Reliable Local Explanations for Machine Listening

Saumitra Mishra, Emmanouil Benetos, Bob L. Sturm, Simon DixonIJCNN 2020 · OpenReview

Abstract

One way to analyse the behaviour of machine learning models is through local explanations that highlight input features that maximally influence model predictions. Sensitivity analysis, which involves analysing the effect of input perturbations on model predictions, is one of the methods to generate local explanations. Meaningful input perturbations are essential for generating reliable explanations, but there exists limited work on what such perturbations are and how to perform them. This work investigates these questions in the context of machine listening models that analyse audio. Specifically, we use a state-of-the-art deep singing voice detection (SVD) model to analyse whether explanations from SoundLIME (a local explanation method) are sensitive to how the method perturbs model inputs. The results demonstrate that SoundLIME explanations are sensitive to the content in the occluded input regions. We further propose and demonstrate a novel method for quantitatively identifying suitable content type(s) for reliably occluding inputs of machine listening models. The results for the SVD model suggest that the average magnitude of input mel-spectrogram bins is the most suitable content type for temporal explanations.

Introduction

As deep neural networks become the standard for "machine listening" tasks like audio analysis, their "black-box" nature presents a challenge for high-stakes applications. To build trust, researchers employ local explanations, which identify specific input features—such as a time segment or frequency band—that drive a model's prediction.

A prominent method is SoundLIME (SLIME), which generates explanations by "occluding" (masking) parts of an audio spectrogram and observing changes in the model's output. However, a fundamental question remains: How should audio be "hidden"? Unlike computer vision, where a pixel can be replaced with grey, "silence" in audio is ambiguous, especially in log-scaled spectrograms where zero might not represent the absence of signal. This study investigates how occlusion strategies and sampling density affect the reliability of these explanations.

Methodology

The researchers evaluated SVDNet, a state-of-the-art Convolutional Neural Network (CNN) for Singing Voice Detection (SVD). The SVD task involves classifying audio segments as containing a singing voice (V) or not (NV).

The SVD algorithm segments input audio into vocal (V) and non-vocal (NV) regions based on ground truth labels.

The study analyzed SLIME through three lenses:

  1. Stability Analysis: Determining the number of synthetic samples ($N_s$) required for consistent explanations.
  2. Sensitivity Analysis: Testing five occlusion methods: setting values to zero, using dataset-wide minimums ($min_{data}$), input-specific minimums ($min_{inp}$), input-specific means ($mean_{inp}$), or Gaussian noise ($N_g$).
  3. Reliability Framework: A novel quantitative approach using audio with known ground-truth labels (from the ccMixter dataset) to measure which occlusion method produces explanations most aligned with the actual target signal.

Interpretable Components and Occlusion

SLIME breaks down the input spectrogram into "interpretable components." These can be temporal segments (vertical slices) or frequency bands (horizontal slices).

Interpretable components for SLIME: (A) temporal segments and (B) frequency bands.

The choice of how to mask these components is critical. The "Zero Trap" occurs when researchers default to zeroing out features. In log-mel-spectrograms, a zero value can be interpreted by the model as a meaningful feature rather than silence, leading to misleading explanations.

Key Results

  • Stability Scale: While previous literature suggested 5,000 samples might be enough, this study found that complex audio models require approximately 50,000 synthetic samples to achieve stable explanations.
  • High Sensitivity: Explanations were found to be highly fluid. Different occlusion types often resulted in zero overlap between the top-3 identified features for the same audio clip, suggesting that masking strategies can inadvertently bias results.
  • Optimal Occlusion ($mean_{inp}$): The reliability tests identified the average magnitude of the input ($mean_{inp}$) as the most effective strategy. It matched ground truth perfectly in 34% of cases and correctly identified at least two of the top three influential components in over 80% of tests.

Conclusion

This work highlights that local explanations for machine listening are highly sensitive to the choice of occlusion content. For the SVDNet model, using the input's mean magnitude proved most reliable. While these specific findings are tied to the SVD task, the sensitivity to occlusion is likely a general property of local explanation methods. Future research will explore if these patterns hold for other architectures and tasks, such as acoustic scene detection or genre classification.

Summary updated

Abstract

One way to analyse the behaviour of machine learning models is through local explanations that highlight input features that maximally influence model predictions. Sensitivity analysis, which involves analysing the effect of input perturbations on model predictions, is one of the methods to generate local explanations. Meaningful input perturbations are essential for generating reliable explanations, but there exists limited work on what such perturbations are and how to perform them. This work investigates these questions in the context of machine listening models that analyse audio. Specifically, we use a state-of-the-art deep singing voice detection (SVD) model to analyse whether explanations from SoundLIME (a local explanation method) are sensitive to how the method perturbs model inputs. The results demonstrate that SoundLIME explanations are sensitive to the content in the occluded input regions. We further propose and demonstrate a novel method for quantitatively identifying suitable content type(s) for reliably occluding inputs of machine listening models. The results for the SVD model suggest that the average magnitude of input mel-spectrogram bins is the most suitable content type for temporal explanations.
BibTeX
@inproceedings{mishra2020reliable,
  author = {Saumitra Mishra and Emmanouil Benetos and Bob L. Sturm and Simon Dixon},
  title = {{Reliable Local Explanations for Machine Listening}},
  booktitle = {IJCNN},
  year = {2020},
  url = {https://openreview.net/forum?id=JIN8Oznl0L},
}
Figure