Spectroscopy Approaches for Food Safety Applications: Improving Data Efficiency Using Active Learning and...

Area of Science:

• Food Science and Technology
• Analytical Chemistry
• Data Science

Background:

• Spectroscopy is crucial for analyzing food quality, safety, and nutrition, but predicting complex properties is hindered by data limitations.
• Machine Learning (ML) offers potential for improved food property prediction, yet large labeled datasets are a persistent barrier.
• Developing data-efficient ML models is essential for advancing rapid and simple food analysis.

Purpose of the Study:

• To explore and evaluate data annotation and model training strategies for enhancing ML data efficiency in spectroscopy-based food analysis.
• To investigate the effectiveness of Active Learning (AL) and Semi-Supervised Learning (SSL) in reducing labeled data requirements.
• To compare a hybrid AL-SSL approach against baseline passive learning and individual AL/SSL methods.

Main Methods:

• Implemented and compared four ML approaches: baseline passive learning, Active Learning (AL), Semi-Supervised Learning (SSL), and a hybrid AL-SSL model.
• Utilized two distinct spectroscopy datasets for evaluation: one for predicting plasma dosage and another for detecting foodborne pathogens.
• Focused on assessing the reduction in the number of required labeled samples for model training.

Main Results:

• Advanced ML approaches (AL, SSL, and hybrid) demonstrated significant improvements in data efficiency compared to passive learning.
• These methods substantially reduced the number of labeled samples needed for accurate model training.
• In specific cases, the number of labeled samples was decreased by more than half, highlighting substantial data savings.

Conclusions:

• Active Learning and Semi-Supervised Learning offer powerful solutions to the data bottleneck in ML applications for food spectroscopy.
• The hybrid AL-SSL approach shows promise for maximizing data efficiency in complex food analysis tasks.
• These findings pave the way for more accessible and cost-effective spectroscopic analysis in food science by reducing reliance on extensive labeled datasets.