Google Deep Dream: What does my network actually see?

We will be having a guest speaker to kick off the new year! Machine learning and AI are becoming more of a viable option to use with software testing practices. This month we are very excited to learn about neural networks with Joe Dumoulin presenting on Google Deep Dream.

In this talk we will look at Neural Networks for object recognition in images (vgg16 and possibly inception). We will talk about how the network learns and how learning affects what the network is seeing, that is, What are the features that the network uses to "understand" that an object is present? This will further lead us to a discussion of gradient-based learning and how it affects the ability of the network to do what we train it to do. If we have time we will show how this leads to the possibility of "fooling" the algorithm into thinking it sees something that isn't present.