Jon and Catie walk through US Patent 10,417,562, Google's patent on batch normalization, a layer added inside a neural network that rescales the numbers passing through it while the network trains. Starting from what a neural network is and what goes wrong during training, they build up to what claim one covers, what it leaves out, why the first, provisional application was filed fourteen days before the research paper, how the patent family grew to nine United States patents, and four drafting lessons for anyone patenting artificial intelligence.
Sources:
US 10,417,562 B2, Batch normalization layers: https://patents.google.com/patent/US10417562B2/en
US 10,628,710 B2: https://patents.google.com/patent/US10628710
US 10,902,319 B2: https://patents.google.com/patent/US10902319B2/en
US 11,281,973 B2: https://patents.google.com/patent/US11281973B2/en
US 11,308,394 B2: https://patents.google.com/patent/US11308394B2/en
US 11,853,885 B2: https://patents.google.com/patent/US11853885
US 11,893,485 B2: https://patents.google.com/patent/US11893485B2/en
US 12,033,073 B2: https://patents.google.com/patent/US12033073B2/en
US 12,705,484 B2: https://patents.google.com/patent/US12705484
US 2025/0013864 A1: https://patents.google.com/patent/US20250013864A1/en
EP 3,251,059 (European family): https://patents.google.com/patent/EP3251059A1/en
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift (arXiv): https://arxiv.org/abs/1502.03167
35 U.S.C. 119: https://www.law.cornell.edu/uscode/text/35/119
35 U.S.C. 154: https://www.law.cornell.edu/uscode/text/35/154
Read the article: https://burbagelaw.com/podcast.html
Narrated with the assistance of AI. Nothing in this podcast is legal advice.