your understanding is not wrong, for the basic version of convolution typically used in neural networks. however, in practice the most optimized neural net libr

your understanding is not wrong, for the basic version of convolution typically used in neural networks. however, in practice the most optimized neural net libraries don't actually implement the operation in that way. as mentioned at the end of the post, if you understand things in terms of functions rather than just multiplying with a sliding window, you can then make use of Fourier transforms, and this enables much faster algorithms to compute the same result.