Description
- Given a one channel gray scale image of size 6 * 6 shown below, if a 5 2 * 2 filters are applied with stride size equal to 2 and no padding to the image, what will be the size of the result? How many trainable parameters will there be in this layer? (each filter has its own bias).
| 1 | 1 | 2 | 3 | 4 | 2 |
| 2 | 5 | 9 | 7 | 6 | 5 |
| 1 | 2 | 0 | 1 | 5 | 4 |
| 1 | 2 | 2 | 1 | 7 | 0 |
| 6 | 4 | 0 | 1 | 5 | 2 |
| 4 | 3 | 3 | 2 | 5 | 1 |
Image
- If there is padding of one pixel around the image above, what will be the result for the first question?
- Apply the following filters with stride equal to 1 to the image above. What kind of possible feature in the image do you think can be captured by these filters?
| -1 | -1 | -1 |
| 1 | 1 | 1 |
| -1 | -1 | -1 |
| -1 | 1 | -1 |
| -1 | 1 | -1 |
| -1 | 1 | -1 |
| 1 | -1 | -1 |
| -1 | 1 | -1 |
| -1 | -1 | 1 |
- What will be the result if applying a 3 * 3 max pooling filter to the image?
- Below is a diagram of a small convolutional neural network that converts a 13×13 image into 4 output values.
The network has the following layers/operations from input to output: convolution with 3 filters, max pooling, ReLu, and finally a fully-connected layer. There are no bias/offset parameters
- How many weights in the convolutional layer do we need to learn?
- How many ReLu operations are performed on the forward pass?
- How many weights do we need to learn for the entire network?
- Would a fully-connected neural network with the same size layers as the above network (13×13 → 3x10x10 → 3x5x5 → 4×1) be able to represent any classifier that the above convolutional network can represent?
- What is the disadvantage of a fully-connected neural network compared to a convolutional neural network with the same size layers?
- Execute hw_9.ipynb. Apply the sample code to a data set of your choosing and document the results in this file and show the results in the Jupyter notebook.
