Hi, everyone.
I’m working on a project that involves fine-grained classification of ~125 plant species. Unfortunately, I haven’t been able to get a sufficient amount of clean image data to get a model to perform as well as I would like. (I seem to be bumping a ceiling at about 75% accuracy.)
Many of the errors occur when the model predicts a closely related species instead of the correct one. In these cases, I often see that the correct species was nevertheless given a fairly high probability (say, 0.1 instead of the usual .001).
I’d like to use these not-quite-zero predictions to offer multiple “best guesses” to the user. But, I’m not sure how to select them. An arbitrary cutoff might work, but the size of these “next-best” predictions fluctuates depending on how confident the model was in the winning result. Any ideas?