The day I found out I'd been training my AI model on garbage data for 6 months

I run a small crew doing painting jobs, but I also mess around with AI on the side. Got this idea to train a model to spot drywall cracks from photos. Spent 6 months feeding it thousands of images I took myself with my phone. Thought I was being clever. Then a buddy who actually knows this stuff looked at my dataset and laughed. He said half my "crack" photos were just shadows from bad lighting. The model was learning shadows, not cracks. That's when it hit me. How do you even know if your training data is any good before you waste half a year?

2 comments

2 Comments

andrewt4110d agoMost Upvoted

Six months is nothing, you probably learned more than if it worked first try.

abbyc3310d ago

Yeah @andrewt41 I used to think that whole "garbage in, garbage out" thing was just tech people being dramatic... but now I get it. I always figured as long as you had lots of data it would figure itself out somehow, like the AI would just magically know what mattered. But that story about the shadows really hit home for me. It's not even about quantity, it's about whether the data actually represents what you're trying to teach it. Makes me realize I probably have blind spots in my own thinking about this stuff too.