[D] Why I abandoned YOLO for safety critical plant/fungi identification. Closed-set classification is a silent failure mode
A 94% accurate plant ID model was abandoned because its architecture couldn't say 'I don't know'.
A developer building an open-source, on-device tool for identifying edible and toxic plants and fungi has publicly detailed why they abandoned popular YOLO (You Only Look Once) object detection models. Despite achieving 94-96% accuracy on target species using iNaturalist data, they discovered a potentially lethal flaw for safety-critical applications: YOLO's closed-set architecture cannot express uncertainty. When presented with an unknown plant (out-of-distribution data), the model is forced by its softmax function to assign 100% confidence to one of its known classes, offering no 'none of the above' option. This silent failure mode is catastrophic for foraging, where a misidentified toxic mushroom could be fatal.
To solve this, the developer engineered a multi-layered, open-set detection pipeline constrained by the Hailo 8L accelerator's 13 TOPS budget for battery-powered use. The new system employs a MobileNetV3 'router' to direct images to one of three EfficientNet B2 specialist models (e.g., for fungi or berries) or reject them upfront. Crucially, it uses 'energy scoring' on raw model logits—a technique from recent AI research—to cleanly separate known from unknown inputs, and incorporates ensemble disagreement and a dedicated 'unknown' class in each specialist. This case highlights a critical, under-discussed limitation of many production AI systems where overconfidence on novel data poses real-world risks.
- YOLO's closed-set design forces 100% confidence misclassifications on unknown inputs, a fatal flaw for safety apps.
- The new pipeline uses energy scoring on logits for OOD detection, a major improvement over softmax confidence thresholds.
- The entire system must run on a Hailo 8L chip with 13 TOPS, prioritizing real inference latency over pure accuracy.
Why It Matters
This exposes a critical blind spot in AI deployment: models that can't express uncertainty are dangerous for real-world, safety-critical decision-making.