Some models in AddaxAI include a feature called taxonomic fallback, a strategy to improve classification accuracy when species-level predictions are uncertain. This post explains how it works and how you can use it effectively in your workflow.
What is taxonomic fallback
Taxonomic fallback is a method that adjusts the prediction level based on confidence. The model first tries to identify the species, but if the confidence is too low, it falls back to a higher-level taxon such as genus, family, or order. This fallback continues until a prediction meets the confidence threshold you define.
Example:
- Model predicts:
- Panthera leo (lion): 30%
- Panthera pardus (leopard): 50%
- Combined confidence for genus Panthera: 80%
- If your confidence threshold is 75%, the model will return genus Panthera, skipping the low-confidence species predictions
- If your threshold is lower, for example 40%, it will return Panthera pardus
This avoids vague results like “unidentified animal” while still giving meaningful biological insight when species-level certainty is low.
How it works
- Aggregation: Model predictions are summed across shared taxonomic groups, such as all species within the same genus.
- Fallback logic: If no species-level prediction meets the threshold, the model returns a broader label with higher confidence.
- Customizable settings:
- You can enable or disable fallback mode via a checkbox in advanced settings.
- You can fix the prediction level (for example, always predict at species, genus, or family level).
- Some models allow filtering by training sample size (for example, only predict categories trained with more than 10,000 images).
Why use it
- Improved accuracy: Broader categories like genus or family tend to be more reliable.
- Better usability: Useful when working with rare or underrepresented species.
- More informative than “unknown”: You still get useful information even when the exact species is unclear.
Working with mixed taxonomic levels
Many real-world models contain labels at mixed taxonomic levels, such as “bird”, “rodent”, and Panthera leo in the same model. In these cases, addaxai will return the most specific available level unless you choose otherwise.
If you select a fixed prediction level such as species or family, the model will adapt. For example:
- If a label in the model is simply “rodent” and you request species-level predictions, the model will return order Rodentia instead of forcing an incorrect species label.