I have always been inspired by Artificially Intelligent systems. For a lot of people, it is a 'black box', but I wanted to dive deep into its inner workings. Tensorflow 2.0 along with the Lucid library enabled me to manipulate the visuals that I generate, and the results were spectacular.
What it does
This is a creative experiment to see what can be generated using just code and concepts of machine learning using Tensorflow2.0. Basically, the code produces feature visualizations of inner layers of neural network architectures. There's a lot of trial and error involved while trying to find patterns which are compelling, thus putting the human in charge of this process. The collaborative effort between the human and the machine produces incredible visuals, something which was never possible before.
How I built it
I used Tensorflow 2.0 along with the Lucid library by Google to visualize inner layers and neurons of an inception network.
Challenges I ran into
A lot of computing power was needed to generate the graphics. I used Google's free collaboratory notebook's GPU runtime to generate the visuals.
Accomplishments that I'm proud of
I was able to import and run big models into Google Collab and execute it in its runtime. I was also able to manipulate and pinpoint certain layers which generated interesting visuals using Lucid and Tensorflow.
What I learned
I learned to work with big pre-trained models. I learned about the inner layers of popular neural net architectures and what is possible in the world of generative art using machine learning. I also learned to leverage a lot of open source python libraries and packages which I had previously never known about.
What's next for Creative Artificial Intelligence with Tensorflow2.0
I plan to create a series of artwork based out of different neural net architectures like ResNet50, AlexNet, and do a comparative and artistic analysis.