Sketched and clustered federated learning with automatic tuning

ChangeBridge: Patent Grants - AI & Computing (G06N)

Published March 31st, 2026

Detected March 31st, 2026

Summary

USPTO issued Patent US12591807B2 to IBM on March 31, 2026, covering a computer-implemented method for automatic adaptive client selection in federated learning using sketch-based clustering and automatic tuning. The patent includes 20 claims relating to optimizing client clusters and sketch dimensions subject to memory and communication constraints.

View original document View source feed page

What changed

USPTO granted Patent US12591807B2 to International Business Machines Corporation for a method of sketched and clustered federated learning with automatic tuning. The patent covers systems where a server sends ML model parameters to clients, receives gradient sketches, computes client similarity, clusters clients, optimizes cluster numbers and sketch dimensions under constraints, and selects a subset of clients for gradient aggregation. Application No. 18149682 was filed January 4, 2023.

Patent grants do not impose compliance obligations on third parties. Technology companies developing federated learning systems should consider reviewing the patent claims for potential licensing implications or to assess freedom to operate. IP counsel may be consulted regarding any overlap with existing or planned machine learning implementations.

Source document (simplified)

← USPTO Patent Grants

Sketched and clustered federated learning with automatic tuning

Grant US12591807B2 Kind: B2 Mar 31, 2026

Assignee

International Business Machines Corporation

Inventors

Arpan Mukherjee, Georgios Kollias, Theodoros Salonidis, Shiqiang Wang

Abstract

A computer-implemented method, a computer program product, and a computer system for automatic adaptive client selection in federated learning. A server sends parameters of a machine learning model to all of clients, where all of the clients compute respective gradients using the parameters. The server receives sketches of the respective gradients, where the sketches are computed by all of the clients. The server uses the sketches to compute similarity between all of the clients and clusters the all of the clients based on the similarity. The server optimizes a number of client clusters and a dimension of the sketches, subject to a constraint of memory consumption, a constraint of communication overhead, and a performance metric. The server determines a subset of the clients that send the respective gradients, by selecting the clients from the client clusters. The server aggregates the respective gradients sent by the subset of the clients.