Course4 - WeiliangGuo/deepleaning_studies GitHub Wiki

CNNs are designed to discover position-invariant features of their inputs whereas RNNs are used for sequential data (with positional information).

image, filter. Padding p, stride s

image shape after performing strided convolutions and zero-padding:

If layer l is a conv layer:

Input: , h stands for height and w for width.

Output:

f^[l] = filter size, p^[l] = padding, s^[l] = stride, = # of filters, here c stands for # of channels.

For each filter:

Activations applied on results of convolutions(m training examples):

Bias: one real number for each filter:

In addition to using convolutional operations in convnets, we also often use pooling layers which are used to reduce the model complexity to speed up computation as well as making feature detection more robust. Common hyper-parameter values for filter size f and stride size are 2 and 2 or 3 and 2. Notice: There are no (NOT hyper-parameters)parameters in pooling layers hence the back-propagation process has no effect in these layers.