Lecture 9 | (1/3) Convolutional Neural Networks
HTML-код
- Опубликовано: 2 авг 2024
- Carnegie Mellon University
Course: 11-785, Intro to Deep Learning
Offering: Fall 2019
For more information, please visit: deeplearning.cs.cmu.edu/
Contents:
• Convolutional Neural Networks (CNNs)
• Weights as templates
• Translation invariance
• Training with shared parameters
• Arriving at the convolutional model Кино
I really like the way the professor motivates CNN, which helps me understand the "why" CNNs does things in that way.
These lectures are amazing! I feel so lucky to have found them.
Amazing lecture! I hope Professor Raj would produce a book on deep learning based on his slides! By far the best Deep learning courses I have found!
Very good.
This is how a professor should introduce a topic. Unfortunately such professors are very rare. I wish the professor doesn't change his approach to teach for lack of attendance or comments otherwise. One who learns under him must be able to innovate more than one who learns cookbooks.
Sir this is a wonderful lecture but I have a couple of questions:
1. In 01:01:53 , Slide 226, I can not understand how you are introducing the factor of (K/L1.L2.L3...Ln) in the formula of calculation of no. of parameters for the distributed CNN. I understand that L1 is the filter size in the 1st Layer, L2 in 2nd layer and so on but why (K/L1.L2...) in the penultimate layer before moving to the FC layers?
2. Also, you evaluated the no. of parameters to be less in the distributed CNN which I understood. However, as per my understanding from the rest of the lecture, the L x L block is moving across the entire image to generate smaller images at subsequent layers which you have shown earlier. That will increase number of multiplications and additions by a great deal. Isn't this a disadvantage over MLPs?
Kindly send a reply as I am a little confused.
18:22 "A basket with photographs where "SOME" of the images have cars"; Isn't this mislabeling instead of weaklabelling? I ask this because, there are images in the first basket which do 'not' have cars (because the word "SOME" was mentioned), but is labelled as having a car (because it is a part of the first basket), thus giving wrong signals to backprop? Could someone please clarify?
To my knowledge, this is so far the most complicated way of understanding CNN. There is a big jump from the initial scanning story to the convolution story...
You may wish to skip to part three if you are not interested in how CNNs relate to MLPs or how they were originally conceived.
To me, this is a very organic introduction from MLPs. Most importantly, it seems to set up "why" things are done the way they are, instead of handing the student tools that work. After all, knowing "why" things are helps one frame future problems that have not yet been solved yet.