Transcriptions
Note: this content has been automatically generated.
00:00:00
we have a question in the room yes
00:00:12
i don't thank you very much for the annex presentation at
00:00:15
i've had a very interesting these physics and fun approaches but have you actually come right to the
00:00:21
are we do not watch because if this gets them people who actually often find that
00:00:25
text based approach just it takes so much longer to train unpacked is
00:00:28
dead and in the end again you get might not necessarily justify at
00:00:33
and the second question would be how much is actually the gain if included data preprocessing for the training data set
00:00:39
um hum rather than just uh that uh the
00:00:43
modelling approach itself with the neural network okay so mm
00:00:48
of course as you were saying d. e. training of those network is kind of
00:00:53
a longer than we've got approaches um but you have two advantages the first thing is that
00:01:00
you don't have to make a simulations and when you make simulation of course you have to plan
00:01:07
uh what to simulate an uh an have
00:01:11
samples there are enough a space enough let's say
00:01:15
one from each other so that your network and training them um a reasonable big problems bins
00:01:23
and the second advantage is that by imposing
00:01:27
some physical constraints you are let's say a
00:01:31
little more sure that it's actually doing what we supposed to do and it's not just throwing
00:01:38
um productions of something does remember the for example or
00:01:43
and that was biased by your training data let's see
00:01:48
but you compare that when i actually got better results yes actually
00:01:53
um i've compared to to and i'd got slightly better results we've d.
00:01:58
a liability approach the cask approach but then again it was performing very well
00:02:05
on a small range of input values but as i was to extend
00:02:09
the range of what is was not performed as well to be honest so
00:02:14
of course the training tiny was much real were so you can think of a more
00:02:19
to the specialist for our um a small range of values but then that's not really interesting
00:02:29
um thank you unfortunately we are running out of time should that we have a cushion