Adding hidden layers to a patternnet hurts accuracy?

I am trying to use patternnet to classify the MNIST handwritten digit dataset.

I expected patternnet(10) to do worse than patternnet([10,10]), but it seems that the accuracy decreases as I add more layers.

Can someone explain why?

Here is my code:

images = loadMNISTImages('train-images.idx3-ubyte'); % initialize figure  
labels = loadMNISTLabels('train-labels.idx1-ubyte'); % initialize figure
labels = labels';                                    % transpose
labels(labels==0)=10;                                % dummyvar function doesn´t take zeroes
labels=dummyvar(labels)'; 

net = patternnet([10,10]); %or patternnet(10)

net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;
net.performFcn = 'crossentropy';

net = configure(net,images,labels);

net = train(net,images,labels);

y=net(images);

perf = perform(net,labels,y)


correctcount=0;
for i = 1:60000
[M, I]= max(y(:,i));
if t(I,i)== 1
correctcount=correctcount+1;
end
end
errorrate = 1- (correctcount/60000)

Answer 1

[Greg Heath]

1. The global minimum is achievable with a single hidden layer.
2. With more hidden layers you add more local minima; most of which are higher than the global minimum.

Thank you for formally accepting my answer

Greg