I can't understand how these visualizations can be real if images are progressively being convoluted and have less and less resolution throughout the convnet
You are thinking the deep CN learning as a "forwarding process", but it is actually a "backwarding process". Instead of thinking the first layer params are generated from a convolution of the image, better think the first layer params are told to be so by the second layer which is again told to be so by higher layers all the way to the loss function of input data. It is this structure that decided the first layer better only "see only a litter fraction of the image" will perform the best for the purpose.
It convolves through the entire image but with it's filters i.e. the weights that are only looking for pattern in the whole image like a diagonal or a straight line we can only see the edges(Activation only values it as edge or not edge).Next is pooling, Imagine it as something that reduces the resolution but shows the same image(The image hear being the edges but just in a smaller resolution).We convolve it again and start looking for shapes like circle through the whole image features which we got from the previous result(i.e. the images) and it goes .LMK if it makes sense
![เขามัทรี - เอ็กซ์ ศุภกฤต - จอนนี่มิวสิค [ Official MV ]](http://i.ytimg.com/vi/AS9ufKuQIVs/mqdefault.jpg)
my logic of understanding- the further you are from images- the more (wider) you see!! :D
great explanation
I can't understand how these visualizations can be real if images are progressively being convoluted and have less and less resolution throughout the convnet
You go backwards to find the patch in original image that is giving maximum response at a certain layer.
why does the first layer see only a little of the image? it convolves through the entire image right?
You are thinking the deep CN learning as a "forwarding process", but it is actually a "backwarding process". Instead of thinking the first layer params are generated from a convolution of the image, better think the first layer params are told to be so by the second layer which is again told to be so by higher layers all the way to the loss function of input data. It is this structure that decided the first layer better only "see only a litter fraction of the image" will perform the best for the purpose.
@@lizeng489 So, in a way, the CNN in breaking the image down moving from higher level to lower level layers?
It is due to something called the receptive field. Look at this documentation from cs231n.
cs231n.github.io/convolutional-networks/
It convolves through the entire image but with it's filters i.e. the weights that are only looking for pattern in the whole image like a diagonal or a straight line we can only see the edges(Activation only values it as edge or not edge).Next is pooling, Imagine it as something that reduces the resolution but shows the same image(The image hear being the edges but just in a smaller resolution).We convolve it again and start looking for shapes like circle through the whole image features which we got from the previous result(i.e. the images) and it goes .LMK if it makes sense
@@lizeng489 Not sure why so many upvotes, this answer is just wrong and irrelevant to the question.
"Pick a unit in layer 1" means 110x110x1 from 110 x 110 x 96 ???
A unit is just one neuron
3:00
Who on earth gives picture of his wife in examples..?😂
How is this a problem when our country is going bananas.
Let's get to this when we solve like millions of actual problems in our lives.
Andrew Ng.
I understand every English words he said in this video, I just don't understand the content.
Do you understand it now?