Jeremiah and Laura Johnson
These works are collaborations between the husband and wife team of
Jeremiah Johnson, a computational mathematician, and Laura Johnson, a
painter. Jeremiah developed a generative neural network that uses a
sequence of linear and nonlinear transformations to produce images
from randomly generated input vectors. Such visualizations can be
helpful in understanding the transformation a neural network produces.
In these works, the parameters of the linear transformations are
sampled from Gaussians. Hyperbolic tangents provide the nonlinear
transformations. The raw images this network produces are often
interesting, but with careful manual cropping to balance the
compositions and to focus attention on compelling aspects, they become
art.
This is a collaborative piece produced in two stages. First, a
generative neural network was used to generate a raw image. The
neural network is a function $\mathcal{F}:\mathbb{R}^p \to
\mathbb{R}^{m\times n}$ that maps a latent vector through a
sequence of linear and nonlinear transformations to produce the
raw image. The parameters of the linear transformations are
sampled from Gaussians, hyperbolic tangents provide the nonlinear
transformations, and the network is not optimized in any way; as
such, the raw output image provides a way to visualize the
transformation the network implements before it has been trained.
We then manually cropped the raw image to balance the composition
and to focus on a particularly compelling aspect of it.