(new) Call for applications
Postdoc position(with fellowship from FAPESP)
* Data efficient methods for plankton image classification *
University of São Paulo (USP), Brazil
Application open until: June 30, 2022
This opportunity is open to applicants of any nationality. The opportunity is linked to a FAPESP sponsored project that is part of a larger transnational project, World Wide Web of Plankton Image Curation (WWW.PIC), promoted by Belmont Forum and with participating teams from France, United States, Brazil and Japan. This is an ongoing project. The postdoc fellow will be part of the Brazilian team, supervised by Prof. N. Hirata, and will be primarily working at Institute of Mathematics and Statistics and occasionally at the Oceanographic Institute, both at University of São Paulo, São Paulo, Brazil.
Research topic description
Modern imaging technologies allow acquisition of in situ plankton images in large scale. Recognizing and estimating their distribution on distinct taxonomic classes is of great importance for better understanding the marine ecosystem. Machine assisted methods for the recognition of these organisms are fundamental for timely processing and analysis of these data.
In this scenario, machine/deep learning techniques emerge as a promising tool. However, the data-hungry nature of deep learning models is a challenge for their effective use, specially considering data that may have distinct characteristics depending on many factors such as season of the year, imaging technology, resolution, location where images are acquired, weather conditions, unknown species, and so on.
The main goal of the postdoc project is to develop machine learning based computational methods to speed up both annotation and classification of plankton images and at the same time minimize the required effort from the expert. That is, we do not wish to simply restart training for each new batch of data nor we would like to rely on experts manually labeling thousands of images for each situation. We would like to be able to effectively reuse previously generated knowledge to quickly produce a classifier adapted to the new imaging or use conditions.
The focus of the research should be on semi-supervised and user-interaction based methods. Clever user interaction to guide the data labeling and machine training processes and also for validating the results are desirable features for the solution. There is no strong constraint of approaches to be investigated as long as an effective solution is targeted. Developed methods will be integrated into EcoTaxa. As this project will be within an international collaboration, we expect to have access to distinct batch of images, coming from different equipment and imaging conditions.