Step 3:  

clear;
clc;
close all;


mynetwork=feedforwardnet([7,7]);  % Create the network
mynetwork.layers{1:2}.transferFcn= 'logsig';  % Change the activation functions for the hidden layers (from default tansig to logistic logsig)
mynetwork.layers{3}.transferFcn= 'tansig'; % Change the activation function for the output layer (from pure linear to tansig)

Change the learning rule to stochastic gradient descent (or to another one of the ones shown in the list during lecture).  Some of these rules are faster than others, and some are more accurate.  Some are designed for smaller networks and some for larger complex networks.  The stochastic gradient descent training rule is a good compromise among the many training rules and can be used in most cases.

mynetwork.trainFcn = 'traingd';