Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Characterizing Continual Learning Scenarios and Strategies for Audio Analysis

Ruchi Bhatt, Pratibha Kumari, Dwarikanath Mahapatra, Abdulmotaleb El Saddik, Mukesh Saini

TL;DR

The paper systematically evaluates continual learning for audio-based monitoring by constructing domain-incremental and class-incremental sequences from the DCASE datasets and benchmarking a wide range of CL and non-CL baselines. It provides a structured evaluation framework, reveals that replay-based methods (notably Replay) consistently outperform others in both DI and CI settings, and highlights the differing strengths of regularization versus rehearsal approaches across scenarios. The study contributes a public, scenario-rich dataset design and practical insights for deploying adaptive audio monitoring systems in non-stationary environments. The findings have direct implications for real-world surveillance and anomaly detection where distribution shifts and new classes emerge over time.

Abstract

Audio analysis is useful in many application scenarios. The state-of-the-art audio analysis approaches assume the data distribution at training and deployment time will be the same. However, due to various real-life challenges, the data may encounter drift in its distribution or can encounter new classes in the late future. Thus, a one-time trained model might not perform adequately. Continual learning (CL) approaches are devised to handle such changes in data distribution. There have been a few attempts to use CL approaches for audio analysis. Yet, there is a lack of a systematic evaluation framework. In this paper, we create a comprehensive CL dataset and characterize CL approaches for audio-based monitoring tasks. We have investigated the following CL and non-CL approaches: EWC, LwF, SI, GEM, A-GEM, GDumb, Replay, Naive, Cumulative, and Joint training. The study is very beneficial for researchers and practitioners working in the area of audio analysis for developing adaptive models. We observed that Replay achieved better results than other methods in the DCASE challenge data. It achieved an accuracy of 70.12% for the domain incremental scenario and an accuracy of 96.98% for the class incremental scenario.

Characterizing Continual Learning Scenarios and Strategies for Audio Analysis

TL;DR

The paper systematically evaluates continual learning for audio-based monitoring by constructing domain-incremental and class-incremental sequences from the DCASE datasets and benchmarking a wide range of CL and non-CL baselines. It provides a structured evaluation framework, reveals that replay-based methods (notably Replay) consistently outperform others in both DI and CI settings, and highlights the differing strengths of regularization versus rehearsal approaches across scenarios. The study contributes a public, scenario-rich dataset design and practical insights for deploying adaptive audio monitoring systems in non-stationary environments. The findings have direct implications for real-world surveillance and anomaly detection where distribution shifts and new classes emerge over time.

Abstract

Audio analysis is useful in many application scenarios. The state-of-the-art audio analysis approaches assume the data distribution at training and deployment time will be the same. However, due to various real-life challenges, the data may encounter drift in its distribution or can encounter new classes in the late future. Thus, a one-time trained model might not perform adequately. Continual learning (CL) approaches are devised to handle such changes in data distribution. There have been a few attempts to use CL approaches for audio analysis. Yet, there is a lack of a systematic evaluation framework. In this paper, we create a comprehensive CL dataset and characterize CL approaches for audio-based monitoring tasks. We have investigated the following CL and non-CL approaches: EWC, LwF, SI, GEM, A-GEM, GDumb, Replay, Naive, Cumulative, and Joint training. The study is very beneficial for researchers and practitioners working in the area of audio analysis for developing adaptive models. We observed that Replay achieved better results than other methods in the DCASE challenge data. It achieved an accuracy of 70.12% for the domain incremental scenario and an accuracy of 96.98% for the class incremental scenario.
Paper Structure (31 sections, 4 equations, 6 figures, 7 tables)

This paper contains 31 sections, 4 equations, 6 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Continual learning
  3. Continual learning scenarios
  4. Domain-incremental scenario
  5. Class-incremental scenario
  6. Continual learning techniques
  7. Continual Learning for Audio Analysis
  8. Baselines
  9. Continual learning baselines
  10. Elastic Weight Consolidation
  11. Learning without Forgetting
  12. Synaptic Intelligence
  13. Gradient Episodic Memory
  14. Averaged Gradient Episodic Memory
  15. Greedy sampler and Dumb learner
  16. ...and 16 more sections

Figures (6)

  • Figure 1: Overview of a continual learning pipeline. This figure shows the available possible CL scenarios, the data preparation for different scenarios in subsets called tasks, broad categories of CL strategies based on literature, the well-known metrics used to evaluate the performances, and different ways of comparative analysis.
  • Figure 2: Benchmarking dataset for DI scenario. Here the CL model aims to distinguish between normal and abnormal machine sounds in domain shift conditions.
  • Figure 3: Benchmarking dataset for CI scenario. Here the classification model aims to distinguish among machine types in class incremental fashion. Here only normal sound of considered machine types is available.
  • Figure 4: The train-test matrix for T=5.
  • Figure 5: Comparison of different approaches in terms of average accuracy over all episodes after each training session in DI scenario.
  • ...and 1 more figures