Loading…
15
A guy at the Denver AI meetup said my training set was too clean, and he was right by a lot
Six weeks ago I showed my image classifier at a small meetup in Denver and a guy named Raj looked at my samples and said I had maybe 40 usable edge cases out of 2,000. I figured he was being picky, but I went home and added 300 messy photos, blurry ones, weird lighting, half cropped stuff. My accuracy on real world test shots went from 71 percent to 89 percent, which I did not expect at all. Now I want to know how the rest of you decide what counts as enough ugly data before you call a model done.
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.