Artificial Intelligence (AI) gets boost from Growth Fund
The National Growth Fund announced last week that it will make a quarter of its €4 billion budget available to strengthen research and innovation in the Netherlands. Part of that money will be spent on the key technology Artificial Intelligence (AI).
But how intelligent is AI really? And can this new technology already be applied 'just like that'? My answer to this question is a resounding no. AI is far from being 'finished'; we are still missing a number of essential ingredients. Do not focus only on research at universities. Higher vocational education (hbo) can also make an important contribution, together with industry and partners.
The cork on which AI runs today is computing power and data, lots of data. Self-learning algorithms such as neural networks are fed large data sets and are able to discover patterns in them. Based on this technology, we can program data-driven decisions. In fact, these algorithms simply calculate correlations, the relationship between X and Y. But an AI that really understands things, that is, establishes cause-and-effect relationships and develops a mental model of how the world works, is still a long way off.
For example, today's data-driven translation algorithms do their job without understanding the content of the text. Take the example of the self-driving car. After thousands of hours of training with extremely large amounts of data, a neural network has learned to recognise objects on the road and instruct the car to avoid them or stop for them. But a self-driving car doesn't think: Hey, I see a ball rolling onto the road, let me slow down because there could be playing children behind it. We humans can do that, with only 40 hours of driving lessons.
Another element that is still missing is ethics. In practice, it is unclear how and whether we can link AI to human values. Reports regularly appear in the media that AI discriminates or wrongly marks someone as a suspicious person. Often this has to do with patterns already hidden in the data that the algorithm is trained with. For example, a year ago Amazon came into the news negatively because of their woman-unfriendly recruitment and selection algorithm. It turned out that this algorithm mainly recruited men for vacancies for software developers. A look into the data showed that these "biases" were also present in Amazon's hiring policy.
The final argument for why AI is not yet mature is of a very practical nature. AI is not yet a true engineering discipline. The steps to create an AI application have a high 'trial-and-error' content and are often not linked to a thorough development process.
An experimental nature suits AI. However, building an impressive 'demonstrator' or a research prototype is not the same as creating a smart product or service for a large group of end users. Software professionals and managers of ICT companies today do not have sufficient tools to deal with this properly. How do you organize the collection and storage of data? When are predictive models out of date and do you need to retrain them?
I would therefore argue that not all AI funds from the National Growth Fund should be used for scientific research projects. Let universities work on the next generation of AI models with causality and other fundamental concepts. But we also need to invest right now in better embedding AI in society. With respect to both AI engineering and ethics, there is much to be gained. Field labs and the hybrid learning environments, such as the Fontys ICT Innovation Lab at Strijp T, in which the institutes for higher professional education work closely with the business community and government organisations, are extremely suitable for experimenting with this and for investigating what does and does not work in practice.
Gerard Schouten is a lecturer in AI & Big Data at Fontys University of Applied Sciences in Eindhoven.This opinion piece was published in the Eindhovens Dagblad.