A gentle introduction to cooperative learning of Pl@ntNet’s Artificial Intelligence algorithm

Par Benjamin Charlier (Unité MIAT – INRAE Toulouse)

Deep learning models for plant species identification depend on large, annotated datasets. The Pl@ntNet system facilitates global data collection by allowing users to upload and annotate plant observations, though this crowdsourced approach introduces noisy labels due to varying user expertise. Reaching consensus on labels is essential for effective model training, yet the scale of data—across observations, users, and species—makes traditional aggregation methods difficult to apply. Pl@ntNet addresses this with a label aggregation strategy that cooperatively trains plant identification models by estimating user expertise as a trust score. This score reflects each user's ability to correctly identify species. We present a case study on the Pl@ntNet dataset to assess the current aggregation strategy and investigate possible improvements. Given the dataset's complexity—with many tasks, users, and species—our findings show that the proposed approach consistently outperforms existing aggregation methods.