Abstract
For many machine learning applications, a common input representation is a spectrogram. The underlying representation for a spectrogram is a short time Fourier transform (STFT) which gives complex values. The spectrogram uses the magnitude of these complex values, a commonly used detector. Modern machine learning systems are commonly overparameterized, where possible ill-conditioning problems are ameliorated by regularization. The common use of rectified linear unit (ReLU) activation functions between layers of a deep net has been shown to help this regularization, improving system performance. We extend this idea of ReLU activation to detection for the complex STFT, providing a simple-to-compute modified and regularized spectrogram, which potentially results in better behaved training. We then confirmed the benefit of this approach on a noisy acoustic data set used for a real-world application. Generalization performance improved substantially. This approach might benefit other applications which use time-frequency mappings, for acoustic, audio, and other applications.
Introduction
Deep learning models are notoriously overparameterized, possessing significantly more parameters than available data points. This often leads to "singularities" or linear dependencies within the network, resulting in a non-convex loss function that is difficult to optimize and prone to poor generalization. While internal regularization techniques like Rectified Linear Units (ReLU) are standard for managing these issues within hidden layers, the input representation itself is often overlooked.
In acoustic modeling, the standard input is a spectrogram derived from a Short-Time Fourier Transform (STFT). However, the STFT is inherently redundant due to overlapping windows and frequency correlations. This redundancy introduces linear dependencies at the very first stage of the pipeline, potentially handicapping the model before the first calculation is even performed.
The Complex Clipping Method
To address this, the authors propose Complex Clipping, a regularization technique applied to the input stage of deep networks processing complex-valued signals. In a standard pipeline, the complex STFT is converted to a magnitude-squared representation. Complex Clipping replaces this step by zeroing out specific regions of the complex plane.
As illustrated in the figure above, the technique defines a cone-shaped "support region" in the complex plane (Real vs. Imaginary axes). If a complex value falls inside the cone, its magnitude is retained; otherwise, it is set to zero. This is essentially a 2D, complex-valued analogue to the ReLU function. By aggressively zeroing out roughly half of the complex input space, the method introduces sparsity and reduces the "noise" of linear dependence while maintaining essential physical symmetries.
Experimental Setup
The researchers validated this approach on a challenging real-world task: detecting COVID-19 from cough sounds. The dataset consisted of 5,284 high-quality audio recordings (SNR > 15 dB) collected via phone calls. Each sample was processed into a 4-second segment, and model predictions were compared against gold-standard PCR test results.
One of the most significant findings during preprocessing was that Complex Clipping reduced the largest singular value of the input data by approximately 70%. In linear algebra terms, this indicates that the input matrix became much better conditioned, significantly stabilizing the optimization problem for the neural network.
Results and Discussion
The impact of regularizing the input stage was profound. By "cleaning up" the input and removing redundant linear dependencies, the model's ability to generalize to new data improved dramatically.
As shown in the ROC analysis, the Area Under the Curve (AUC) jumped from 0.77 using a standard spectrogram to 0.93 with the complex clipped version. The authors also noted that the specific orientation of the clipping cone (e.g., rotating it 90 degrees) had negligible impact on performance, suggesting that the primary benefit stems from the act of reducing dependency rather than the specific phase information removed.
Conclusion
Complex Clipping demonstrates that regularization at the "front door" of a model is just as critical as regularization within its hidden layers. By addressing the inherent redundancies of the STFT, this method provides a more stable numerical representation for audio data. While currently validated on acoustic cough detection, this approach holds promise for other time-frequency applications, including music information retrieval, speech recognition, and radio frequency (RF) signal analysis. Future work may explore the theoretical optimality of the cone shape and its integration with modern architectures like Transformers.
Summary updated