Naturally, animals exhibit activities in an imbalanced manner. For example, lions spend a lot of their time resting. When the data annotation is performed chronologically, this results in an equally imbalanced labeled dataset. Data from the majority classes is often left out during the classification to improve the balance between the activity classes. Discarding labeled data is wasteful -- why label that data if it is discarded later?
Active learning balances the dataset during the data annotation process and prevent wasteful work on labeling. Active learning promises to not only obtain a more balanced dataset but also speed up the process. Active learning aims to select raw data that is dissimilar to the data that is already labeled. For example, when most of the data consists of the animal grazing, then it might be beneficial to have an active learner select data that is different from grazing and query the annotator for a label, opposed to manually scanning all the footage to find other behavior.
The goal of this project is to investigate active learning strategies in an existing labeling application using IMU data from animals. An AI classifier needs to be implemented that runs in the background during the labeling process. It is up to the student to design an intelligent labeling application.
20% Theory, 60% Simulations, 20%Writing
Jacob Kamminga, firstname.lastname@example.org, room ZI5011