While Neural Networks (NNs) are widely used and produce astonishing results on different kinds of problems, there are those special cases where you have to deal with little training data, highly restricted power budgets or tiny memory. We will take you on a behind the scenes tour around well-known NN building blocks and show you how in depth analysis of the NN training process will assist successful design and training, especially in those non-standard cases. You will see how building blocks interact to create or suppress properties of NNs and how to supervise successful training. We will use our Open Source Software Barista to show you some live examples.
Required audience experience
The talk aims at an audience from intermediate level upwards. A basic understanding of the training and inference stage of standard layers (i.e. Perceptron, Convolution, Pooling) is expected.
Objective of the talk
When and why do we want to achieve properties like sparsity or maximum separation?
Which design choices lead to beneficial properties?
How to deal with little training data?
Techniques to supervise NN training and how this leads to better designswith regard to quality, size and/or speed.
You can view Soren’s slides below: