How is creativity generated in a currently rule based neural network?

Question

I recently became interested in how creativity is generated in NN. My understanding of NNs is that the output always known, given that it they are trained with target values, but how does one train a network to be creative, I mean in such cases would the task be to create something novel, but the actual target would not be known. How does creativity express itself in a rule based system, such as NN?

One would need to redefine the cost-function, but how does one imply creativity in a rule based system?

Answer 1

Your question is a bit broad. Some examples of creativity and how it's achieved (in all of them there is an ordinary convolutional network trained via supervised learning, but applied differently):

• Google Deep Dream uses backpropagation to the image to tweak the image towards maximum activation in the convolutional layer. As a result, it finds lots of "things" in the image and enhances them:

  

• Generative Adversarial Networks use two competing networks (generator and discriminator) to generate samples from the training manifold:

  

• Neuro style method captures the style of one image via Gram matrix of convolutional layer activations and trains (another) neural network with a special loss function that minimizes the joint distance to the incoming image (content) and to the Gram matrix (style). The result: