These three things are the same story. US20220051104A1 (published February 17, 2022; assigned to a Microsoft research team) describes mapping classic ML pipelines onto neural-network frameworks, tagged G06N 3/0454, G06N 20/20 (ensemble methods), G06N 5/003. The pitch: don't rewrite your old models, just run them on the new hardware.
Connect the dots. A decade of enterprise value sits in non-deep-learning models, gradient-boosted trees, logistic regressions, feature pipelines. Those don't naturally run on GPUs. But if you can recast them as tensor operations, they ride the same accelerators that serve transformers, getting big speedups for free.
“Methods, systems, and computer program products are provided for generating a neural network model. A ML pipeline parser is configured to identify a set of ML operators for a previously trained ML pipeline, and map the set of ML operators to a set of neural network operators.”— U.S. Patent Application 2022/0051104 A1 source
Follow both the money and the IP and the business logic is obvious. Microsoft sells the cloud those accelerators live in. Anything that pulls more workloads, including legacy ML, onto GPU infrastructure increases utilization of expensive hardware the company already bought. A patent that bridges old models to new chips is a patent that fills data-center capacity.
This reframes the is the AI buildout overbuilt debate. If the same accelerators can serve both frontier models and the long tail of classic ML, the hardware is less of a one-trick bet. The 2022 filing is an early, concrete move toward making AI silicon a general acceleration platform, not just a transformer engine.
House caveat: a publication describes a method, and whether traditional pipelines actually benefit depends on the model and batch size. But as a marker it's clean, by early 2022, a hyperscaler was patenting ways to drag the entire ML back catalog onto its AI hardware. Old models, new chips, same bill.