README.md

neuron

Status: active.

This crate is the executable companion for 03 Neuron and 04 Learning.

It models the first honest training system in the course with:

• semantic newtypes for inputs, weights, bias, prediction, target, loss, and gradients
• a single-neuron forward pass with manual backpropagation
• SGD updates and epoch-level dataset training helpers
• tiny boolean datasets such as OR and AND
• token-target utilities that bridge scalar supervision to next-token loss
• a tiny bigram next-token model using token -> embedding -> lm_head -> logits

Layout

src/
  bigram.rs
  dataset.rs
  neuron.rs
  optimizer.rs
  token_targets.rs
  lib.rs
examples/
  train_bigram_cycle.rs
  train_or_gate.rs
  token_targets.rs

Run

cargo test --manifest-path code/neuron/Cargo.toml

cargo run --example train_or_gate --manifest-path code/neuron/Cargo.toml

cargo run --example train_bigram_cycle --manifest-path code/neuron/Cargo.toml

cargo run --example token_targets --manifest-path code/neuron/Cargo.toml

Scope

This crate is intentionally educational, not production-grade. It does not include:

• mini-batch loaders
• momentum or Adam
• autograd
• matrix-based multi-layer backpropagation
• full tokenizer or vocabulary tooling
• Transformer training

The goal is to make the first training loop and the token-target bridge explicit.