A series of beetles generated by a neural network
The project consists of a series of 25 beetles generated by a neural network using pix2pixHD.
The network learned to generate beetle drawings from abstracted representations. After training, it is capable of coming up with new beetles from fictive beetle abstractions I create myself.
The title, Genesis, refers to the process of manually creating datasets for generative machine learning (ML) projects. ML requires large amounts of data which, in many cases, is hard to come by. This is why we fall back to a few public repositories and, possibly, why research papers prove concepts by generating kittens or celebrity faces. We can also crowdsource datasets, but this comes with its own limitations and politics.
I sourced my own dataset at the Reanimation Library, a library of misfit books in Brooklyn. My beetles are based on "A Book of Beetles", published by Josef R. Winkler (author) and Vladimir Bohac (illustrator) in 1965. In a laborious process I scanned the book, transformed the data into a pix2pix training set, and used it to train the neural network.
The result, after many hours of work, is 25 bugs. My dataset is way too small for any serious application. Ultimately, the project is a reflection on where ML datasets come from and the politics inherent in their genesis.
Class: Major Studio 2, Spring 2018, MFA DT, Parsons School of Design
Advisor: Ernesto Klar
Open Source: pix2pixHD