Loading…
19
Vent: a guy at the Austin AI meetup asked me what my model does when it's unsure, and I couldn't answer
So I went to this small AI meetup in Austin 2 weeks ago, maybe 30 people in a back room of a coworking space on Congress, and during the break this older guy with a laptop covered in stickers starts chatting with me about my side project. I've been messing with a fine tuned model that sorts customer support tickets, nothing huge, maybe 4,000 training examples. He asks how it handles stuff it wasn't trained on, and I said it just kind of guesses. He goes quiet for a second and then says "so it never tells you it doesn't know?" and I literally had nothing. I pulled up my logs on my phone right there and found 12 cases in the past week where the model gave a confident answer that was flat out wrong, and my app just passed it through like it was fine. We talked for another 20 minutes about calibration and refusing to answer, and he showed me a paper on uncertainty estimates that I'd never even heard of. Now I'm rebuilding my whole output layer and it's way more work than I expected. Has anyone here actually shipped something with a real confidence threshold, and did it hurt your accuracy at all?
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.