Reassessing deep learning (and meta-learning) computer vision as an efficient method to determine taphonomic agency...

Area of Science:

• Paleontology
• Archaeology
• Computer Science

Background:

• Taphonomic research reconstructs preservation and modification processes of paleobiological remains.
• Recent critiques questioned the reliability of deep learning (DL) for analyzing bone surface modifications (BSMs) due to dataset limitations and potential model bias.
• These critiques may have overlooked how small, unbalanced datasets can lead to underfit models and coding errors can introduce bias.

Purpose of the Study:

• To reassess the efficiency and resolution of DL in taphonomic research.
• To address critiques regarding DL model reliability for BSM analysis.
• To investigate potential biases and overfitting in DL models applied to taphonomic datasets.

Main Methods:

• Replicated original DL models as baseline models for comparison.
• Developed optimized DL models to address issues of poor-quality images and validation set overfitting.
• Implemented enhanced image data augmentation, k-fold cross-validation, and few-shot learning (supervised and model-agnostic meta-learning) for robust validation.

Main Results:

• Replicated DL models showed consistent and comparable outcomes to original baseline models.
• Optimized models yielded virtually invariant results, even with recently generated BSM images as test sets.
• Findings reinforce the efficacy of the original DL models in differentiating BSM, refuting claims of methodological overfitting.

Conclusions:

• The original DL models demonstrate nuanced efficacy in differentiating BSM, despite dataset limitations.
• Critiques regarding DL reliability in taphonomy may stem from methodological oversights in handling small, unbalanced datasets.
• Future research with larger, higher-quality datasets can further enhance DL model generalization and reliability in taphonomic studies.