Due on Feb 17

In this “quiz”, I would like you to try to play with Netscope, a visualization tool for neural network architectures.

1. Try to study the preset GoogLeNet model and answer the following questions:

   1. How many pooling layers are there?

   2. What is the size of the output after the first conv layer? And how many parameters in that layer?

2. Now, please take a look of the Caffe Zoo wiki (if you train a good model, you should put it up there also!). Follow the link of ResNet, probably the most sophisticated recognition model at the moment, please download deploy.prototxt (the network structure file) of the 50 layer model. And visualize it by copying and pasting to the Netscope editor. And please answer the following questions:

   1. What is the size of the output after first “block” layer (before first pooling)?

   2. Describe what a scaling layer (e.g., scale_conv1) does. You may want to check out the caffe layer description.