On 03/09/2011 10:57 AM, spir wrote:
On 03/09/2011 01:52 AM, Andrei Alexandrescu wrote:
On 3/8/11 4:11 PM, %u wrote:
Uh... how helping fix compiler bugs? Could we help with that? I
feel that's *much* more important than benchmarking, for instance,
since it doesn't make sense to benchmark something if it has bugs.
:\
The funny thing is that sometimes it makes perfect sense, as
benchmarks _do_ push the limits of, for instance, GC and may reveal
a latent bug ;)
Those are a very specific class of bugs -- bigger bugs like compiler
errors with handling templates are completely unrelated to
benchmarking, and they can be a deal breaker for many people.
I don't think anyone cares about *speed* as much as *correctness*...
would you rather have your 50% accurate program be twice as fast, or
have your 100% accurate program be half as fast?
In machine learning it's very common to trade off accuracy for speed.
Accuracy is not correctness. A result can be inaccurate and correct inside a
tolerance field, which is precisely one common path for machine learning. If
the program were incorrect, the machine would not learn (what one expects it to
learn).
Sorry, I was unclear. I meant inaccuracy and incorrectness can often two
different notions, depending on the topic. Just like simplicity and difficulty.
While people often mistake one for the other.
Denis
--
_________________
vita es estrany
spir.wikidot.com
