Alibaba Mama Open‑sources Curvature Learning Framework and Federated Learning Solution

On September 15, Alibaba Mama announced the open‑source release of two AI technologies: the Curvature Space Learning Framework and a Federated Learning Solution. The open‑source launch aims to enhance data‑privacy protection, reduce storage consumption by about 80%, and improve user‑request matching precision by roughly 15%, with potential applications beyond the internet industry.

Curvature Space Learning Framework

Curvature measures how much a space bends; the closer the curvature to zero, the flatter the space. Inspired by the sci‑fi novel "The Three‑Body Problem," where curvature is used to build a curvature‑based spaceship, the framework leverages curved spaces to model graph data more accurately than traditional Euclidean space.

The framework provides a complete deep‑learning pipeline—including manifolds, operators, models, and Riemannian optimizers—allowing users to migrate models into curvature space. Experiments on the Cora graph dataset show an ~8% increase in prediction accuracy when replacing Euclidean space with curvature space.

In Alibaba Mama’s own business, the technology has been applied to Taobao search advertising. By modeling interactions among billions of merchants and users in curvature space, the system achieves directional data amplification and precise segmentation. After full deployment, storage consumption dropped by 80% and matching precision rose by 15%.

The approach is expected to benefit other domains such as cloud movement tracking, navigation, logistics, and resource flow mapping, leading to better weather forecasts, navigation, logistics efficiency, and equitable resource distribution.

GitHub repository: https://github.com/alibaba/Curvature-Learning-Framework

Federated Learning Solution

Federated learning, introduced by Google in 2016, enables machine learning across multiple devices while preserving privacy. Alibaba Mama’s open‑source solution emphasizes privacy protection and encrypted computation, establishing APP‑island information links to build models that support high concurrency, encryption, usability, and productization in large‑scale sparse scenarios.

Key features include:

1. Massive‑scale high availability : Cloud‑native implementation supports data intersections at the hundred‑billion scale, multiple verification methods, streamlined training protocols, and robust fault‑tolerance.

2. Encrypted privacy protection : Combines data‑security and computation‑security techniques, offering various privacy‑preserving schemes for optimal security‑performance balance.

3. Enhanced convenience : First open‑source release of both horizontal and hierarchical aggregation models, with a visual web interface for task development, pairing, scheduling, and management, greatly improving iteration efficiency.

The solution powers Alibaba Mama’s Unidesk product, helping brands such as Proya, KAZILAN, VINO, Huaxizi, and Xiuzheng achieve notable business gains. For example, Huaxizi saw a 15% increase in brand ROI within two months after adopting Unidesk.

Future extensions target finance, healthcare, and other sectors, demonstrating high universality.

Commitment to Open‑source

Since 2015, Alibaba Mama’s technical team has open‑sourced large‑scale deep learning, graph learning, and reinforcement learning technologies, including the industrial‑grade training engine XDL (2018) and the large‑scale graph deep‑learning framework Euler (2019, Euler 2.0 in 2021). The CTO emphasizes that AI is a new productive force, and sharing these capabilities maximizes collective benefits.

Follow the Alibaba Mama Technology account for more updates.