3 ms·
Agreed The "most popular" (and maybe click-bait like) the content, the more crap you'll get. Especially when people think they know the subject but they don't.
by raverbashing 6y ago
Agreed
The "most popular" (and maybe click-bait like) the content, the more crap you'll get. Especially when people think they know the subject but they don't.
Most common offender nowadays: ML and Data Science. I've found articles with glaring errors, techniques that don't make sense, very shallow tutorials, etc (Disclaimer: nobody is immune from making mistakes, not even me)
And then you get "oh but you start training and then your loss go to zero, right?! This means it works, right?" Yeah. No.
If your training loss goes to zero, your are probably overfitting or the problem you're trying to solve is too simple. If your testing loss goes to zero, you're probably doing something wrong. Or my favourite: your loss is very low but what you're trying to predict has result A 99% of the time and result B 1% of the time. What's your loss if you only predict result A?
I mean, if even the myths about /dev/random and /dev/urandom keep circulating around, can you imagine about other stuff?
- GuiA 6y agoI took a grad level course on data mining/machine learning back in the early 2010s, before the current gold rush, and while there was some level of theory and programming, the professor spent a lot of his time discussing these finer points. The (apocryphal) example of armoring planes in the wrong spots [0] sticks with me to this day. It’s human nature to focus on the more immediate, tangible outcomes - eg going through a pytorch tutorial - over spending time trying to really understand the problem they are trying to solve. Great teachers try to guide their students there, and great students are perpetually calling themselves out mentally to try and stay on the track of understanding rather than rote repetition. But it’s very hard. [0]: http://www.ams.org/publicoutreach/feature-column/fc-2016-06 http://www.ams.org/publicoutreach/feature-column/fc-2016-06