Understanding convolution

Convolution is the central concept behind the CNN architecture. In simple terms, convolution is a mathematical operation that combines information from two sources to produce a new set of information. Specifically, it applies a special matrix known as the kernel to the input tensor to produce a set of matrices known as the feature maps. The kernel can be applied to the input tensor using any of the popular algorithms.

The most commonly used algorithm to produce the convolved matrix is as follows:

N_STRIDES = [1,1]
1. Overlap the kernel with the top-left cells of the image matrix.
2. Repeat while the kernel overlaps the image matrix:
    2.1 c_col = 0
    2.2 Repeat while the kernel overlaps the image matrix:
        2.1.1 set c_row = 0
        2.1.2 convolved_scalar = scalar_prod(kernel, overlapped cells)
        2.1.3 convolved_matrix(c_row,c_col) = convolved_scalar
        2.1.4 Slide the kernel down by N_STRIDES[0] rows.
        2.1.5 c_row = c_row  + 1
    2.3...