Convolution

Question 1

What is the mathematical operation called convolution?

Answer 1

Sliding window dot product between a kernel, a small matrix, and the pixel matrix.
Reminder: dot product is the summation of element-wise multiplication:
a * b = sum(a_i + b_i)

Question 2

What is the size of the of the output for convolution between a picture of N pixels and a kernel size of K

Answer 2

N - K + 1

Question 3

Feature Map / Activation Map

Answer 3

The output of the convolution between an input and a kernel

Question 4

Convolution Layer

Answer 4

a convolution layer computes the convolution between k kernels and its input with optionally a bias, yielding k outputs.
The learned parameters are the kernels and the biases.

Question 5

Pooling Layer

Answer 5

Reduces the dimensionality of its input by an unlearned, simple filter.

Max Pooling: take the maximum value from a given window (for example, a 2x2 window from a 4x4 matrix).

Question 6

Name the components of a CNN

Answer 6

1. Convolution Layers
2. Pooling Layers
3. Fully Connected Layer

The fully connected layer is designed to flatten the output to the desired dimensionality. For classification, it will undergo a softmax to get a probablisitic output.