Edifice Notebook Index

Overview

These Livebook notebooks demonstrate Edifice‘s neural network architectures through interactive, runnable examples. Each notebook is self-contained — pick any one and start exploring.

Setup: Each notebook has two setup cells — use “Standalone” for a normal livebook server, or “Attached to project” if you launched via ./scripts/livebook.sh (recommended for GPU/CUDA).

Notebooks

Getting Started

• Edifice: A Guided Tour — A comprehensive walkthrough of Edifice for newcomers. Covers classification, tabular data (TabNet with feature importance), uncertainty estimation (MC Dropout), time series forecasting (Mamba vs GRU), the architecture zoo, decision boundary comparison, and generative models (VAE). Heavy on charts and plain-English explanations — no ML background required.

• Architecture Zoo — Tour of Edifice’s 111+ architectures. Build MLPs, transformers, Mamba, Hopfield networks, and more with the unified Edifice.build(:name, opts) API. No training — just see what each architecture looks like and what shapes it produces.

• Training an MLP — Your first end-to-end training loop. Train an MLP and TabNet on a 3-class classification task using Axon.Loop.trainer. Covers data generation, training, and evaluation.

Sequence and Language

• Sequence Modeling — Compare Mamba, GRU, and GatedSSM on sine wave prediction. Shows how different sequence architectures handle temporal patterns, with overlay plots of predictions vs ground truth.

• Small Language Model — Build a character-level language model from scratch. Trains a GPT-style DecoderOnly transformer and Mamba on novel excerpts, with autoregressive text generation at different temperatures. Includes architecture swap comparison.

• LM Architecture Shootout — Head-to-head comparison of 8 architectures (Decoder-Only, Mamba, S4, GRU, xLSTM, MinGRU, Griffin, RWKV) on next-character prediction. Same corpus, same hyperparameters, same training loop — only the architecture changes. Includes loss curves, perplexity comparison, training time benchmarks, and text generation samples.

Graphs and Structure

• Graph Classification — Compare GCN, GAT, and GIN on synthetic graph classification. Demonstrates how different graph neural networks learn from node features and graph topology, with structure visualizations and accuracy comparison.

Generative Models

• Generative Models (VAE) — Train a Variational Autoencoder on 2D crescent moon data. Covers the VAE loss (reconstruction + KL divergence), the beta trade-off, latent space visualization, and generating new samples by decoding random latent points.

Comparison and Benchmarking

• Architecture Comparison — Head-to-head comparison of 8 architectures (MLP, TabNet, Bayesian, MCDropout, EBM, Hopfield, SNN, ANN2SNN) on 2D classification. Includes accuracy table, timing comparison, and decision boundary visualizations.