Semi-supervised learning for enhancing rare species birdsong recognition

Read also: Writing a Master's Thesis in Language Technology or Computational Creativity

The project will be carried out in collaboration with Benjamin Cretois at the Norwegian Institute for Nature Research (NINA). NINA is Norway’s leading institution for applied ecological research, with broad-based expertise on the genetic, population, species, ecosystem and landscape level, in terrestrial, freshwater and coastal marine environments. The student(s) will be collaborating with researchers at the Miljødata department who focus on employing advanced technologies to study and protect biodiversity, with a particular emphasis on bioacoustics and bird species.

Problem Description: 

NINA are currently using an off-the-shelf deep learning-based bird classifier (BirdNET) that effectively identifies common bird species. However, its performance significantly drops for the rarer species, limiting our understanding of biodiversity in diverse ecosystems. Properly accounting for rare species is critical to establish ecosystem health and it is essential to develop automatic methods efficient at detecting and classifying them.

The project aims to refine the model using advanced machine learning techniques such as fine-tuning and few-shot learning strategies, potentially exploring other innovative methods to enhance species recognition, particularly for underrepresented species. The expected outcome includes improved model accuracy and recall for rare species, contributing to more comprehensive biodiversity assessments.

Data: 

The student(s) will have access to the Sound of Norway (https://thesoundofnorway.com) dataset. It contains approximatively 7 terabyte of audio files recorded by BUGG recording devices at 41 different sites throughout Norway. All of the detections made by BirdNET have been verified by an ornithologist. Moreover, a few files have completely been reviewed by the ornithologist to compute the model recall. This dataset should provide a robust foundation for evaluation of the methods that will be developed.

Risks / Challenges:
Improving model capabilities to generalise well to rare events (rare bird calls in the case of this subject) is a hot topic and numerous methods are being developed. The primary scientific challenge lies in the method's ability to generalise well to rare species with limited examples.