Question

I am writing a machine learning code and I struggle to find the way of having some operations with matrix manipulations instead of iterative way with basic for loops. Do you think using matrix other than iterations make so much difference or it is ignorable performance difference?

Answer 1

Historically loops in matlab were very slow. However, in the versions of Matlab that have the new JIT compile loops can be quite fast.

In Matlab its advised to avoid loops whenever possible because the language as a whole is designed for vector based operations. When writing matlab code its considered bad style to loop over a vector instead of using vector based math.

Good matlab code:

[a b] = deal( rand(10,1) );
c = a+b;

Bad matlab code:

[a b] = deal( rand(10,1) );
c = zero(10,1);
for i = 1:10
  c(i) = a(i) + b(i);

Both of these implementation are “correct”, however 99% of matlab programmers will use the first implementation. Additionally any matlab programmer would see the first implementation and know exactly what the code means.

Regarding performance, its hard to say before hand if a vector operation will be faster than a loop, as it really depends on the implementation details. However, in my experience vector functions are rarely slower than loops. I’ve encountered algorithmic problems that I couldn’t solve without a loop. When I discovered the vector based solution it reduced the computation time down from several minutes to less than a second: How can I optimize this indexing algorithm