Our image AI training dragged on way past the planned finish line

We built a system to spot defects in photos and gave it a two-week training window. But the model kept needing more tweaks and data, which added another month. Now we're scrambling to meet client deadlines because of that hold-up. How do you keep AI projects from running over time?

2 comments

Log in to join the discussion

Log In

2 Comments

wesley7617mo ago

My buddy tried training a model to count apples in orchard photos. He figured a week of training would do it. The thing worked great on his test set from one farm, but completely failed when they tried it on a different orchard with different lighting. They spent three more weeks just collecting pictures from like six other farms. Real world stuff always messes with the clean data.

4

james_allen247mo ago

Wait what, @wesley761? It totally failed just because the lighting was different? That's wild. I can see it maybe struggling between a sunny day and a rainy one, but you're telling me a model trained on, like, Fuji apples in California sun couldn't spot a Gala apple in Oregon morning fog? It makes you realize all the tiny stuff we don't even see. The angle of the trees, the color of the dirt, everything. That's brutal.

2