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
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










