I have been doing fast AI for coders for awhile now and I was able to follow the code in my own notebook. But when I tried to apply structured data approach to real life application I never get any good result and I don’t know how to fix the problem. I am getting a bit discouraged because I felt like I am not making much progress except following the lecture and doing the workbook step by step. Anyone had similiar problems and was able to overcome it? Any tips or pointers would be greatly appreciated.
I’m in a similar position, I’m now comfortable recreating the example structured data workbooks and applying the strategy to different data but the results I’m getting are a little under whelming. Looking forward to any pointers you get.
Hey Patrick:
I found this link, https://www.youtube.com/watch?v=F1ka6a13S9I&feature=youtu.be , check out Andrew Ng’s talk beginning at about 21 minute, it talks about what to do if your model doesn’t produce the desirable result. Let me know what you think after watching it.
I also liked this recent paper on Airbnb’s transition from traditional ML to DL. The upshot was that it takes time to make the transition, and while CV classification such as dogs v cats is a fairly easy recipe, structural data takes a bit more thought about what you are trying to accomplish
https://screenshots.firefox.com/8zOD2qpYeQxwToIZ/arxiv.org
I will say that a real world fraud model using embeddings took several months to get to a viable state, but none of the traditional ML variants even came close.
3 Likes
Thanks for the article Ralph. It is good to know that I am not the only one who struggle with this. As Andrew Ng points out in his talk above, DL is an art with science. But he also shows ways to trouble shoot your model depending on the error ratio between validation and testing. Which was very helpful.