Training neural audio classifiers with few data

TL;DR

This paper explores various supervised learning strategies, including regularization, prototypical networks, and transfer learning, to improve neural audio classifiers trained on small datasets, demonstrating transfer learning's effectiveness and prototypical networks' promise.

Contribution

It systematically evaluates the effectiveness of regularization, prototypical networks, and transfer learning for small dataset audio classification tasks.

Findings

Transfer learning significantly improves performance on small datasets.

Prototypical networks perform well without external data.

Regularization alone offers limited benefits.

Abstract

We investigate supervised learning strategies that improve the training of neural network audio classifiers on small annotated collections. In particular, we study whether (i) a naive regularization of the solution space, (ii) prototypical networks, (iii) transfer learning, or (iv) their combination, can foster deep learning models to better leverage a small amount of training examples. To this end, we evaluate (i-iv) for the tasks of acoustic event recognition and acoustic scene classification, considering from 1 to 100 labeled examples per class. Results indicate that transfer learning is a powerful strategy in such scenarios, but prototypical networks show promising results when one does not count with external or validation data.