How do you teach machines to see?

Autonomous vehicles recognise stop signs and endoscopes can detect cancer. An expert on artificial intelligence explains how this works.

VIDETE: DFKI project for machine vision
VIDETE: DFKI project for machine vision DFKI

VIDETE! – See! That is the name of a project at the German Research Center for Artificial Intelligence (DFKI). Here AI experts from the Augmented Reality Research Group in Kaiserslautern are developing new machine vision techniques. The goal is to enable machines to reliably recognise moving objects using only one camera. Until now, systems of this kind have required at least two cameras.

How can machines learn to see?

Project leader Gerd Reis explains: “How humans see and how machines see have very little in common.” “Learning to see” by machines is a mathematical process. What it involves is teaching robots how to understand scenes with the aid of neural networks – basically, a form of software. For example, in traffic situations they have to be able to recognise cars and their movements and correctly interpret traffic signs.

There are an extremely large number of variables in machine learning with neural networks. The researchers compare outcomes with the expected results and analyse the level of discrepancy. Then they take the system back to the last point at which there was agreement. That is how the system learns the connection between input and correct output.

What fields of application are there for machine vision?

Above all, in autonomous driving and in industry. But applications also include the detection of cancer using medical imaging techniques, the evaluation of competition scenes at sporting events and assistance systems for older people (Ambient Assisted Living). Reis explains: “It is important here to be able to differentiate between a person who has collapsed in an armchair unconscious and needs immediate help or someone who is just relaxing and reading a book.”

What are the challenges confronting AI researchers?

VIDETE is also conducting research on the justification for a machine’s decisions. How can you ensure that the artificial intelligence system always recognises traffic signs correctly even when they are covered in dirt? Or what enables an endoscopic device to identify tissue changes as cancer? Another algorithm checks the decision-making process before humans use the results. “It’s like a medical second opinion,” says Reis.


Newsletter #UpdateGermany: You would like to receive regular information about Germany? Subscribe here to: