Fast threshold optimization for multi-label audio tagging using Surrogate gradient learning

TL;DR

This paper introduces SGL-Thresh, a gradient-based method for automatically optimizing decision thresholds in multi-label audio tagging, significantly improving F1 scores on multiple datasets using pre-trained neural networks.

Contribution

The paper presents a novel surrogate gradient learning approach for threshold optimization, enabling fast and scalable F1 maximization in multi-label audio classification.

Findings

SGL-Thresh outperforms baseline and heuristic methods in F1 score.

Achieved 54.9% F1 on AudioSet, surpassing 50.7% with default thresholds.

Method is fast, scalable, and applicable to large tag sets.

Abstract

Multi-label audio tagging consists of assigning sets of tags to audio recordings. At inference time, thresholds are applied on the confidence scores outputted by a probabilistic classifier, in order to decide which classes are detected active. In this work, we consider having at disposal a trained classifier and we seek to automatically optimize the decision thresholds according to a performance metric of interest, in our case F-measure (micro-F1). We propose a new method, called SGL-Thresh for Surrogate Gradient Learning of Thresholds, that makes use of gradient descent. Since F1 is not differentiable, we propose to approximate the thresholding operation gradients with the gradients of a sigmoid function. We report experiments on three datasets, using state-of-the-art pre-trained deep neural networks. In all cases, SGL-Thresh outperformed three other approaches: a default threshold value (defThresh), an heuristic search algorithm and a method estimating F1 gradients numerically. It reached 54.9\% F1 on AudioSet eval, compared to 50.7% with defThresh. SGL-Thresh is very fast and scalable to a large number of tags. To facilitate reproducibility, data and source code in Pytorch are available online: https://github.com/topel/SGL-Thresh