KACPER MUTKEPortfolioPolski

VIDEO MAPPING / NEURAL NETWORKS · 2025

CONV2D

MULTIVERSE-METAVERSE · 11th MEDIATIONS BIENNALE — YOGYAKARTA · KALISZ · KANTON · POZNAŃ

Open in the portfolio: photos and video
CONV2D

A video projection in which the image is produced by passing letters through an untrained neural network.

The letters pass through successive layers of the network: convolutional, maxpool and dropout. The network has been taught nothing, so it does not recognise characters; its layers act as filters that break a letter down into numerical values and noise. Colour comes from mapping the internal data of the network directly to RGB channels, with no chosen palette, so the image literally shows the mechanism at work. On every run the layers receive random initial values and no sequence repeats. The starting point is the letter as a sign and as a visual form. For the network it remains only an arrangement of numbers.

NETWORK
untrained layers: conv2d · maxpool · dropout · fully connected
COLOUR
direct mapping of tensors onto RGB
SOFTWARE
TouchDesigner · PyTorch
PRINCIPLE
random initialisation each iteration — no repetition