Datasets:
SPRINGLab
/
asr-task-data

Dataset card Data Studio Files
xet
 Community
1
audio
audioduration (s)
0.62
28.4
text
stringlengths
2
287
with neural networks is of the way that ah we would be starting down understanding as
comes into it but before starting down any of these aspects over there the first introduction
which we need to have very clear in our minds is about what do we define as something called
as learning now if you go down by the very classical definition
on machine learning is ah what outlines it out and the outline is something like this
that a computer program is said to learn from certain experience e with respect to a certain
class of task t and a performance measure p so if you see there are three attributes
factor which is called as an experience e there is a particular task which it has to
perform t and there is a performance measure p
classification right so we are not doing any other task so this this was like if i want
to just find out whether there is a ball in the image or there isnt a ball then thats
as my experience was increasing which is my number of epochs over which i was translating
that means that i am somehow able to measure my performance and see that the performance
is increasing so as the performance increases it becomes more and more accurate and accordingly
to learn if this performance on a task t which is of my classification as measured by this
learning is all centered around in fact human learning is also quite similar there also
as as human beings when we say that we are learning about something then the whole task
of learning is when we are able to really getting more
standard definition now once we have been able to define that one lets look into trying
to demystify what this would mean now lets get down with a very basic problem
have you been what have you seen what have you learned whatever you experience and then
humankind gains over there so if this is the image which is given down
even there i mean this this blind person to somehow know and contemplate on your own experience
of this one or a very simple thing which is called as an image captioning problem as of
today so what will happen is something like this
that as you would see that initially a computer program it if it is a so it will be doing
over there what you would see is that as its able to understand and recognize each block
increasing along that one now going down through that one what happens is that in the next
instant that it will be able to identify some more objects over there and they are those
and there is a great wall tower and finally it can identify these different
equivalent of whatever it has identified over here now the interesting aspect which happens
to that experience now as it took all of this a good amount of time coming down over there
to the sentence so there is a lot of error so finally when you go closely on the sentence
now now that we know that this is what essentially we meant down when we were saying down that
its learning something the next objective is to understand what was it learning and
how was it learning more than what it is actually how does it actually go on to learn this one
so lets again get back over there so as you see in the whole image you would be getting
down the image first and then the first objective over there is to break it down into some number
it comes and lets say that this is breaking down an image into its salient segments now
to identify some of these segments or what is also called as an objectification task
then the machine is able to find out that there are certain number of inanimate objects
that there are humans over there it will try to recognize humans find out who is who actually
and this is essentially what what this machine is able to do but the question is even bigger
the question is that we know that how it was learning was by doing something of this sort
and the deeper it keeps on going so that is over the hierarchy as it keeps on going which
keeps on climbing climbing climbing up to the description of the scene so as it keeps
it and now the aspect of deep learning says that
as its able to go down so its obviously gaining this depth by gaining looking at more number
of images getting down more and more experience and accordingly its its performance is increasing
the major question which we have as of now so ah i would give you a few seconds to actually
ponder on this one whether its unique or not do you think there can be a non unique way
would make a replica of this itself now let me just remove certain of these connections
these blocks is still the same whereas the order in which the blocks were connected somehow
here as you look into over here what happens is that you can still put down image it will
goes on to recognize and some of you can even say that we can pull
the inanimate things over there and thats thats perfectly fine i mean thats also another
possibility of doing it so as you see what happens is essentially it turns out that there
so as researchers for us its a very interesting point because we know that there can be multiple
a product development perspective its really really a very dicey situation because if you
have non unique ways of solving a problem that means you will have to explode down each
and every possibility of solving out that problem and find out which is the best possible
solution in order to achieve a solution to this problem
only way of doing and and and this this problem this this challenge which we have over here
way of solving this problem and yet more another way of in fact there are certain interesting
papers which do come out in conferences which called as yet another way of solving this
but then the point is is that the only issue which comes out or or can there be some other
ways of doing it as well so as it turns out this is not the only challenge
on this small block itself ok so that should be enough to say whether there is so this
of a body part recognition or run down one classifier which can identify which which
and this is a body part recognition now once i have my body part recognition what you can
do is between these body parts i can draw down lines and find out what are the distance
these distance relationships the angles in which they vary and then using that these
are very pretty different over there then the posture because i they they dont always
different the posture is different the distance between the legs and the hands are different
the the angles at which these things are connected they are also pretty different and thats what
which is like if human beings are present over there in black and whites this is what
ways and two non unique ways of detecting humans and as it turns out that you can have
speech and and apparently it turns out that there is no unique way of doing it
processing which is from your sentences can you make inferences out of it or say today
it knows that it has to put down get todays date and generate a query to our website on
now from there there are interesting problems on hierarchical and transfer learning as well
and so we would eventually go down a bit later on into what this transfer learning and hierarchical
learning is all about and it it does exist in the field of medical imaging and image
scope for researchers was for a longer duration of time but today if you see with the advent
what we come down to is lets come down to the most consistent solution available by
dilemma and for that whats done is something of that sort so say you have this image captioning
problem over there so what i can do is i can take an image i can organize all the pixels
subsequent nodes over there and now finally what it would do is that there is it would
generate some sort of an output which would say that there is a great wall behind and
there and if you look into this one what what this
the pixels and from pixels it will translate to some alternate representations by clubbing
all the pixels together into one representation than another and then subsequently as it goes
certain labels over there now carefully getting back this model actually
on on simple neural networks so we will get down into exactly what how the mathematics
layers and each of this is what is called as a hidden layer the reason its hidden is
of these layers like its its no target output which comes out the target output only comes
so these output layer and the input layer to which you give an input and you draw an
output from is what is called as the visible layers and inside all of these intermediate
as the hidden layers over there now as you get a multi layer perceptron what comes down
is that you will also have to train a multi layer perceptron