THEME
AI & Machine Learning Patents
Expired and disclosed patents that shaped the AI / ML stack.
9 episodes
- AI & ML PATENTS #6Asking a Database in Plain English: A 1989 Patent That Posed the NL2SQL QuestionAI & ML Patents Memo #6 — US5197005A, Intelligent Business Systems, filed 1989Filed in 1989, US5197005A describes a rule-based system for querying databases using natural language — a knowledge base plus expert system approach to the problem that LLMs now solve with probabilistic generation. The question is the same; the method is fundamentally different. Primary source URL confirmed; full text not yet read.
- AI & ML PATENTS #5Replacing Synaptic Weights with Filters: A 1991 NASA Patent on Time-Series Neural NetworksAI & ML Patents Memo #5 — US5253329A, NASA, filed 1991Filed in 1991, US5253329A by NASA researchers Villarreal and Shelton replaces scalar synaptic weights with adaptive digital filters, encoding temporal dependencies directly into network connections. A prior example of the same problem intent that later led to LSTMs and Transformers. Primary source URL confirmed; full text not yet read.
- AI & ML PATENTS #3Running Backpropagation on Dedicated Hardware: A 1993 Philips Patent and What It Tells UsAI & ML Patents #3 — US5517598A, US Philips Corp, filed 1993Filed in 1993, US5517598A describes a hardware architecture where the same processor structure handles both forward inference and error backpropagation, using transpose matrices to route error signals. The problem it posed — running learning on dedicated hardware — overlaps in intent with modern GPU and TPU design, though the underlying architecture is fundamentally different. Full text retrieved from Google Patents.
- AI & ML PATENTS #5Bell Labs Filed a Multi-Resolution Symbol Recognition Patent in 1992. The Design Logic Looks Familiar.AI & ML Patent Note #4 (memo) — US5337372A, AT&T Bell Labs, coarse-to-fine symbol recognitionFiled 1992 by LeCun and Wu at AT&T Bell Labs, US5337372A describes a symbol recognition system that uses low-resolution feature arrays to narrow candidates, then confirms matches at higher resolution — reducing computation. Patent URL confirmed; full text unread.
- AI & ML PATENTS #4LeCun's 1994 Patent Trained Neural Nets to Ignore Rotation and Shift Using Tangent VectorsAI & ML Patent Note #3 (memo) — US5572628A, Lucent Technologies, tangent-vector invariant trainingFiled 1994 by LeCun, Denker, Simard, and Victorri at Lucent Technologies, US5572628A describes a method for making neural networks invariant to transformations like rotation, scaling, and translation — using tangent vectors instead of data augmentation. Patent URL and abstract confirmed; full text unread.
- AI & ML PATENTS #2Bell Labs Filed a Patent in 1989 That Described Weight Sharing — The Same Assumption Modern CNNs Are Built OnAI & ML Patent #2 — US5067164A: LeCun et al.'s 1989 AT&T patent on hierarchical constrained neural networksFiled November 1989 by John S. Denker, Richard E. Howard, Lawrence D. Jackel, and Yann LeCun at AT&T Bell Laboratories, US5067164A describes a layered network for optical character recognition using weight sharing: 90,000 connections represented with approximately 2,600 free parameters (97% reduction). Google Patents full text retrieved. Comparing the design problem to modern CNNs 36 years later.
- AI & ML PATENTS #3IBM Filed a Statistical Translation Patent in 1991. Here Is What Problem It Was Trying to Solve.AI & ML Patent Note #2 (memo) — US5477451A, IBM's statistical machine translation systemFiled 1991 by Peter F. Brown, John Cocke, Frederick Jelinek, and colleagues at IBM Research, US5477451A describes learning translation probabilities from parallel corpora. Patent URL confirmed; full text unread. This is a record of what the design was trying to do — not a claim that it leads to LLM translation.
- AI & ML PATENTS #2The Patent That Ranked Every Web Page: PageRank, Filed 1997AI & ML Patent Note #1 (memo) — US6285999B1, Lawrence Page's linked database ranking methodFiled 1997 by Lawrence Page, US6285999B1 describes a method for ranking nodes in a linked database by computing how often a random link-follower would land on each page. Patent URL confirmed; full text unread. Exploring connections to LLM search and graph-based RAG.
- AI & ML PATENTS #1Amazon Filed a Patent in 1998 That Described Exactly How 'Customers Who Bought This Also Bought' WorksAI & ML Patent #1 — US6266649B1: the design logic behind item-to-item collaborative filtering at scaleFiled 1998 by Gregory Linden, Jennifer Jacobi, and Eric Benson, US6266649B1 describes a recommendation system that precomputes item-to-item similarity tables offline and serves personalized recommendations at low latency by looking up and combining those tables in real time. Expired 2018. Distinct from embedding-based recommendation but shares the same design problem.