,
4 tweets,
1 min read
Read on Twitter
One of the biggest failures I see in junior ML/CV engineers is a complete lack of interest in building data sets. While it is boring grunt work I think there is so much to be learned in putting together a dataset. It is like half the problem.
Our own vision is so full of perception hacks that you really don't see what you're actually looking at. Moreover, taking one or two samples and saying, "oh yeah, that's the problem" is fraught with sample bias.
If you are building an ontology you really don't understand the problem until you see a lot of the corner cases, which means combing through the data to find the stuff that happens 0.1% of the time.
At this point I consider my job to be 95% building tools to scrape, mine, move, annotate, review, and preprocess data sets, and about 5% doing the actual "machine learning".
