L12.2: DeepDream

In this video I continue looking at the topic of Generative Deep Learning, with an example of image generation. The Google DeepDream technique is an old technique (circa 2015), but it was one of the first examples of image generation using deep networks. Similar techniques are used in more modern versions of image generation systems. DeepDream is similar to the visualization technique we discussed previously, to try and visualize what a filter in a convolution layer does but defining a loss function of the output activation of a layer filter, and optimizing an initially random image to maximize the output activation of that one filter. This gives us some idea of exactly what features a particular filter in a convolution layer responds to. The DeepDream method uses the same idea with a few simple differences. With DeepDream you try to maximize the activation of entire layers, not just a single feature filter. Thus you define a loss function to maximize that combines the activations of all features from multiple layers of an existing trained convnet. Also we are interested in generating an interesting image that takes the features of an existing image and "dreams" or adds interesting features to existing elements of an image. So we start a DeepDream with an existing image instead of an initially random set of pixels. Resources: Textbook: Chollet (2022). "Deep Learning with Python (2ed)". Manning. https://www.amazon.com/dp/1617296864/?bestFormat=true&k=deep%20learning%20with%20python&ref_=nb_sb_ss_w_scx-ent-pd-bk-d_de_k0_1_15 CSci 560 Class Repository: https://github.com/csci560-nndl/nndl Contains video slides and iPython notebooks for this course. 00:00 Introduction 01:37 DeepDream algorithm 04:23 Implementing DeepDream in Keras 05:21 DeepDream loss function 10:56 Summary