Managing Edge Deployment of Large Deep Learning Models in industry

MEDLI (Managing Edge Deployment of Large Deep Learning Models in Industry) focuses on the challenge of running powerful deep learning models at the edge of networks, close to the data source. This enables real-time decisions without dependence on the cloud.

The problem: modern AI models are becoming increasingly large and complex, making them difficult to deploy on edge devices with limited computing power and memory. In addition, companies struggle with:

• compressing and adapting pre-trained models (transfer learning)
• choosing suitable edge hardware and software
• monitoring and maintaining AI models after they have been rolled out

MEDLI focuses on industrial end users, technology and software suppliers, and system integrators who want to deploy AI directly on the shop floor or in production environments.

MEDLI develops practical tools and guidelines to make it easier for companies to apply deep learning at the edge. The four key objectives are:

• Faster development of edge AI models
  Use pre-trained models, transfer learning, and techniques such as pruning and quantisation to reduce the size of models while maintaining accuracy.

• Easy selection of hardware and software
  Overviews, decision trees, and guidelines for selecting the right combination of edge hardware (GPU, TPU, CPU, etc.) and deployment tools, tailored to application requirements.

• Monitoring of AI models in operation
  Methods and tools to monitor model health (accuracy, errors, drift), with automatic signals for adjustment or retraining.

• Generic demonstrations and use cases
  Practical example projects — such as visual inspection or time series analysis (vibration, sensors, audio) — including documentation and step-by-step plans as a blueprint for implementation.