Trustworthy ML and Generating Music Traditions
I was attending to another meetup on 05/12/2016. The meetup’s title is ‘Trustworthy ML and Generating Music Traditions’.
The first part was about the trustworthy ML by Peter Flach and it was really useful. I almost understood the whole presentation :). The takeaways were for me the followings:
Do not use the Precision Recall curve unless you exactly know that this really helps in your case. The main reason is that the Precision Recall curve can easily fool you and you will pick the wrong model to work further with it. I really don’t know the background of this statement, but the professor said that the diagram does not represent the data well, because it is ‘logarithmic?’.
If you don’t know deeply the mathematics of the algorithms that you are using then you can get false result very easily and the worst part of it that maybe you will not know that is false result. (I was a little bit afraid when I heard that because I just realised how far I’m to become a Data Scientist, but the next thought was that okey I’m also excited because there’s a lots of thing to learn :) ) And this statement leads to the 3. point.
In the ML everybody uses algorithms that are maybe not in the final format (so we can improve it or invent new one) and those algorithms are just equations so the Data Scientist should know how to use them to get the right result. We can get a result very easily but if we don’t consider those result and reconcile them with the expected result or even we just don’t try to understand the result and verify it then the whole ML is just a childish game without any confidence.
The second part was about generating music by ML algorithms by Bob Sturm. It was a funny presentation and also showed that the AIs can be fooled very easily and we can build an AI that is able to detect the right genre of the music but if the AI is not robust enough than with a little noise the AI gives a wrong answer. Here is the presentation https://highnoongmt.wordpress.com/2016/12/06/my-london-machine-learning-meetup-2016-slides/