Microcontrollers, edge TPUs, and specialized constrained hardware require ML tuning before they can be deployed. We provide state-of-the-art mechanisms for pruning and quantizing models, especially neural networks, but also other architectures for use on embedded devices.
SERVICE DESCRIPTION
Moving computations to the "edge" may require additional steps to use models efficiently. These steps include model pruning to accommodate limited memory resources and real-time requirements, and quantization or hyperparameter optimization to account for different architectures.
Our algorithms are able to handle different types of embedded systems, from minimal MCUs (Arduino, esp32, RISC-V) to specialized edge TPUs such as NVIDIA Jetson systems. We have experience in developing embedded machine learning applications, especially for industrial sensing applications, environmental sensing, and wearable and mobile computing. We can provide expertise for the most novel hardware designs employing different types of processing such as neural networks on printed electronics.
With the help of our computing resources, we will match customized machine learning to fit your hardware needs based on the concrete data to be processed. Applications may range from mobility to wearable computing to industrial machinery.
The tasks are performed by experienced ML researchers at KIT. KIT is "The Research University in the Helmholtz Association". As one of the largest scientific institutions in Europe, Germany's only university of excellence with national large-scale research facilities combines a long university tradition with program-oriented cutting-edge research. Since KIT also has a focus on innovation and technology transfer, our experts have many years of experience from applied industrial projects.
SPECIAL ACCESS CONDITIONS
Conditions and requirements for participation in an experiment within the Open Calls:
By participating in an EUHubs4Data Open Call, you are initially only applying for funding that originally comes from the European Commission and is awarded by the coordinator exclusively in its own name under the conclusion of a sub-grant agreement. This sub-grant agreement does not establish a contract with KIT, neither through your application nor through a possible positive funding decision.
KIT will therefore - also in your own interest - conclude a separate written agreement with you at the start of the experiment (based on our sample cooperation agreement. If you decide to propose the participation of KIT and SDIL infrastructure in your experiment, you must respect the following conditions. We provide this information in advance to ensure maximum transparency: please contact us if you have any questions. In the unlikely event that you are unable to conduct your experiment with our participation, we will attempt to assist you in selecting alternative services before the experiment begins.
Please note that contrary to the name "service", the above description is not a genuine commercial offer, but a listing of exclusive contributions as part of a genuine eye-to-eye collaboration.
For genuine commercial offerings related to the above topics, please feel free to contact us any time outside of the Open Calls.
PREREQUISITES
Existing ML model with training data (e.g. from a previous task within an experiment).
CASE EXAMPLES
You may come to us if you have identified a fitting target hardware architecture for your product and want to fit your exisiting running machine learning model to that device.
SUCCESS STORY
In this example we fitted an advanced image recognition task into a low cost hardware together with an innovative local startup for traffic monitoring: https://www.sdil.de/en/projects/e-scooter-detection-2
SERVICE CAN BE COMBINED WITH
This service can be used together with any PaaS or SaaS service that can run tensorflow efficiently. If no other infrastructure is selected, KIT can provide fitting infrastructure.