Journal
AI
Volume 4, Issue 3, Pages 574-597Publisher
MDPI
DOI: 10.3390/ai4030031
Keywords
applied machine learning; few-shot learning; in-the-wild data processing; ecology applications
Ask authors/readers for more resources
This paper explores the effectiveness of few-shot learning in a real-world problem where labels are hard to obtain. By training an FSL network using large public datasets and only labeling a few images per new species, the paper achieves the classification of various animal species in the wild. The paper also discusses the challenges and constraints posed by uncurated data and evaluates the potential real-world usefulness of different FSL networks.
Few-shot learning (FSL) describes the challenge of learning a new task using a minimum amount of labeled data, and we have observed significant progress made in this area. In this paper, we explore the effectiveness of the FSL theory by considering a real-world problem where labels are hard to obtain. To assist a large study on chimpanzee hunting activities, we aim to classify various animal species that appear in our in-the-wild camera traps located in Senegal. Using the philosophy of FSL, we aim to train an FSL network to learn to separate animal species using large public datasets and implement the network on our data with its novel species/classes and unseen environments, needing only to label a few images per new species. Here, we first discuss constraints and challenges caused by having in-the-wild uncurated data, which are often not addressed in benchmark FSL datasets. Considering these new challenges, we create two experiments and corresponding evaluation metrics to determine a network's usefulness in a real-world implementation scenario. We then compare results from various FSL networks, and describe how factors may affect a network's potential real-world usefulness. We consider network design factors such as distance metrics or extra pre-training, and examine their roles in a real-world implementation setting. We also consider additional factors such as support set selection and ease of implementation, which are usually ignored when a benchmark dataset has been established.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available