Deep Learning Showcase

Here you can find a few examples of what we were able to do with a server setup for deep learning.
Contents:

• Coloring b&w images
• Bizarre image with Google's DeepDream

With a deep learning server we were able to train a network to take a black and white image (on the left) and add color back to it. This was a achieved by training a Convolutional Neural Network on black and white images and the colored version of each image. Over time it was able to learn what a black and white image should look like like in color, allowing us to feed in an image of our own and have it predict what the original looks like.

Heres the link to the original github project:
https://github.com/richzhang/colorization

Instructions: Can you figure out which image is the original? On the left is the black and white version of the image. One of the images on the right is the original and one of them has been colored via the convolutional neural network.
Click on the one that you think is the original and figure out if you were correct!

Generate psychedelic images algorithmically! This showcase warps Death Valley into a trippy stadium using the Google DeepDream algorithm and the Tensorflow framework.

By looking at the features expressed through deep dream, we can start to build an intuition about underlying features. Objects that are expressed in the image align more with the network's weights.

Repeated calling then excenuates the features further as we have already added them once.

You may want to download the ipython notebook due to its size. It has visualizations for a few layers ran so that you don't need to!

Heres the link to the github project with more tutorial information:
Feature Vizualization with Deep Dream

Death Valley

This is original image of Death Valley, the pride and joy of Clemson!

Death Valley Flowerfied

This is the image after running it through the deep dream algorithm multiple times, enhancing the features that express flowers.

This is expressing lower level node features in the deep learning network.

We can see that there is a lot of merging of aspects of the image.

Death Valley tiles

This is expressing high level node features in the deep learning network.

We can see that there are only minimal changes to the image now.

Looking at features through random noise enhancement

This is an interesting high level (near top of network) feature. We can find interesting channels by finding weights that maximize the sum of the outputs. This weirdly patterned image was visualized through running deep dream algorithm version on random noise.

We can see how those features manifest into its changes to the stadium.