I managed to get this to work by reducing the learning rate:
x = rand(75,11);
y = randi(5, 5, 75);
P = x';
T = y;
% Set the learning rate to 0.001
net = linearlayer(0,0.001);
net = configure(net,P,T);
net.IW{1}(5,11);
net.b{1};
[net,a,e,pf] = adapt(net,P,T);
net = train(net,P,T);
By default, linearlayer does not use backpropagation when training. Instead, it uses the Widrow-Hoff learning rule, so the training might behave differently compared to other networks (fitnet for example uses backpropagation by default).
It's also worth noting that linearlayer is learning only a linear function (that is, a weight and a bias), so it won't fit data with a non-linear relationship very well.
