Artificial Intelligence for Small Devices

In the HALF project, we are developing energy-efficient hardware that enables artificial intelligence to evaluate patient data on mobile devices.

Energy-efficient Hardware for Artificial Intelligence

The Goal: Artificial Intelligence Fits on the Smallest Devices

Especially in the medical field there are promising applications of artificial intelligence (AI). For example, AI systems can support doctors in making diagnoses and recommend more suitable medication for the patient. In addition, portable devices that can record and monitor a person's state of health have been available for some time. What is new is the trend towards carrying out the necessary, complicated partial diagnoses directly on the terminal device in order to detect a critical state of health as quickly as possible.

A popular example of such a device is the Applewatch, which is capable of recording an ECG of the wearer. In this way, the device can detect an approaching heart attack and sound the alarm early. This technology has the potential to save lives.

The algorithms for evaluating patient data can be very computationally intensive, which results in high power consumption. However, the runtime and thus the reliability of a mobile system depends on its energy consumption. For mobile applications, therefore, the energy-efficient execution of the evaluation algorithms on the hardware has the highest priority.

The Competition: ECG Data is Automatically Analyzed with the Lowest Energy Consumption

Within the framework of a pilot innovation competition of the Federal Ministry of Education and Research (BMBF), a system for the reliable detection of atrial fibrillation is to be developed which is to be as energy-efficient as possible. The focus is on the hardware that will be developed specifically for this task. So-called Field Programmable Gate Arrays (FPGA) are prescribed as the hardware platform for this competition. FPGAs are special chips whose circuits can be reconfigured during operation and thus allow the fast realization of any circuit on concrete hardware.

The Fraunhofer ITWM was selected together with our partner, the Microelectronic Systems Design Research Group at the TU Kaiserslautern, to participate in this competition.

Our Approach: Energy-Efficient Hardware Starts with the Algorithm

For this competition, we choose a holistic approach that looks at the application and the hardware at the same time, thus achieving the optimal execution of an optimal algorithm. The project name HALF - Holistic AutoML for FPGAs - reflects the core aspect of our approach.

An important finding is that the AI model not only determines the quality of detection, but also very decisively the energy consumption on the hardware. Instead of focusing only on the hardware and relying on known AI models, we use state-of-the-art methods to automatically search for AI models. In doing so, we combine this search with hardware-specific criteria so that we can find optimal models for the underlying hardware platform.

The difficulty is to find a balance between a lightweight and energy-efficient model on the one hand and its reliability and robustness on the other.

The software tools we have developed enable an almost fully automated search for the best design in a multi-stage process. Starting with the ECG data set, so-called Auto-ML (Automated Machine Learning) methods search for the best model for the given data. During the search, both unimportant parts of the model are avoided and hardware-specific criteria that determine energy consumption are taken into account. An optimal design for the FPGA platform is generated for the model thus found. Finally, the FPGA is configured and is then capable of automatically evaluating ECG data.

Our Project Partner:

The Microelectronic Systems Design Research Group at the TU Kaiserslautern provides the know-how to map the AI models to the hardware. In particular, the hardware architecture generation methods developed by the EMS Group are used to exploit the full potential of the FPGA hardware platform.