I spent 6 months training my AI model on bad data before I figured it out

I was building a chatbot for my gardening club in Portland and kept getting weird plant identification results. Turns out I accidentally included a bunch of photos of fake plastic plants from a craft store in my training set. Has anyone else had their model give totally off-base answers because of bad input data?

2 comments

2 Comments

nina_johnson861mo ago

Good catch on realizing the issue, but fake plastic plants and real plants actually do share some visual features. Those craft store photos weren't totally useless, they probably just confused your model more than they helped.

west.anna1mo ago

Ugh, this is so spot on. A buddy of mine tried training a model on fake aquarium plants and ended up with a model that flagged every single silk flower arrangement as "real." It was a mess. So yeah, @nina_johnson86 is right - the shared shapes mess things up even if the textures are different.