Name of the participant: Martin Rapp
Description of the IT research project: The number of devices that interact (partially) autonomously with the real world is growing rapidly, not least due to the rapid growth of IoT. Machine learning is often used to control the interaction with the environment. Examples are process predictions in intelligent production, suggestions for user actions in smartphone apps or predictions in the field of autonomous driving. If several devices perform a similar task and interact with a similar environment, these devices can learn cooperatively and thus improve the quality of the learned models.
Sending training data to a central server (e.g. in the cloud) may not be practical for various reasons, such as high latency, or privacy. Therefore a trend towards distributed learning can be observed. Each device continues to train with its own experience. To benefit from the experience of other devices, the devices exchange the updated models.
In order to be able to realize devices in large numbers, each individual device must be produced at low cost and require little electrical power and energy. This means that these devices can be very resource limited in terms of available computing power, electrical energy, electrical power, memory or bandwidth. These limitations vary in type and characteristics between different devices and thus form a heterogeneous system. Highly resource constrained devices are not able to continuously train complex models locally, making distributed learning difficult.
The goal of this project is to develop techniques that enable cooperative machine learning on heterogeneous distributed systems consisting of resource constrained devices.
Software Campus partner: KIT, Huawei
Implementation period: January 2020 to April 2022
