Implementing DeepDream

Another use for trained CNNs is exploiting the fact that some intermediate nodes detect features of labels (for instance, a cat's ear, or a bird's feather). Using this fact, we can find ways to transform any image to reflect those node features for any node we choose. This recipe is an adapted version of the official TensorFlow DeepDream tutorial (refer to the first bullet point in the next See also section). Feel free to visit the Google AI blog post written by DeepDream's creator, named Alexander Mordvintsev (second bullet point in the next See also section). The hope is that we can prepare you to use the DeepDream algorithm to explore CNNs, and features created in them.

Getting ready

Originally, this technique was invented to better understand how a CNN sees. The goal of DeepDream is to over-interpret the patterns that the model detects and generate inspiring visual content with surreal patterns. This algorithm is...