That time a dev meetup in Austin showed me a better way to train models

I was at this small AI meetup in Austin last month where some guy demoed his custom image classifier using transfer learning on a tiny dataset. He got 95% accuracy with only 50 images by fine-tuning a pretrained model instead of starting from scratch. Has anyone else messed with transfer learning for specific niche projects and seen big improvements?

2 comments

2 Comments

danielmason27d ago

Transfer learning is wild for niche stuff. I trained a model to spot different types of wood rot on houses with maybe 30 pics of each kind and it nailed it way better than I expected. Saved me from having to hunt down thousands of images nobody has.

troythompson26d ago

I mean I get what you're saying @danielmason and transfer learning is definitely cool, but I feel like people oversell how easy it is. Like with wood rot, sure you got it working with 30 pics, but what if someone has a weird type of rot from a different climate or a house built in a totally different decade? My buddy tried something similar for spotting rust on old cars and it totally fell apart on anything that wasn't a Ford from the 70s. Maybe it's just me but I think these models are way more fragile than people admit, especially when you're dealing with real world stuff like houses that have all kinds of dirt and shadows.