Slow walk through a convolutional neural network @animations_ag
Slow walk through a convolutional neural network  @animations_ag
Uploaded September 2026 | Updated November 2026, 1 hour ago
Inference of a handwritten digit of the MNIST dataset represented as a 28x28 pixel raster image is shown in a hopefully clear way. The convolutional neural network, CNN, is scarcely trained for the sole purpose of the final probabilities to be slightly more spread out. Further, the architecture is just an example, although it seems working pretty well for sufficiently trained weights/filters.

This is a re-upload. This since the original had some annoying re-encoding glitches I couldn't stand.

Music by @gpcbass , called SevenEight made yesterday.

Languages/tools: Python, JavaScript, GPT 6 Astra, FFmpeg
Slow walk through a convolutional neural networkChaos simulation using 1000 uncoupled double pendulumsBebop drone takeoffAnnoying maze billiardOptical illusion by Fourier epicyclesSupersonic mouse slides into mooseFourier series animation square waveBilliard of expressions, fruits, symmetry, etc.Billion ball cardioid billiard (chaotic)Large ball clusters of interacting balls dropped on a flat surfaceBad stringartFourier series animation semicircle
Alexander Gustafsson |

Slow walk through a convolutional neural network

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER