Listen to Interpret: Post-hoc Interpretability for Audio Networks with NMF

Abstract

This paper tackles post-hoc interpretability for audio processing networks. Our goal is to interpret decisions of a network in terms of high-level audio objects that are also listenable for the end-user. To this end, we propose a novel interpreter design that incorporates non-negative matrix factorization (NMF). In particular, a carefully regularized interpreter module is trained to take hidden layer representations of the targeted network as input and produce time activations of pre-learnt NMF components as intermediate outputs. Our methodology allows us to generate intuitive audio-based interpretations that explicitly enhance parts of the input signal most relevant for a network’s decision. We demonstrate our method’s applicability on popular benchmarks, including a real-world multi-label classification task.

An Explainable Proxy Model for Multilabel Audio Segmentation

“…with a two-step optimization process [8]. In this work, we consider the sparse NMF implementation [24] as suggested in [7]. In our segmentation proxy model, the dictionary W is pre-learned while the activation H is extracted by a neural model Ψ and referred to as an embedding.…”

“…In this work, the f model is pre-trained with frozen weights and serves as a teacher for the proxy model. We use a similar approach as [7] where the proxy model is composed of two functions. Let Ψ be a function that maps a sequence of D-dimension feature vectors S ∈ R D×T to the embedding H ∈ R K×T…”

“…Recently, a few explainable models have been developed like APNet [5], which extends the training of prototypes to the audio domain, and post-hoc visualization of explanations obtained from Shapley values [6]. In the architecture proposed in [7], the authors explain a black box audio classifier with a proxy model which is optimized to classify audio scenes while reconstructing the audio with the nonnegative matrix factorization framework [8].…”

“…We propose to train an explainable proxy model from a pretrained multilabel segmentation model (designated as the teacher). The architecture is inspired by [7]. The proxy is trained following a teacher-student approach, commonly used in knowledge distillation [21].…”

“…The former inputs a spectrogram and the latter uses the teacher's Wavlm outputs. Contrary to [7], we consider frame-level, i.e. time segmentation, instead of utterance-level classification.…”