Adaptive Metric Learning for Acoustic Data

Project Type: Classification
Area/Sector: Acoustic
Company/Partner: Meridian
Academic Institution: Dalhousie University
Lead Investigator: Dr. Stan Matwin
HQP: Xiang Jiang, Dr. Oliver S. Kirsebom

DeepSense and Meridian are working together to developa model trained on a sample-sufficient domain and afterwards adapt it to the sample-deficient acoustic domain.

This project will develop a sample-efficient algorithm for few-to-medium-shot image classification by first meta-learning the representation from the source domain and adapting the representation to the target domain. The advantage of this algorithm is twofold: 1) the model learns the representation from data-abundant tasks using natural images; 2) the model maintains its compatibility towards the acoustic data classification task. Therefore, the trained model will be capable of classifying different species of marine animals with a few labeled acoustic samples.