BioEncoder: A metric learning toolkit for comparative organismal biology

Area of Science:

• Biological image analysis
• Deep learning
• Biodiversity informatics

Background:

• Conventional deep learning (DL) methods struggle with large biodiversity datasets.
• Challenges include unbalanced classes and subtle phenotypic differences.
• Existing DL toolkits may not be flexible enough for diverse biological data.

Purpose of the Study:

• Introduce BioEncoder, a user-friendly toolkit for deep metric learning.
• Overcome limitations of conventional DL in biological image analysis.
• Facilitate practical applications of advanced deep learning in biology.

Main Methods:

• BioEncoder employs metric learning to focus on relationships between data points.
• It offers taxon-agnostic data loaders for diverse datasets.
• Features include custom augmentation and simple hyperparameter tuning via configuration files.

Main Results:

• BioEncoder provides a flexible and user-friendly approach to deep metric learning.
• The toolkit is designed for ease of use across various biological image datasets.
• It enables learning from data relationships rather than strict class separability.

Conclusions:

• BioEncoder democratizes access to advanced deep metric learning techniques.
• The toolkit bridges the gap between complex DL pipelines and biological research applications.
• It has the potential to unlock new research avenues in biological image analysis.