Big O computation of CNN architecture in deep learning

Hi,

I need to compute the complexity of AlexNet architecture theoretically using Big O notation. How should I compute that? Thanks in advance

Answer 1

[Shivani Dixit]

Hello,

Generally, we do not disclose the inner workings of built-in MATLAB functions, which includes providing specific values for the Big-O complexity that can be expected.

To find the information about typical algorithms, you can try to empirically determine an estimate. The following example gives a reference where we determine the Big O complexity of the built in MATLAB function “eig()” :

T = []; N = 100:10:1000; 

for n = N, disp(n) 

   A = rand(n); 

   tic; eig(A); t = toc; T = [T t]; 

end 

figure; plot(log10(N),log10(T)); grid on; 

As the figure window gets generated,

1. Go to "Tools > Basic Fitting" and choose a linear fit.
2. This will provide the equation with slope ‘a’ of the line, which is the exponent of n in the Big-O notation: t = O(n^a).

This code gives an example how the execution time of a built in MATLAB function “eig()” depends upon the input.

You can try to extend same sorts of experiments to CNN architecture to get a rough computation of Big O for the same.