Open Set Audio Classification Using Autoencoders Trained on Few Data.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Signal Processing

Background:

• Open-set recognition (OSR) challenges classifiers with unseen classes during training.
• Few-shot learning (FSL) addresses limited positive samples in recognition systems.
• A new audio dataset facilitates research in combined OSR and FSL.

Purpose of the Study:

• To propose and evaluate an audio system addressing both open-set recognition and few-shot learning challenges.
• To develop a robust method for identifying known audio classes and rejecting unknown samples.
• To investigate the system's performance under varying conditions of class openness and sample availability.

Main Methods:

• A three-step approach involving high-level audio representation and feature embedding.
• Utilizing two distinct autoencoder architectures for feature extraction.
• Employing a multi-layer perceptron (MLP) on latent space representations for classification and rejection.

Main Results:

• The proposed audio OSR/FSL system demonstrated validity across extensive experiments.
• Superior performance was confirmed compared to a baseline transfer learning system.
• The system effectively handled multiple combinations of openness factors and few-shot conditions.

Conclusions:

• The developed system offers a promising solution for audio open-set recognition and few-shot learning.
• The autoencoder and MLP combination proves effective for distinguishing known from unknown audio classes.
• This work advances the capabilities of audio recognition systems in practical, data-scarce, and open-world scenarios.