ramnet-study

Code and frozen results for Forgotten Architectures #1 — Weightless Neural Networks on Modern Compute. Paper: https://vulkgryph.com/research/nn-revival/papers/weightless-neural-networks Series: https://vulkgryph.com/research/nn-revival/

A controlled, honest re-run of weightless neural networks — Aleksander's WiSARD and Kanerva's Sparse Distributed Memory — on binarized MNIST, together with two of our own extensions (gradient-trained contents; a frozen/plastic dual-mode hybrid) and small-model baselines. Predictions are pre-registered and frozen; three seeds; results frozen and hashed.

Arms

arm	what it is
wisard	faithful write-based WiSARD (fixed n-tuple RAM addressing)
trained_wisard	same fixed addressing, gradient-trained contents
reservoir	fixed random binary projection + trained readout
sdm	Kanerva Sparse Distributed Memory
mlp	one-hidden-layer dense baseline
hybrid	frozen (gradient) + plastic (one-shot write) sections for class-incremental continual learning

Reproduce

Requires a recent Rust toolchain and Python 3 (standard library only).

1. MNIST. Place the four standard MNIST IDX files in data/mnist/:

data/mnist/train-images-idx3-ubyte
data/mnist/train-labels-idx1-ubyte
data/mnist/t10k-images-idx3-ubyte
data/mnist/t10k-labels-idx1-ubyte

(Standard, widely-mirrored dataset. The runners require real MNIST and abort rather than silently falling back.)

2. Run the campaigns:

cargo run --release -- --full         # baseline 4 arms (Q1 sample-efficiency, Q2 WiSARD N-scaling)
cargo run --release -- --new-arms     # wisard / trained_wisard / sdm (Q1, Q3, Q4)
cargo run --release -- --hybrid-full  # hybrid continual learning (Q5)

3. Regenerate the paper's tables from the frozen results:

python3 gen_tables.py

Every figure in the paper is generated from results/*_frozen/*.json by this script — no result numbers are hand-typed.

Results & reproducibility

• Frozen results live in results/*_frozen/ (JSON + Markdown). Their SHA-256 hashes are recorded in the paper.
• Predictions were written in SPEC.md before each multi-seed run and are append-only — outcomes are added; predictions are never edited.

Scope

This studies weightless architectures in their original form under modern controls, plus our own labeled extensions. It is a toy-scale study on binarized MNIST with small models — not a SOTA claim (a CNN would beat every arm here). See the paper's Scope and Limitations sections.

Contact

See vulkgryph.com.

License

Copyright 2026 Vulkgryph LLC. Code licensed under Apache-2.0 — see LICENSE and NOTICE.

---

Code and experiments produced with AI coding agents under the author's direction; figures are generated from the frozen results by gen_tables.py and audited by the author.