Fraunhofer Institute for Integrated Circuits IIS
Fraunhofer Institute for Integrated Circuits IIS

TinyML

© Fraunhofer IIS / Gauthier
TinyML Demonstration

Tiny Machine Learning (TinyML) is a research area in machine learning and describes the optimization as well as execution of AI-based processing chains on embedded systems.  

Our embeddif.[ai] technology is an established brand for TinyML. Here you can find several use cases for embeddif.[ai].

For us, artificial intelligence (AI) is not only the use of deep neural networks with millions of parameters, which require a not inconsiderable amount of energy already during their development (training), but also during operation (inference, scoring), but we also include classical machine learning with very efficient processing chains in the solution space.

At the ADA Lovelace Center, we develop prediction models for the energy requirements of AI on embedded hardware platforms to optimize the trade-off between performance and energy requirements of battery-powered sensor systems in very fast development cycles. Only the fast prediction of the energy demand allows the integration into a multi-criteria AutoML (automatic machine learning) solution.

In industry projects, we provide our customers with automation solutions for AI development for ultra-low power applications or advise data science teams on optimized workflow and reliable testing in seminars and workshops.

TinyML Advantages

Energy efficiency & resource conservation

The TinyML needs up to a thousand times less power to run ML applications compared to a standard GPU. The reason for this is, among other things, the local execution of the models instead of sending the data back and forth. For this reason, the TinyML devices can run for years without batteries in some cases, depending on the use case.

The size of the batteries required and thus the use of valuable resources is also reduced due to the considerable energy savings.

Real-time

Since the model is run locally, in order to perform inference, the raw data does not need to be sent to the cloud first and then sent back when processed. This reduces output latency as well as communication bandwidth requirements, which in turn enables rapid response.

Independence, Privacy & Security

The user is not dependent on a cloud service provider. Since the data does not have to be shared with external parties, this point ultimately also contributes to the protection of privacy.

In addition, there is no dependence on a communication link. The risk of possible interference during transmission between the embedded system and the cloud is therefore eliminated.

Our offer

You want to run an AI based project and already know that the execution will take place on embedded hardware?   

You want to save energy and thus costs when executing your AI models?

You have an idea for a battery-powered sensor system for the fields of Industry 4.0, sports or service and maintenance and need support?

We offer partial or complete R&D services:

  • Consulting for successful data acquisition and planning
  • Provision or consulting on required sensor technology
  • Development and optimization of AI pipelines for your application
  • Support for integration on your hardware (e.g. delivery of C/C++ code)

 

Do you or your employees need hands-on training on how to run ML projects?

  • Take a look at our seminars
  • We would be happy to use your data for a customized seminar

 

You need a concrete solution for a complex application quickly?

  • Maybe we have already developed the right model or processing chain and it can be directly integrated into your use case or licensed by you
  • We optimize your application through standardized workflows and automation of time-consuming tasks to the greatest extent possible, specifically adapted to your use case
    • Through training and automatic reduction of complex models by removing redundancies we generate optimal models in terms of accuracy and efficiency       
  • We support you with a fast integration

More infos about embeddif[ai].

Applications

 

embeddif.[ai] Tools

Machine learning in embedded sensor modules for cognitive hand tools to detect assembly processes and ensure quality.

KI4Tools
 

embeddif.[ai] Sports

Machine learning in embedded systems, e.g. wearables for sports applications such as fitness, soccer, boxing or basketball.

Sportanwendungen

embeddif.[ai]
Condition Monitoring

Use machine learning in embedded systems to monitor states of plants and machines in order to be able to react at an early stage or to increase efficiency.