Decentralized X: Aggregating Heterogeneous and Decentralized AIs

Current challenges in training AI systems

AI technologies have come into use in various domains. However, it still remains an open question how to prepare a sufficient amount and variety of data for training AI systems to achieve a required performance.

Problem 1: Data distributed unevenly in various places
Problem 2: Difficulty in aggregating client data

Decentralized X: Aggregating Heterogeneous and Decentralized AIs

Framework to aggregate machine learning models trained in heterogeneous and decentralized environments

Decentralized X is aimed at training a high-performance machine learning model by collecting and aggregating models trained at the client side, rather than aggregating the original data owned by each client. This framework allows us to develop AI systems that have the same performance as if they were trained from all the client data.


Key features of Decentralized X

Unlike other related technologies, Decentralized X can have clients holding their own model definition and data, while allowing a server to aggregate these client models to improve their performance.

Google Federated Learning: Collaborative Machine Learning
without Centralized Training Data Thursday, April 6, 2017

Applications to manufacturing
Visual inspection system for detecting defects

AI technologies have come into use in manufacturing, such as visual inspection systems for detecting defects in products. However, it is hard to collect a huge amount of product images with defects as their training data resource because of the fundamental shortage of such defect examples. Moreover, collecting defect images from multiple factories would also be hard as they are often regarded as confidential.

With Decentralized X, we can develop a high-performance visual inspection system that can recognize a variety of defects by sharing and aggregating models trained at each site, without sharing the original training data. With the aggregated model, each site can then detect a new type of defect that it has never observed before.