The Long Detour — Machine Learning Before the GPU

The Long Detour<br>Machine Learning Before the GPU, 1943–2011 · in three volumes<br>assembled and directed by Remy Ochei<br>written with an event-sourced multi-agent harness

1943 · McCulloch–Pitts

1958 · perceptron

1974 · expert systems

1986 · backprop

1995 · SVM<br>2011 · the eve of the GPU<br>▶ watch it being written — the tape, all 13,539 events, replayed

Volume I — A Narrative History of Machine Learning, 1943–2011

Two Founding Bets: Symbols and Connections

The Rule-Based Ascendancy: Expert Systems and the Knowledge Engineering Dream

The Connectionist Counterattack: Backpropagation and Parallel Distributed Processing

Neural Networks Go to Work

The Statistical Turn: Bayesian Networks and Probabilistic AI

The Kernel Insurgency: Vapnik and the Statistical Learning Theorists

Committees of Weak Learners: The Ensemble Revolution

Reinforcement Learning and the Behaviorist Thread

Evolution as Optimizer: Genetic and Swarm Algorithms

The Empiricist Turn: Text, Web, and Recommendation

Vision Without Understanding: Statistical Pattern Recognition

Epilogue: On the Eve of the GPU

Volume II — A Graduate Text

Linear Discriminants and the Perceptron

The Multilayer Perceptron and Backpropagation

Convolutional and Self-Organizing Architectures

Energy-Based Networks: Hopfield Nets, Boltzmann Machines, and the Helmholtz Machine

Recurrent Networks and Sequence Learning

Statistical Learning Theory and the VC Framework

Kernel Methods and the Support Vector Machine

Bayesian Networks and Probabilistic Graphical Models

The EM Algorithm and Latent Variable Models

Decision Trees and Rule Induction

Ensemble Methods: Bagging, Boosting, and Random Forests

Reinforcement Learning: Value Functions and Temporal-Difference Methods

Evolutionary and Swarm Algorithms

Unsupervised Learning: Clustering, Dimensionality Reduction, and Manifolds

Semi-Supervised and Transductive Learning

Volume III — A Practitioner's Guide to Small-Data, Low-Compute Learning

Learning from a Handful of Examples: Instance-Based and Local Methods

Feature Engineering and Selection Before End-to-End Learning

Trees and Forests: Interpretable Workhorses

Support Vector Machines for Small, Messy Data

Handling Imbalanced, Costly, and Drifting Data

Ensembles and Model Combination in Practice

Cross-Validation, Model Selection, and Honest Error Estimates

Case-Based and Analogical Reasoning as an Alternative to Learning from Scratch

Fuzzy and Neuro-Fuzzy Systems for Ill-Specified Domains

Practical Bayesian Networks and Expert Systems

Clustering and Unsupervised Discovery on a Budget

Recommender Systems, Time Series, and Domain Case Studies

HARPY · SPEECH UNDERSTANDING · CMU 1976<br>VOCABULARY ......... 1,011 WORDS<br>NETWORK STATES ..... 15,000<br>COMPILE TIME ....... OVER 13 HRS, DEC-10 (KL)<br>THROUGHPUT ......... 80 X REAL-TIME<br>PROCESSOR .......... 0.35 MIPS PDP-KA10<br>STATUS ............. IT WORKED.<br>$5<br>PER SENTENCE.

— cost target, HARPY speech understanding system, 1976

THE ALGORITHMS WERE READY.<br>THE MACHINES WERE NOT.

the print edition, concepted — die-cut sprocket holes,<br>red-sprayed edges. A rendered mockup, not yet an object.

Colophon

This book was drafted, revised, and fact-checked by an event-sourced<br>multi-agent system — an author, an adversarial critic, a planner, and an<br>arbiter — working over a corpus of ~5,000 primary papers, directed and<br>assembled by its human author. Nothing here rests on trust in the process:<br>the process is published. Every sentence traces to logged reads of the<br>primary sources; every quotation was machine-verified verbatim against the<br>paper it cites; every number in an experimental rerun comes from an executed,<br>independently reproduced computation; and the full construction record — the<br>tape — replays to exactly this text.

Words176,659<br>Sections117<br>Construction events on the tape13,539<br>Executed experiments398<br>Critiques filed and resolved175

Preprint, text only — figures are forthcoming. Source, tape, lab, and<br>harness:<br>github.com/doInfinitely/long-detour.

Even the jacket keeps receipts: the cover concepts and the<br>design<br>transcript are in the repository, including the quote check that caught a<br>mocked-up cover inventing archival telemetry. Jacket copy passes the same<br>quote check as the prose.

remyochei.com ·<br>The Bracket Studio