hello welcome to the video this has taken forever to make don't think oh I'm
just gonna make a small little animation and then it'll be cute god I've spit so
outside like this I really hope you guys enjoy this video and yeah let's just
jump right into it there's a lot of things I don't understand chief among
them probably being colors I'm colorblind and it makes it really tough
to do stuff like color correct videos I still feel like I get around it pretty
much but you are still going to have to cut me some slack on this one I can't
actually see these colors so I've had to sort of rely on what other people have
told me about them and sometimes people agree on colors sometimes people don't
I started thinking after the video I made about sounds and how you could make
new instruments based on nothing but an AI and then the existing instrument that
you can put into it as an input we could do exactly what we did with the images
of the cats we can move between sounds we know and listen to what the model
does in spaces that we don't know now let's listen to bass and flute from
ensign what it does in my mind's eye is makes a
really big bass flute and I thought why can't you do the same with colors and
obviously and a lot of people have been thinking the same where they've not
really been getting great results at least in my opinion and the people I've
showed this to it seems like they pretty much just used existing colors and then
given those colors names but I think see erratically it should be possible to
actually get new colors we've never seen before the question however is will we
be able to see those colors because they might be existing colors already in
nature that we just simply can't see all of this is very theoretical and we can't
really prove much of it and I've tried to look around for papers and that's not
a lot even fixing colorblindness seems to be put down too if your red and green
all about it we can fix it with this pair of glasses if you're anything apart
from that you're screwed but I still found it super interesting if you could
create new colors using one simple AI just pop some colors in there let it
train see what it comes up with it would have to be a Gann Network a generative
adversarial Network where you basically have to a is competing against each
other both being trained on a training set of data and the data you would put
into this would have to be existing colors so we would always be able to see
the output from the Gann Network the interesting bit here is you basically
have two different AIS that are competing trying to outsmart each other
one is generating new images or colors in this case based on noise or some kind
of input it could really be anything like a string of numbers or or anything
and then you would have the other AI that would be trying every time that it
got a picture to determine if the picture was from the training set or if
it's a new picture and if so if the new picture is good enough to be part of the
training set and if that's the case then it outputs the picture in the other end
and that's when we get the actual picture or once again the color in this
case we could use this to create new colors just like they've done with new
musical instruments and everything else really the one thing about this is using
other types of networks such as convolutional networks
don't get great results so what we end up having is these networks that
basically just put out an existing color with AB funky name attached to it
because all the colors in the training set has a name on it and it's kind of
fun to see but it's also a little bit sad that the whole world goes crazy for
this I had to make a video on it because it is true the colors are generated from
nai and the AI has named all of the colors
but for instance in this article we talk about all of these colors and then they
say they prefer beige and brown and it's not really the case in a lot of the
articles I found out this this was misinterpreted as the AI having a
preference for certain colors beige or brown in this case and that's simply not
the case a convolutional network will start to figure out what type of images
get through and then create more images like those so in this case if beige and
brown are the first colors then the initial neural network will start
creating more colors like these this is why it's convolutional networks have a
lot of issues with faces and animals because once one animal or one face has
been approved it will start creating more images like these so it will
basically send hundred pictures of the same person let me know what you guys
think in the comments down below if you have a better way to do this I would
love to hear about it we could use something like an LS TM which is like a
long short-term memory system the memory in this case I don't see a lot of reason
to have maybe we should use an RNN if this is all just mumbling to do then let
me know what else would you use this on if you had the choice to do it on
anything because honestly the limitations are nowhere to be found here
we could try this on pretty much anything and it might work but yeah
that's gonna be it for this video guys and I'd love to hear what you think
about this animated beer as well I call him Jason I been playing around with him
for a while I don't know it takes how long to bake but I really I find them
quite cute and I don't know let me know what you guys think if you want to see
some old stuff from me there's a video up here if you want to see my last video
there's some stuff up here and yeah thank you so very much for watching I'll
see you guys next week
Không có nhận xét nào:
Đăng nhận xét