Workflow Training for designing Neural Network Node-Layouts

Hello NN Professionals,

I am wandering whether there is a training program which helps to get strong in designing Neutral Networks the right way. This is only about the Node Structure.

What I found is:

• Designing Neural Networks
• 14 DESIGN PATTERNS TO IMPROVE YOUR CNN
• Periodic Table of DL Patterns
• The Neural Network Zoo

But this is only the description of the workflow and patterns. What I am looking for is a training which strongly focusses on design patterns of node structure but also the simple basics with about 10 nodes only.

For example: You want to train a CNN to perform visual planning on a Maze of size 28x28 an 4 action (n,e,w,s) agent with view 5x5 and walls of size 2. How does the most efficient structure of nodes looks like if a block is 2x2 pixel?

• 8 Nodes in linear layer followed by 4 Node as output
• 16 Nodes as Input followed by 8 Nodes and finally 4 Nodes as output
• ....

Another Example: You want to train your Network on MNIST Dataset. What network structure fits best if you like to detect numbers 0-9 consisting of 28x28 pixels?

• Convolution Layers of size 28x28 and 10 output layers
• Input Kernel of 5x5 followed by 10 filtered CNN of 24x24 followed by Max Pooling 2x2 and 10 12x12 Pooling Nodes...
• ...

@editors: please make this example more reasonable and erase this line :)

This is not tested - but simple should show what I mean.

Currently I am mostly unsure why I am doing things - I simple have an idea which makes sense to me at that moment. I like to become sure in knowing what I'm doing - this needs training.

Hope someone knows such a training like Multiple Choice + explanation ... or someone is able to start such a quiz (maybe @ KaHoot we used this in a Reinforcement Learning Lecture - it helped a lot)

Another Idea would be to collect all interesting free training capabilities for students and people who have to learn and be trained about designing NNs.