Hewlett Packard Enterprise (HPE) is launching HPE Swarm Learning, a software for decentralized AI training. This allows different locations or organizations to share AI training results with each other without exchanging raw data. […]
“Swarm learning is a new, powerful AI approach that has already made progress in addressing global challenges – such as improving healthcare and detecting anomalies in fraud detection and predictive maintenance,” said Justin Hotard, executive vice president and general manager of HPC &AI at HPE. According to Howard, HPE is making a significant contribution to the spread of swarm learning by providing a solution that is suitable for larger organizations to work with, innovate and increase the performance of their AI models – while at the same time complying with their ethical, data protection and regulatory standards.
HPE Swarm Learning Solves problems of Centralized AI Training
Today, AI model training usually takes place at a central location with centralized data sets. However, this approach can be inefficient and costly when large amounts of data need to be sent to one place. In addition, data protection or concerns about data sovereignty often prevent data from being centralized. The result may be that too little data is available for AI training.
HPE Swarm Learning enables organizations to use distributed data sources for AI training without transferring the source data. Instead, they share AI training results in the form of model parameters. This procedure is organized via a blockchain. For example, it controls the inclusion of swarm members and the recurring selection of a member who merges the model parameters in the respective training cycle. This gives the swarm network stability and security. In addition, large amounts of data can be tapped for AI training without compromising data protection or data sovereignty.
The possible areas of application of HPE Swarm Learning include, for example:
- Hospital: these can share AI training results from CT and MRI scans, for example, or from gene expression data with other hospitals in order to improve the diagnosis of diseases and at the same time protect patient data.
- Banks and financial service providers: they can combat credit card fraud by exchanging fraud-relevant model parameters with other financial institutions.
- Production sites: with swarm learning, they can improve their predictive maintenance by collecting training results from sensor data from several production sites.
For example, the first users of HPE Swarm Learning include the following organizations:
A team of cancer researchers at the University Hospital of RWTH Aachen University has conducted a study to improve the diagnosis of colorectal cancer by applying AI to image processing. The aim is to predict genetic changes that can lead to cells becoming cancerous. The researchers trained AI models with HPE Swarm learning with patient data from Ireland, Germany and the USA and compared the prediction performance with two independent data sets from the UK. The results showed that swarm learning outperformed the AI models that were trained only on local data.
TigerGraph, a provider of graph analysis solutions specializing in fraud detection, combines HPE Swarm Learning with its own analysis tools to detect unusual transactions in credit card transactions faster. The combined solution increases the accuracy of training ML models from huge amounts of financial data from several banks and branches in different locations.