From Llull's combinatorial wheels and Leibniz's calculus of reasoning to today's large language models: ideas, programs, machines, games, people, companies and the policy around them. Moments are tagged (symbolic AI, machine learning, neural networks, robotics, games, language, vision, policy & safety, people, industry) and colored by their first tag; the field's eras, including both AI winters, are shown as spans. Each moment links to its sources on Wikipedia; summaries are our own, and dates given only as a year or month are marked approximate where uncertain.
Compiled by an AI assistant (Claude) through the OpenChronology API and reviewed by the account owner. Coverage of the most recent years stops at the assistant's knowledge.
Duration correction: Ongoing subjects previously drawn through a present/research cutoff are represented at their original starts, with no end asserted. Existing moments at those coordinates are reused; their attached cards retain the former duration’s notes, sources, tags and media. Bounded historical and coverage intervals remain durations.
84 moments · 1305 CE – 2025-02-10 CE · by @athan · @athan
Residual networks
neural networks
vision
Kaiming He and colleagues at Microsoft Research introduced residual networks, using skip connections to make very deep models easier to optimize. Instead of requiring each stack of layers to learn an entirely new transformation, the design supports learning a residual relative to an existing path. Strong ImageNet results in 2015 helped establish the architecture’s influence. Greater depth alone was not the complete explanation; optimization and the network structure mattered. Residual connections later appeared in many model families beyond the original image-classification setting. **Timeline anchor:** 1 December 2015 CE. **References:** [Source 1](https://en.wikipedia.org/wiki/Residual_neural_network).