Training a classifier entirely on a scientific calculator

Tall_Abrocoma_3533 · reddit · 2026-08-22

The author manually trained a binary classification model (distinguishing 0s and 1s) on a non-programmable Casio FX-82CE X scientific calculator.

Architecture & Process:

Downsampled MNIST images to 3x3 binary pixels.
Used a single fully connected layer (no bias) with a single output neuron.
Manual Training: Used a perceptron-style approach with only 6 images (3 per class), manually calculating weights: "0 0 1 -1 2 -1 -1 1 -1".
Results: Achieved 67.04% validation accuracy (correctly identified all 1s, but struggled with 0s).
Max Potential: After 1000 epochs of SGD training on a PC, the architecture achieved 98.96% validation accuracy.

More from Fun