Artificial intelligence on embedded devices

Insight Brief

Artificial Intelligence | AI – especially the subfield of Machine Learning (ML) or Artificial Intelligence (AI) – is conquering more and more areas of our everyday lives. The ability of these algorithms to learn independently from data sets opens up new horizons. AI algorithms are ahead of humans, especially when it comes to evaluating big data and searching for ‘hidden’ patterns.

Introduction

There is huge potential for AI algorithms in medical technology. Anonymized treatment data can be condensed into therapy recommendations. Device diagnostics can be improved and service life increased. Control and regulation tasks are also already being taken over in part by neural networks, decision trees, or the like.

However, medical technology also places high demands on AI and ML systems. Wrong decisions can have drastic consequences. The behavior of a neural network may be harder to predict than that of a human-developed algorithm. There is an urgent need for regulatory requirements to ensure the safety of AI systems in medical technology. IMT AG employees have therefore been working in the standards committee IEC PT63450 – “Artificial Intelligence-enabled Medical Devices – Methods for the Technical Verification and Validation” since the beginning of the year.

PF-300 Pro flowmeter for the calibration of medical ventilators and anesthesia machines with innovative AI algorithms

AI algorithms for computationally intensive applications

The processor power of embedded devices in medical devices is often orders of magnitude lower than on a PC for multimedia applications or in an industrial production plant. Typical processors in embedded systems are therefore often not powerful enough to run complex learning tasks on them in a meaningful way. Nevertheless, a large part of popular AI algorithms can be easily integrated into them if the training is done by a PC or server. This is because in most cases the training is far more computationally intensive than the subsequent application (inference).

This is true for artificial neural networks (KNN) as well as for support vector machines, decision trees, and many more. Bayesian networks are a counterexample: Here, inference is also very computationally intensive. This is due to the fact that for interference conditional probabilities have to be calculated by numerical integration over several variables.

IMT Analytics AG, a customer of IMT AG and also located in Buchs, Switzerland, manufactures flowmeters for the calibration of medical respiratory, and anesthesia equipment. In newer models such as the PF-300Pro, innovative AI algorithms are used, for example, to compensate for environmental influences on flow measurements or to reduce the size of large look-up tables.

Training and application splitting when using AI algorithms. The application can often easily take place on embedded systems, while the training often requires more computing power

For example, a look-up table with over 200,000 support points could be compressed to 31 coefficients using an artificial neural network. This achieves a memory reduction of 99.98 % with a maximum error rate of 1 %. This reduces both costs and computing time. Due to such an algorithm, IMT Analytics can assert itself as the market leader in flow measurement technology in the future.

The potential of AI in medical technology is enormous, but there are still many open questions: How can AI algorithms be interpreted, explained, and verified? How do we define test coverage in this context and how do we realize it? How do we ensure a representative data basis so that demographic, cultural, or even gender-specific differences are covered as well as possible?

Such questions urgently need to be clarified in the field of medical technology. Several normative bodies have already begun to address them. Initial recommendations are expected shortly.

Summary

Artificial neural networks on embedded systems

How can artificial neural networks be applied to embedded systems? These three ways are possible to bring Artificial Intelligence, Machine Learning, and Embedded Systems to a common denominator:

Manual Implementation:
Perceptron-type artificial neural networks (ANNs) are simpler than you might think from a mathematical perspective. Simple networks can be implemented with a few loops and standard mathematical operations. However, if one codes the ANN oneself, every change of the network architecture also entails an adaptation of the code. This is costly and error-prone.

TensorFlow Lite:
TensorFlow is one of the most widely used machine learning frameworks. If an ANN has been trained with TensorFlow, it can be deployed on an embedded system using TensorFlow Lite.

X-Cube-AI:
ST Microelectronics offers X-Cube-AI to its customers, an easy-to-use product that promises the same thing as Tensorflow Lite: to apply an ML system trained with Tensorflow/Keras on an embedded system with an STM32 processor. At IMT, the X-Cube AI was tested with the example network. It is very easy to use – easier than Tensor Flow Lite – and convincing with functionality. Disadvantage: Only processors of the STM32 series are supported.

Do you need assistance in the area of machine learning or artificial intelligence in embedded devices or your next project?

Contact our experts at IMT – we are here to assist you with your challenges.

Go back