How are the number of featuremaps and units pooled determined in a Convolutional Neural Network?

Hi All,

I'm wondering how the number of feature maps in various layers and the number of units pooled in the pooling layers are determined. Is the entire process just trial and error? Or is there a way to determine what will be an optimal number of feature maps or the optimal number of units pooled in various layers.

For example, let's take Lecun's Hand Written Digit Recognition paper. Link is below.

In case you don't click the link, the net takes a 28x28 pixel input and transforms it into 4 24x24 feature maps, with each feature being a 5x5 pixel section from the original input. The first pooling layer is 4 12x12 planes, with each unit in a plane being the pooling of four units from the corresponding feature map. The third layer is 12 8x8 feature maps with each feature taking 5x5 units from the corresponding pooling layer. Finally, the 4th layer is 12 4x4 planes where each unit in a plane pools 4 units of the corresponding feature map. image representation.

So how did Lecun decide to use 4 feature maps in the first layer? Why did he decide to pool 4 units in his next layer, instead of say...5 or 3? In the second feature map layer, why did he choose 12 maps? Why not 8? It would have been easier to do 8. Why did he choose 5x5 units for this layer of feature maps? In the 4th layer, why did he choose a 4x4 plane? Why did he choose to pool 4 units again? Why did not choose to create 1 more layer of feature maps and 1 more layer of pooling?

Is the determination just random trial and error? If it isn't trial and error, how do these parameters get determined to be the optimal set? (Note that I am using parameter here differently than how it is normally used in CNN. Obviously the weights will be determined by back propagation / gradients).

http://yann.lecun.com/exdb/publis/pdf/lecun-90c.pdf