The Potential of Swarm Learning (HPE)

The Expert Group took place as a virtual meeting on February 15, 2022. We were joined by 12 participants.

Tim Geppert from ZHAW opened the meeting and introduced Alexander Volk, Hartmut Schultze and Roger Fontana from HPE. He also introduced the speaker for the upcoming meeting, namely Andrew Knox. Andrew will introduce the group to the basics of differential Privacy.

Afterwards Alexander Volk, Hartmut Schultze and Roger Fontana, all experts from HPE in the areas of swarm learning, presented the concept of swarm learning. An overview about the content of the talk is given in the following sections. The meeting was closed with a discussion of potential use cases for this technology.

HPE Swarm Learning: Reduce bias in your ML models (and enjoy the side benefits)

By Alexander Volk, Hartmut Schultze, Raymond Freppel, Roger Fontana

The convergence of algorithmic advances, data proliferation, and tremendous increases in computing power and storage has propelled AI from hype to reality.[1] The quality of the results of AI applications are related to the underlying algorithm and the accessible training data. Both include bias, resulting in the (unintentional) inclination toward certain type of results. One of the early examples is the St. George’s Hospital Medical School during 1982 to 1986, where 60 women and ethnic minorities were denied entry due to a new computer-guidance assessment system that denied entry to women and men with “foreign-sounding names” based on historical trends in admissions[2]. This is only one example of many.

Machine Learning is here to stay, and it is upon all of us who are involved in creating and applying ML to reduce biases. To understand the impact of HPE Swarm Learning on the reduction of bias and the possible improvements it might help to have a short look back at the history of model training.

Initially, Local Learning was performed where local models were trained at each data source. This approach resulted in local data bias, plus inaccurate and suboptimal models delivering inferior results. Next came Centralized Learning to reduce bias and improve model accuracy. This improved model accuracy since larger data sets were used for training. However, data aggregation to a central location posed new challenges around data privacy, data ownership, and data movement. So along came Federated Learning to alleviate central learning challenges with a central custodian responsible for aggregating all learnings from multiple data sources while also preserving privacy. The central custodian, however, poses challenges around resilience, scalability, ownership and central power given to such persona. With the central custodian, local high-availability features may be used for resilience. But, if the central aggregator goes down, training stops, and scalability is limited to what the client-server model in a predominantly star topology can support. It’s time for a new modern approach to machine learning.[3]

The HPE Swarm Learning solution is a decentralized, privacy-preserving, and collaborative machine learning framework at the data source. Users of the easy-to-use API which architecture is based on Blockchain can benefit from data accessibility across geographies and across organizations, leading to larger data sets and an increase of their model accuracy. Since Swarm Learning negates copy or movement of training data, it works on heterogeneous infrastructure and therefore represents an efficient way to resiliently scale.

Does Swarm Learning work?

In June 2021, a study proved Swarm Learning’s feasibility. The classifiers[4] of Swarm Learning achieved higher accuracy than those developed at individual sites. In addition, Swarm Learning completely fulfils local confidentiality regulations by design.[5]

While Healthcare and life sciences have several well-suited use cases Swarm Learning, it can be applied cross all industries. The federal government sector can apply Swarm Learning for detecting anomalies and threats, as well as for research collaboration. Financial services organizations apply Swarm Learning to financial fraud detection activities. And in manufacturing, Swarm Learning provides predictive maintenance capabilities.

If you want to learn more about a solution that enlarges your dataset without moving or duplicating raw data while you can keep your data sovereignty and privacy, or simply want to learn more, you can always reach out to HPE. We are here to help.

Further Links:

Nature.com (2021). “A safer way to share health data“

Technical White Paper Swarm Learning

[1] Michael Chui; Vishnu Kamalnath Brian McCarthy (2020). An executive’s guide to AI. McKinsey.

[2] Nitin Aggarwal (2020). Biases in Machine Learning. towardsdatascience.com

[3] Arshad Khan (2022). HPE Swarm Learning: Increase accuracy and reduce bias in AI models. community.hpe.com

[4] In data science, a classifier is a type of machine learning algorithm used to assign a class label to a data input. An example is an image recognition classifier to label an image

[5] Warnat-Herresthal, S., Schultze, H., Shastry, K.L. et al. Swarm Learning for decentralized and confidential clinical machine learning. Nature 594, 265–270 (2021). https://doi.org/10.1038/s41586-021-03583-3