patentml 0.1.0

37 machine learning algorithms reconstructed from expired US patents. Zero dependencies, pure Python stdlib.

patentml

Machine learning from expired patents. Zero dependencies. Pure Python stdlib.

37 modules, 131 classes and functions — every algorithm reconstructed from a United States patent that has expired into the public domain. The patents that built modern ML were filed by IBM, Bell Labs, Microsoft Research, Lockheed, AT&T and Lucent between 1995 and 2006. They have all expired. This library is what they describe, as clean modern Python, with no imports beyond the standard library.

pip install patentml

No numpy. No scipy. No compiled extensions. If it runs Python 3.8+, it runs patentml — locked-down corporate machines, serverless functions, air-gapped environments, Pyodide in the browser, and (with light trimming) MicroPython boards.

Quick start

from patentml import RandomForest, ScalableKMeans, ThompsonSampling, KalmanFilter

# Classification — US6816847 (Microsoft, 1999)
forest = RandomForest(n_trees=25)
forest.fit(X_train, y_train)
labels = [forest.predict(x) for x in X_test]

# Clustering — US6012058 (Microsoft, 1998)
km = ScalableKMeans(k=3)
km.fit(points)

# Bandits — US6981040 (Utopy, 2000) [919 forward citations]
bandit = ThompsonSampling(n_arms=4)
arm = bandit.select()
bandit.update(arm, reward=1.0)

# State estimation — US6795794 (Univ. Illinois, 2002)
kf = KalmanFilter(dim_state=2, dim_obs=1)

What's inside

Family	Modules
Evolutionary & global optimisation	genetic algorithm, genetic programming, grammar GP / grammatical evolution, linear GP, particle swarm, differential evolution, CMA-ES, simulated annealing, ant colony, Bayesian optimiser / EDA, neuroevolution
Neural networks	mini neural net (mini-batch backprop), Conv1D, SimpleRNN, GRU cell, SGD/RMSProp/Adam/AdamW optimisers
Classifiers	decision tree, random forest, AdaBoost, SVM (SMO), online Bayes, naive Bayes, KNN (+ BallTree), gradient boosting
Ensembles	voting, stacking, bagging, weighted
Clustering	scalable & hierarchical k-means, DBSCAN, OPTICS, EM / Gaussian mixture, spectral, mean shift
Reinforcement learning	Q-learning, SARSA, function-approximation Q, actor-critic A2C, PPO-lite, ε-greedy / UCB1 / Thompson / EXP3 / LinUCB bandits
Probabilistic	Bayesian network, hidden Markov model, Gaussian process regression & classification, kernel density estimation
Anomaly detection	isolation forest, one-class SVM
NLP	TF-IDF + naive Bayes text pipeline, word2vec SGNS, PMI embeddings
Recommenders	memory-based & Bayesian collaborative filtering
Dimensionality & features	PCA, randomised SVD, vector quantisation (LBG, product quantiser), scalers, mutual-information ranking, forward selection
State estimation	Kalman filter, extended Kalman filter

Provenance

Every module documents its source patent: number, assignee, filing year, and forward-citation count. Highlights:

Patent	Assignee	Algorithm	Citations
US5613012	SmartTouch (1995)	Voting ensemble	1,182
US6981040	Utopy (2000)	Bandit selection	919
US6161130	Microsoft (1998)	Online classifier	896
US6556983	Microsoft (2000)	Word embeddings (PMI + SGNS)	645
US6192360	Microsoft (1998)	TF-IDF + naive Bayes	364
US6317707	AT&T (1998)	Mean shift + KDE	269
US6931384	Microsoft (2001)	Gaussian process regression	258

The full list of ~40 source patents is in the package docstring: python -c "import patentml; print(patentml.__doc__)".

All source patents are expired. The implementations are original code, MIT licensed.

Why

Modern ML stacks are heavy, opaque, and supply-chain risky. Sometimes you need one algorithm — a Kalman filter on a microcontroller, a bandit in a serverless function, k-means in a browser — without 200 MB of compiled wheels. And sometimes you want code you can actually read: every module here is a single self-contained file you can audit in one sitting.

These algorithms earned their citations the hard way. They still work.