[Read Paper] Minimizing Computation in Convolutional Neural Networks
Minimizing Computation in Convolutional Neural Networks
Properties of CNN Computation
Computation Optimization
In this section we show how to extend the Strassen algorithm from Normal MM to reduce the computational workload of our Convolutional MM.
We can see that each recursion will reduce the number of multiplications by 1/8, but will incur 18 additions on the submatrices. The overhead of 18 additions could completely eliminate the benefits brought by the multiplication savings in normal MMs.