The number of mathematical concepts that Graphics involves is staggering, combined with physical concepts I think graphics has the most coverage of inter-disciplinary topics in computer science.
keenan
@dranzer Yeah, it can leave your head spinning. But being pushed to consider all these perspectives really opens you up to lots of creative possibilities---and gets you talking to all sorts of interesting people!
Arthas007
I was always wondering why even sometimes numerical integration is better than closed forms answers(like sometimes cos, sin, log, etc), because in real practice we still have to calculate them by using tayler series or sth, so we will get floating errors anyway
keenan
@Arthas007 Yes, this is a great point. At the end of the day, you have to evaluate everything in floating point, and there will inevitably be some kind of error. Finding the scheme that yields the lowest error in practice is challenging. Just like performance tuning, the right answer is: to know how well something works, you really have to "profile" it. I.e., you really have to run it on real data and see how well it works. (And can often be surprised at how reality does not match up with your expectations!)
The number of mathematical concepts that Graphics involves is staggering, combined with physical concepts I think graphics has the most coverage of inter-disciplinary topics in computer science.
@dranzer Yeah, it can leave your head spinning. But being pushed to consider all these perspectives really opens you up to lots of creative possibilities---and gets you talking to all sorts of interesting people!
I was always wondering why even sometimes numerical integration is better than closed forms answers(like sometimes cos, sin, log, etc), because in real practice we still have to calculate them by using tayler series or sth, so we will get floating errors anyway
@Arthas007 Yes, this is a great point. At the end of the day, you have to evaluate everything in floating point, and there will inevitably be some kind of error. Finding the scheme that yields the lowest error in practice is challenging. Just like performance tuning, the right answer is: to know how well something works, you really have to "profile" it. I.e., you really have to run it on real data and see how well it works. (And can often be surprised at how reality does not match up with your expectations!)