Making devices smarter with Artificial Intelligence (AI)

What is a “neural network” and how does it work?

An artificial neural network is an imitation of our brain (biological neural network), in both construction and operation. It is used, among other things, for photograph recognition. For example, we can make a system “smart” so that it can determine whether or not a given photo is of a dog. You develop a neural network by training it with a long series of examples. The underlying technology is a long, simple calculation that is fine-tuned on the basis of examples until it gives a correct result for all input values. Neural networks are really good at recognising patterns, and after training even new photos can be properly assessed.

Prediction vs actual behaviour

We can use neural networks to tackle problems differently. Take, for example, hob ventilation, in which the electronic controller is operated using capacitive touch sensors. These sensors are located under a thick glass plate, and in addition there is also moisture, dirt and interference from the induction hob itself, for example.

We investigated how well a neural network can be trained to recognise the right signal (e.g. switch on) among all this “noise”. To do this, we needed a large number of training examples. We created an “artificial finger”, which operates the sensor using compressed air, to automate the collection of data.

Binary variant possible?

Our electronic controllers often use micro-controllers that are not too adept at handling, for example, decimal numbers. For this reason, we evaluated the binary variant of neural networks for use in our electronic controllers. In this setup, the calculations are done using only one and zero (binary).

To visualise this, we implemented a demo of our own embedded AI framework. You draw a number from 0 to 9 on a touchscreen and in the background it is passed on to a binary neural network. This binary neural network works out the number you mean and shows the result.