EMDS-7-FSCIL: a benchmark for Few-Shot Class-Incremental Learning in environmental microorganism recognition.

Area of Science:

• Microbiology
• Computer Science
• Machine Learning

Background:

• Deep learning models struggle with recognizing new environmental microorganisms due to data scarcity and high annotation costs.
• The plasticity-stability dilemma hinders incremental learning in dynamic environments.
• A dedicated benchmark for Few-Shot Class-Incremental Learning (FSCIL) in environmental microbiology is currently lacking.

Purpose of the Study:

• To establish the first FSCIL benchmark for environmental microorganism recognition.
• To propose a unified evaluation protocol for assessing FSCIL methods on microbial datasets.
• To provide a reproducible platform for comparing and advancing FSCIL techniques in this domain.

Main Methods:

• Development of the first FSCIL benchmark for environmental microorganism recognition using the EMDS-7 dataset.
• Systematic reproduction and comparative evaluation of 10 representative FSCIL methods.
• Comprehensive performance analysis using metrics like per-session accuracy, average accuracy, and performance drop rate.

Main Results:

• SAVC and FACT demonstrated the highest overall accuracy among the evaluated FSCIL methods.
• PFR showed more stable performance but with a lower accuracy ceiling.
• CLOSER and BiDist exhibited significantly weaker performance, highlighting method-specific limitations.
• FSCIL methods successful on general image benchmarks do not directly translate to environmental microorganism recognition.

Conclusions:

• The developed benchmark and evaluation protocol are crucial for fair comparison of FSCIL methods in environmental microbiology.
• Task-specific adaptations are necessary for existing FSCIL methods to effectively recognize environmental microorganisms.
• This foundational work will accelerate future research in FSCIL for microbial identification and monitoring.