Realized I was feeding AI wrong data for 3 months straight

I was training a little model to sort transmission parts at my shop, kept getting junk results. Spent like 3 months wondering why it thought a torque converter was a brake rotor. Finally noticed my training images had bad lighting and weird angles from my phone camera, not the overhead setup I use for work. Total facepalm moment. Anybody else have a stupid simple fix that took forever to notice?

3 comments

3 Comments

victorb171mo ago

Read somewhere that a guy training a model to spot bad welds had the same problem. Turns out he was using photos from his phone too, but the lighting in his shop was totally different from his desk setup. I mean, it's crazy how small things like that mess up the whole thing. Like the AI just learns "oh this is what the world looks like" from your training data, not what you actually want it to do. Idk, maybe it's just me but I feel like everyone runs into this at some point if they're doing stuff from scratch.

skyler431mo ago

Yeah that welding story hits a little too close to home honestly. I once spent a week training a model to tell the difference between my cat and a rug, and it just learned to look for sunlight because all the good cat photos were taken near a window. AI is basically that friend who overanalyzes everything and still gets it wrong because you forgot to mention the obvious stuff.

karen_perry3810d ago

Oh come on. That's just bad dataset work, not some deep AI mystery. I've trained loads of models on phone photos with no issue because I actually hold the lighting constant. If your cat model learned sunlight, that's on you for not mixing in dim shots of the rug. People blame the AI for being dumb when they're the ones feeding it garbage. It's like complaining your calculator gave you the wrong answer because you typed in 2+2 and hit multiply by accident. The welding story? Guy should've taken shop pics at the same time of day. Not the model's fault.

-1