team Deep Dreamers
Team Lead: Shuhong (Shu) Chen
AI4ALL Facilitators: Harpreet Multani, Savva Petrov
Team members: Chirag, Anthea, Nafees, Zenon, Zameel, Andrew
Project Overview:
In our project, our goal was to generate an image using artificial intelligence. We used neural networks, deep dreams, and machine learning to teach the computer to create new artwork given some images.
The computer created new artwork (?) given the image of dog and Mona Lisa.
Another computer created artwork of Van Gogh
Deep Dream of Starry Night
Deep dream of Dogs Playing Poker
Deep dream of an Owl
Project Question:
How can computers learn how to generate new images?
Actions Steps:
We first worked on being able to adapt a model to better fit a data set, using loss tracking to find the "line of best fit" with the lowest mean squared error. We then replaced the dataset with our own images and a pretrained model. Using a selected category from the model, the loss then becomes the difference in the pretrained model and our images. This results in the code attempting to "fix" our input images, resulting in distortions.
We also briefly worked with style transfer, which was mostly the same process except our loss had to be the differing styles of two images (usually the unique texture of an image), resulting in an image that looks to be the combination of two images.
Results:
Everyone was able to create some level of "trippyness" in their images
Deep Dreams:
Style Transfer:
