Advanced Computing in the Age of AI | Monday, July 22, 2024

Increasing AI model accuracy and reducing model bias with swarm learning 
Sponsored Content by HPE

Discussing swarm learning – what it is and what it does – always brings up images from nature. And for good reason. That’s where the concept draws its inspiration. Think of swarms of starlings buzzing around in the sky, or fish swimming likewise in the water. The word "swarm" was inspired by the fact that various creatures, often for their own protection, exhibit a kind of decentralized behavior unrelated to the movements of their flock leader. And that's the basis for the idea of swarm intelligence.

Today, we define swarm learning as a decentralized machine learning solution designed to enable enterprises to harness the power of distributed data while protecting data privacy and security.

Swarm learning leverages the computing power at or near the distributed data sources. It ensures security using tested blockchain technology and protects privacy by sharing insights captured from deep learning models running at the source data, instead of the raw data itself. The data stays local. Only the learnings are shared. This results in improved models with less bias as they have access to a lager set of data – and solves data privacy, data ownership, and efficiency concerns.

With newly announced HPE Swarm Learning, the industry’s first privacy-preserving distributed machine learning solution,1 swarm learning benefits are now available to model developers. A permissioned blockchain is leveraged in HPE Swarm Learning to securely onboard members and dynamically elect the leader. This brings resilience and security to the swarm network.

Overcoming the challenges of a centralized approach to machine learning

As it tends to be done today, machine learning has many components. There’s data collection, usually at the edge. Then there’s data aggregation, done at the core data center where the AI model training is done. The model is then moved back to the edge for inferencing to generate predictions. This generates a continuous cycle, with a tremendous amount of resources needed to move data back and forth between the edge and the data center. This approach signals some significant challenges.

#1 Data is not complete due to data privacy issues and regulations that prevent sharing. Data ownership by different organizations with differing data policies leads to limited data sharing. In additional, demographic differences build bias into training models.

#2 Bias is present in local data due to demographic difference. This means models trained at local nodes will have bias built in.

#3 The centralized model training design requires data to be moved to a central location – with the learnings then distributed to the nodes for implementation. This results in mass data movement and data duplication. More issues arise in networks where connectivity is poor or intermittent data movement is limited and not synchronized.

You’re left dealing with low efficiency as you deploy models into production. You end up straining network, storage, and compute resources. In addition, regulations and data privacy regulations prevent data movement and drive a lack of data privacy. The resulting models can be suboptimal – with delayed time to insights, lower overall accuracy, and biased models.

A close look at two industry examples – healthcare and banking – shows more specifically how swarm learning solves these challenges.

Disease detection in healthcare

Modern healthcare organizations require improved accuracy for diagnosis and decisions from their AI solutions. And to achieve accuracy, AI requires access to larger data sets to reduce local data biases. Without data aggregation to one location, results from suboptimal models lead to inaccurate decisions.

Aggregating data in healthcare can be particularly challenging because data privacy regulations such as HIPAA and GDPR inhibit sharing data. Data ownership prevents sharing between hospitals and across geographies. Even when data can be aggregated, inefficiencies occur due to the cost of moving very large diagnostic image data utilizing precious bandwidth, and the fact that data may have to be duplicated utilizing precious storage resources.

With HPE Swarm Learning, healthcare organizations can perform AI model training with large sets of distributed data at the data source – with no movement of data. Collaboration between data sources can happen without compromising privacy. Raw data is not shared. Data ownership concerns are eliminated, as each data owner participates as an equal partner. The improved models provide more accurate disease classification with reduced bias plus overall improved time-to-accuracy.

In this specific use case focused on examining lung diseases for multiple patients, the patient data was spread across three geographies. Models at one hospital in each location failed to detect infrequently observed diseases. The infrequently observed lung diseases mean that the data had local bias. The swarm learning model was able detect these diseases where the hospitals had limited data and remove local data bias for the category. Even with sufficiently available data, the swarm learning model is either better or at par with any individual model.

More specifically, lung x-ray images were pre-labeled for four diseases. The models were trained locally at each of the three hospitals. Each location with fewer images for the particular disease had a lower accuracy of ~10%. Even where sufficient images were available, the accuracy was ~60%. Swarm learning improves model accuracy and subsequent patient diagnosis and achieved an accuracy of ~70%. This is significantly better in cases where there were fewer images and close or at par where sufficient data was available.

Credit card fraud detection in banking

The intent of a recent demo use case was to develop a machine learning model to detect frauds in credit card transactions. A publicly available dataset was used to build the model in a simulated environment with three credit card networks. Fraudulent transactions were split across the different networks.

As it was, individual credit networks couldn’t see all the fraudulent transactions, and transaction data could not be shared across credit networks. As a result, individual credit networks could not identify all the fraudulent transactions.

By using collaborative learning, swarm learning was able to detect all fraudulent transactions. Significant performance was gain as over individual models with limited data and biases.

These healthcare and banking use cases demonstrates just two of the many ways in multiple industries that swarm learning provides a powerful approach to AI – one that blends local and global insights while preserving data privacy and data ownership.

HPE brings new breakthrough AI solutions for speeding data-first modernization from edge to cloud, enabling scaling up AI to industrial-sized global applications. We make AI that is data-driven, production-oriented, and cloud-enabled – available anytime, anywhere and at any scale. Our solutions support today’s enterprises as well as financial services, health and life sciences, and manufacturing. HPE Swarm Learning brings your business a decentralized, privacy-preserving framework for performing machine learning model training at the data source

1An analysis as of April 13, 2022, of competing offerings that claim privacy preservation found that they use a federated architecture reliant on a central server.