This was the sentence that I was wondering about.
When designing a neural network architecture for execution at Akida, one should take into account a number of additional limitations concerning the layer parameters (e.g., maximum convolution size is 7, while stride 2 is supported for convolution size 3 only) and their sequence.
My layman's interpretation of CNN -
In convolution the system takes manageable size bites of the pixel data for processing.
The processor examines small segments of the whole pixel array to see if there is a pattern which it recognizes. It does this by looking at each pixel in association with their neighbouring pixels.
Think of an array of pixels in a camera photoreceptor, say 100 columns x 100 rows (10000 pixels) - starting at the top left corner, draw a box enclosing, say 7x7 pixels. That is what the convolution size means.
Now move the box 1 pixel to the right - that is a stride of 1.
This results in the 7 pixels in the first column being excluded from the next processing step and a new set of 7 pixels to the right being included.
So a stride of 2 moves the box 2 columns to the right, dropping 2 columns of 7 on the left and adding 2 columns of 7 to the right.
When the box has scanned the whole first 7 rows, it can drop down 2 rows and repeat the process.
