Why is my deeper CNN not learning

Hello,

I am trying to build a CNN to identify certain parameters in images taken using holography. I first started using the built-in AlexNet and got great results. However, I think AlexNet is a bit overkill for the task and I wanted a simpler network for faster classification on slower computers. When I build a very simple network as such

layers = [ ...
	imageInputLayer([150 150 1])

	convolution2dLayer(11, 24, 'Padding', 'same')
	reluLayer

	maxPooling2dLayer(2, 'Stride', 2)

	convolution2dLayer(5, 48, 'Padding', 'same')
	reluLayer

	maxPooling2dLayer(2, 'Stride', 2)

	convolution2dLayer(3, 48, 'Padding', 'same')
	reluLayer

	maxPooling2dLayer(2, 'Stride', 2)

	fullyConnectedLayer(100);
	fullyConnectedLayer(19)
	softmaxLayer
	classificationLayer];


% Training options (default)
% Gradient descent with momentum
opts = trainingOptions('sgdm', ...
      'Plots', 'training-progress', ...
      'Verbose', true, ...
	'MaxEpochs', 100, ...
	'MiniBatchSize', 32, ...
	'Momentum', 0.9, ...
	'Shuffle', 'every-epoch', ...
	'InitialLearnRate', 0.001, ...
	'LearnRateSchedule', 'piecewise', ...
	'LearnRateDropFactor', 0.25, ... 
	'LearnRateDropPeriod', 10, ...
	'CheckpointPath', checkpointPath);

The network learns fine and reaches an accuracy of around 85-90 percent fairly quickly. However, as soon as I introduce any extra layers between the max pooling layers it takes a really long time for the network to learn. See below

layers = [ ...
	imageInputLayer([150 150 1])

	convolution2dLayer(11, 24, 'Padding', 'same')
	reluLayer

	maxPooling2dLayer(2, 'Stride', 2)

	convolution2dLayer(5, 48, 'Padding', 'same')
	reluLayer

	maxPooling2dLayer(2, 'Stride', 2)

	convolution2dLayer(3, 48, 'Padding', 'same')
	reluLayer
      convolution2dLayer(3, 48, 'Padding', 'same')
	reluLayer

	maxPooling2dLayer(2, 'Stride', 2)

	fullyConnectedLayer(100);
	fullyConnectedLayer(19)
	softmaxLayer
	classificationLayer];

What could be the reason for this? It is worth to note that there is not a lot of variation between the images, so changes are subtle between classes. Any tips for building CNNs that can handle subtle changes well?

Answer 1

[Kenta]

Constructiong a good CNN architechture from scratch is not easy. If a heavy network such as AlexNet is not necesarily required, how about using SqueezeNet? It is very light and can get good accuracy.