The Carl Zeiss Foundation is funding the establishment of a new artificial intelligence (AI) research centre in Mainz, Germany, that will focus on determining why modern machine learning methods are so effective.
The new Emergent AI Center is being established at the Institute of Computer Science at Johannes Gutenberg University Mainz (JGU), and will receive €3m over the next three years from the optoelectronics company.
AI research centre takes on machine learning “mystery”
A core focus on the AI research centre will be unlocking greater understanding about machine learning, one of the most widely used types of AI that in many fields now matches human-level performance.
“Modern machine learning methods allow us to solve many problems using computers which just a few years ago, without doubt, only humans were able to handle successfully,” said Professor Michael Wand of the Institute of Computer Science at JGU.
“Of course we know exactly how the systems work, but why they work so well remains a mystery.”
Machine learning advances have occurred due to the development of what are known as deep artificial neural networks. These are able to perform basic pattern recognition, and with training over time can be taught to recognise increasingly complex patterns, creating a ‘deeper’ network.
While this approach is now very effective – providing the training data used is suitably robust – the nature of why this pattern recognition works remains unclear.
Essentially AI researchers know that machine learning involves pattern recognition, which is made more accurate the more a system is trained. This means that the data has to not only be statistically similar, but that the learning process the system goes through has to include knowledge of this similarity.
But recognising what these patterns are and how the neural networks used in machine learning identify and use them remains unclear – and it is this that the new AI research centre hopes to tackle.
“Which patterns natural data have in common and how neural networks exploit this prior knowledge is still an open scientific problem,” said Wand.