency is assigned a weight between 0 and 1. A feed-forward neural network ($\mathcal{N}$) is trained to analyze the signal and predict these weights, using a sigmoid activation function to ensure the $[0, 1]$ range. The term "Pseudo" refers to the fact that because the weights are continuous, the operator is not strictly idempotent, though it maintains the structural rigor of classical filtering.

Experimental Setup

The model was evaluated using the University of Rochester Multi-Modal Music Performance (URMP) Dataset, focusing on individual instrument tracks. Three experimental environments were created:

  1. Clean: Original recordings with no added noise.
  2. Shuffle: A shuffled copy of the signal is added to itself, creating significant frequency overlap.
  3. Outliers: Samples are randomly scaled by a factor between 1 and 2, creating noise directly correlated with the underlying signal.

Key Results and Insights

The PPO was compared against standard POs and Denoising Autoencoders (DAEs). As shown in the performance metrics below, the PPO consistently achieved lower Mean Squared Error (MSE) across challenging noise regimes.

Normalized Test Set MSE for DAE, PO, and PPO models across Clean, Shuffle, and Outliers datasets using standardization scaling.
  • Superiority in Correlated Noise: In the "outliers" dataset, the PPO outperformed the best DAE by over 10% in terms of MSE. Rule-based POs failed significantly here due to the lack of a clear frequency cutoff.
  • Adaptive Transparency: On clean data, the PPO performed as well as an "all-pass" filter, demonstrating that the neural network learns to weight everything as 1 when no noise is detected, avoiding signal distortion.
  • Structural Advantage: Unlike "black box" DAEs, the PPO maintains the structure of a classical filter. The neural network only handles weight selection, while the actual transformation relies on fixed, well-understood mathematical operators. This structure allowed the PPO to scale consistently with model depth without the overfitting seen in some DAE configurations.

Conclusion

The PPO successfully bridges the gap between mathematical rigor and deep learning flexibility. Future work will focus on computational efficiency and applying the PPO to specialized fields like materials science and bioinformatics, where noise signatures are partially known but highly complex.

Summary updated

Abstract

Traditional frequency based projection filters, or projection operators (PO), separate signal and noise through a series of transformations which remove frequencies where noise is present. However, this technique relies on a priori knowledge of what frequencies contain signal and noise and that these frequencies do not overlap, which is difficult to achieve in practice. To address these issues, we introduce a PO-neural network hybrid model, the Pseudo Projection Operator (PPO), which leverages a neural network to perform frequency selection. We compare the filtering capabilities of a PPO, PO, and denoising autoencoder (DAE) on the University of Rochester Multi-Modal Music Performance Dataset with a variety of added noise types. In the majority of experiments, the PPO outperforms both the PO and DAE. Based upon these results, we suggest future application of the PPO to filtering problems in the physical and biological sciences.
BibTeX
@misc{weiss2021pseudo,
  author = {Matthew L. Weiss and Nathan C. Frey and Siddharth Samsi and Randy C. Paffenroth and Vijay Gadepally},
  title = {{The Pseudo Projection Operator: Applications of Deep Learning to Projection Based Filtering in Non-Trivial Frequency Regimes}},
  year = {2021},
  eprint = {2111.07140},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.07140},
}
Figure