A neural network that doesn’t just see images — it remembers them. it learns the essence of handwritten digits from mere coordinates.
This isn’t your usual convolutional wizardry. Instead of filters and feature maps, we’re training a fully-connected neural network to memorize an MNIST digit — by mapping normalized (x, y) coordinates to pixel intensities.
Yes, the network becomes the image.
It’s like asking: “What would a neural network dream of if it only knew where on the canvas it was?”
For each grayscale MNIST image:
- Normalize the (x, y) pixel coordinates to [-1.0, 1.0].
- Use them as input to the neural network.
- Train the network to output the corresponding grayscale pixel value.
- Once trained, sample more densely to reconstruct a higher resolution version of the image.
- No image processing libraries — the network is the renderer
- Learns pixel intensities from coordinates alone
- Upscales MNIST digits to arbitrary resolutions
- Saves output as .png files for easy viewing
Yes, the blurry digits meditate into high-res serenity.