Convolution layers

The values in a kernel are selected during training.

By the nature of matrix multiplication the map will get smaller since the border pixels are always going to be 0.

Attribute

Num of filters: detects number of aspects

Kernel size:

Stride: how many pixels moves per step, aka it might move 3 pixels to the right

Padding: what happens at the borders the 0s in the image above

Pooling layers

There is no need to keep extra data points.

Pooling reduces the size of the matrix.

Types:

• max pooling: takes the highest value in an area

• avg pooling: takes the avg in an area

Drop-out layers

The process of during training to completely ignore some layers.

• Makes it faster

• Helps avoid overfitting

Note all layers are connected in prod.

Dense layers

Flattening is when you turn multidimensional data into a single dimension.

The num of neurons at the end matches the number of classes, each layer should decrease the number of neurons to sooth out the transition.