Sorry this is so late. Some combination of procrastination, real life, and holidays made this happen so late. Mainly procrastination. What was explained well? What are you still confused about? ... My wife and I lost 100 lbs combined. . . . . . . . . She lost 120 lbs.
I cannot believe that my ML teacher failed to properly explain what pooling was, I had never understood until I saw your video and I can't believe how simple the idea is. Thank you so much for this great series on cnns
This is a simply and effective explanation. Thanks a lot for your time and passion. I will be watching more videos like this. You are a natural talented professor!
I cannot believe what just happened. You just explain me such a difficult topic in very very simple language. I am very impressed with you. I have, not particularly about pooling, but I have watched many ML and deep learning videos on RUclips and you are just very smooth. I am very sorry to see that you are not posting now-a-days, please do keep posting take some time out and do some favour for us learning candidates. We are students who want to learn ML/AI concepts. Thank you for this wonderful explanation and do post more videos.
If pooling is just picking the highest values in a sub Matrix then is the output of the pooling layer still just as feature rich as it was before it or does it lose meaning(or features)?
Hi, Thank you for making such good videos. I am enjoying re-visiting DL concepts. I would like to ask the following question: (1) What would happen if I connect a FC layer before my output size becomes 1X1, lets say around 56X56? How does that changes my network interpretation? And (2) What good book would you suggest for an intermediate level learner, I am in need of something that allows me to design my networks more intuitively. Thank you and regards!
Sure thing! 1) You could reshape your output to be one dimensional and feed that into your FC layer. I've seen plenty of papers do something similar, but usually it's when the output is already very small (e.g. 4x4). You have to keep in mind that the output of a CNN is actually 3d (height, width, num filters), and normally you increase the number of filters as your network get's smaller, so if you end on an output of (56,56, 256), the number of weights you will have is much larger than (1, 1, 2048), for example. How does it change your interpretation? FC's don't care about local patterns and are 1D - they find all patterns they can, regardless of how they are related in the image. Now, instead of the FC looking at features representing the whole image, it's looking at features representing blocks of the image. For example, an FC might find that 5 particular blocks of your input are useful for classifying digits, even if those 5 blocks are not at all what you think should be used. 2) I am not a book learner, so I can't recommend you and books. I learn from practical experience (e.g. coding tutorials, fun projects, work), seeing what other people did, and looking up specific questions online (e.g. what is the point of a 1x1 convolutional filter?), then working through what I think that means to how the network should be designed and then asking more questions and repeat.
Sorry this is so late. Some combination of procrastination, real life, and holidays made this happen so late. Mainly procrastination.
What was explained well? What are you still confused about?
...
My wife and I lost 100 lbs combined.
.
.
.
.
.
.
.
.
She lost 120 lbs.
lol
I cannot believe that my ML teacher failed to properly explain what pooling was, I had never understood until I saw your video and I can't believe how simple the idea is. Thank you so much for this great series on cnns
Glad it helped!
i rarey comment on videos, but you are just the best! explained the whole lecture in a couple minutes only!
Oh my god I can say, never learned anything this simple and deep. Great video
This is a simply and effective explanation. Thanks a lot for your time and passion. I will be watching more videos like this. You are a natural talented professor!
I cannot believe what just happened. You just explain me such a difficult topic in very very simple language. I am very impressed with you. I have, not particularly about pooling, but I have watched many ML and deep learning videos on RUclips and you are just very smooth. I am very sorry to see that you are not posting now-a-days, please do keep posting take some time out and do some favour for us learning candidates. We are students who want to learn ML/AI concepts. Thank you for this wonderful explanation and do post more videos.
Good explanation. I think it might be good idea to also explain how to propagate backwards through a pooling layer.
man i wish i saw your videos earlier thank for the simple and clear explination
Very informative
But why does it work in the first place? We basically lose informations by using pooling.
Indeed he should have mentioned this problem, which is ok
you are awesome!!!!! thank you
Great explanation. Thank you
When we want to do 2x2 max pooling for 11x11 feature map, what would we need to do with the extra rows and columns at the right and bottom edges?
Another great video! Thank you!
My pleasure!
Great explanation.
Ah, so that's why Ford gets a goofy face after killing and spawning him again too many times in Bonelab with the Gibable Guys mod
Superb
If pooling is just picking the highest values in a sub Matrix then is the output of the pooling layer still just as feature rich as it was before it or does it lose meaning(or features)?
Why and how does pooling increase the receptive field?
Hi,
Thank you for making such good videos. I am enjoying re-visiting DL concepts. I would like to ask the following question:
(1) What would happen if I connect a FC layer before my output size becomes 1X1, lets say around 56X56? How does that changes my network interpretation?
And (2) What good book would you suggest for an intermediate level learner, I am in need of something that allows me to design my networks more intuitively.
Thank you and regards!
Sure thing!
1) You could reshape your output to be one dimensional and feed that into your FC layer. I've seen plenty of papers do something similar, but usually it's when the output is already very small (e.g. 4x4). You have to keep in mind that the output of a CNN is actually 3d (height, width, num filters), and normally you increase the number of filters as your network get's smaller, so if you end on an output of (56,56, 256), the number of weights you will have is much larger than (1, 1, 2048), for example.
How does it change your interpretation? FC's don't care about local patterns and are 1D - they find all patterns they can, regardless of how they are related in the image. Now, instead of the FC looking at features representing the whole image, it's looking at features representing blocks of the image. For example, an FC might find that 5 particular blocks of your input are useful for classifying digits, even if those 5 blocks are not at all what you think should be used.
2) I am not a book learner, so I can't recommend you and books. I learn from practical experience (e.g. coding tutorials, fun projects, work), seeing what other people did, and looking up specific questions online (e.g. what is the point of a 1x1 convolutional filter?), then working through what I think that means to how the network should be designed and then asking more questions and repeat.
@@IntuitiveML Thank you great help!
when you write 100x100 or 3x3 what is the unit of this numbers?
magnifique