Explaining siamese Networks in few-shot learning for audio data

Machine learning models are not able to generalize correctly when queried on samples belonging to class distributions that were never seen during training. This is a critical issue, since real world applications might need to quickly adapt without the necessity of re-training. To overcome these limitations, few-shot learning frameworks have been proposed and their applicability has been studied widely for computer vision tasks. Siamese Networks learn pairs similarity in form of a metric that can be easily extended on new unseen classes. Unfortunately, the downside of such systems is the lack of explainability. We propose a method to explain the outcomes of Siamese Networks in the context of few-shot learning for audio data. This objective is pursued through a local perturbation-based approach that evaluates segments-weighted-average contributions to the final outcome considering the interplay between different areas of the audio spectrogram. Qualitative and quantitative results demonstrate that our method is able to show common intra-class characteristics and erroneous reliance on silent sections.