Question about the old way we used to train image models

I was cleaning out some old project folders from 2018 and found a script where we had to manually label thousands of pictures of street signs. It's wild to think we used to spend weeks on that, and now I keep seeing people act like you still need a massive labeled set for every new task. That's just not true with modern self-supervised learning, and it matters because it changes what a small team can build. How did you first notice this shift in your own work?

2 comments

2 Comments

paul2332mo ago

Remember labeling medical charts by hand for hours. The tedium was real. Finding a self-supervised method for my own side project last year felt like unlocking a cheat code. It completely changed the scale of what I thought was possible alone.

ruby5612mo ago

Right? "Manually label thousands of pictures" gave me flashbacks. I found a folder from 2019 full of cropped cat ears I had to box by hand. The new tools feel like actual magic.