Hello :) Today is Day 28!
A quick summary of today:
- covered CNN by Andrew Ng
Today’s material was a bit hard, but still fun ^^
I learned about widely used CNN architectures (LeNet, AlexNet, VGG-16, ResNet, MobileNets)
1. Basics
depending on the filter hyperparam, we can learn different features
During training, the w1….w9 parameters can be learned, allowing us to know the different features of the image
If you look at this picture, the model is composed of a conv layer and a pool layer. Among the pool layers, there are average and max layers, and the calculation is different.
If I want to learn more about architectures, I will need to read some research papers.
2. Widely used CNN architectures
-
LeNet
7 layers: 3 conv, 2 pooling, 2 dense. Used for handwritten digit recognition
-
AlexNet
8 layers: 5 conv, max-pooling layer, 3 dense
-
VGG-16
conv + pool + dense layers
-
ResNet
different architectures: ResNet-18, ResNet-50, ResNet-101, and ResNet-152
ResNet’s ‘skip connections’ allowed us to train deeper learning neural network model.
3. Other
-
1x1 convlutions
I need to read the paper for this one - https://arxiv.org/abs/1409.4842
-
MobileNets
There is big cost difference.
I need to apply what I learned today in Kaggle some day.
That is all for today!
See you tomorrow :)