sarthusarth / deepdream

Google's Deep Dream model in keras

Deep Dream

Deep Dream is an experimental project by google, for visualising the layers within a deep convolution neural network. Apart from visualising an interesting applications involves transforming images into imaginary stuff.

Demo

Understanding the Convolution layers :

Convolution is a mathematical operation performed over the images ( In Convolution Neural Networks ) and the convolution layers learns to extract features like edges , shapes within the image through back propagation . The layers looks something like the image below .

Deep Dream Approach :

What actually happens in deep dream is the we take pre - trained convolutions neural network ( VGG 16 in our case ) and choose some layers the we want our image to imitate. So our aim is to maximise the output these layers or minimise the negative of these layers . Thus we find the derivative of out loss ( neg of the output of layers taken ) with respect to our input images and apply gradient descent to it .

References

• Colah'sBlog
• Research paper