Hand-rolled neural network in a Swift Playground reads 8×8 emojis

What it does EmojiIntelligence is a hand-written neural network implemented entirely in Swift inside an Xcode Playground. You draw an emoji on an 8×8 grid; the code flattens pink pixels into 64 binary numbers and pushes them through a tiny fully-connected network to produce a classification. It is explicitly a learning exercise, not a production framework.

The interesting bit Instead of importing Core ML or TensorFlow, the author wired up a three-layer network—64 inputs, two hidden neurons, one output neuron—with sigmoid activation and connection weights directly in Swift. That makes the inference pipeline inspectable in a way that massive frameworks rarely allow.

Key highlights

• Built from scratch in Swift with no external ML dependencies
• Runs entirely inside an Xcode Playground on macOS
• Uses a fully-connected architecture: 64 binary inputs → 2 hidden neurons → 1 output neuron
• Applies sigmoid activation to squeeze outputs between 0 and 1
• Open-sourced as an educational toy to demystify how layers and weights actually work

Caveats

• Currently macOS-only; the author reports a bug blocks it from running in iPad Playgrounds
• The README does not specify which emojis are supported, training set size, or accuracy metrics

Verdict Grab it if you are a Swift developer who wants to see how weights and activation functions behave without framework magic. Skip it if you need a real-world classifier or iPad support.