Ahmet was the first PhD candidate associated with SFI CRIMAC (external link) to defend his thesis. The thesis is dedicated to development of machine learning methods for analysis of acoustic data.

Acoustic data from echo sounders are widely used to study fish stocks. These data are large and complex, and analysis is still mostly manual.  Ahmet's thesis addresses the automation of fish detection in such data using deep learning.

 

Self-supervised model for feature learning
The image shows a self-supervised model for learning features in acoustic data, for details see <a href="https://doi.org/10.1016/j.ecoinf.2024.102878 (external link)">Pala et al, 2024</a>. Photo: an open access article distributed under the terms of the Creative Commons CC-BY license,https://doi.org/10.1016/j.ecoinf.2024.102878 (external link)

 

A central challenge is extreme data imbalance. Fish schools appear in only a small fraction of the data, while the majority of data correspond to empty water column, seabed, and other non-fish objects. This makes both supervised and unsupervised learning difficult. The thesis explores methods to mitigate this imbalance, including improved sampling strategies.