Thema1. Alexandra Brand.jpg

AI tools

Artificial intelligence leads – besides discussion – to practical solutions. Five scientists explain which concrete applications they are developing on the basis of AI and which challenges that brings.

Thema5. Alexandra Brand..jpg
Text: Martijn van Calmthout, image: Alexandra Brand

Big Data researcher Jacco van Ossenbruggen, Centrum Wiskunde & Informatica

Doing justice to cultural diversity and historical archives

‘At the National Library of the Netherlands in The Hague, they have digitalised more than 400 years of newspapers. That’s about 150 million individual articles that are publicly accessible via It is impossible to search through those manually, and so Delpher uses a search engine. You enter keywords, and the articles with the most hits come in the top 10 results. Artificial intelligence can be used to accelerate and refine such searches. But, of course, the questions remains whether such a search engine can really find the relevant articles? That is because a historical newspaper archive is also vague and disorderly: digital images can be of poor quality, newspapers can be damaged, there are stencilled wartime newspapers, and pamphlets, and on top of that metadata like the date and title can sometimes be doubtful. Some articles are short, others are long and there are photos without text. A certain incompleteness always creeps into the search results, and you do not want sources to be missed. One of the challenges we face, is developing AI systems that do justice to the cultural diversity and, for example, can find minority viewpoints and compare those with more mainstream opinions.’

AI researcher Guido de Croon, TU Delft

Drones that decide themselves

‘My research focuses on small drones: flying mini robots with the dimensions of a large insect. Their low weight alone makes them safer for a wide range of applications. One such example is monitoring crops in a greenhouse. What you need is an autonomous system that navigates and inspects. A swarm that sets to work itself and notifies you of problems. And that can also be used, for example, to find victims following a disaster.

Traditional robot systems build a map of the environment first and then determine the routes. That takes calculating power and memory, and there is no room for that in small drones. However, the animal kingdom reveals that you can navigate really well without the need for maps. With just these few rules, you can already achieve much: you have to avoid ad hoc obstacles and each other, and you should measure the distance to your starting point. In our case, that is the signal from a Wi-Fi router. We have demonstrated this with a swarm of six drones, weighing just 30 grams, which learned to discover a room by itself. Our dream is that we will no longer need to actively establish the rules, but that the swarm will discover the best rules itself. And those could be very different rules from the ones that people come up with. A form of evolution.’

PhD particle physicist Bob Stienen, Radboud Universiteit Nijmegen

Relationships that no physicist can see

‘People – and so also physicists – can see patterns in two or three dimensions quite easily. That’s handy if you need to spot a tiger in the forest. However, that doesn’t help to detect the relationships between the many parameters that play a role in the case of colliding particles. That’s different for a computer: how many dimensions a dataset has does not matter that much because it is mainly a question of processing power. I’m investigating how we can make use of this in particle physics. The particle physicist hunts for phenomena or particles that do not occur in the Standard Model. We are searching for deviations within enormous quantities of measurements using the detectors at the Large Hadron Collider of CERN. That can be done in two ways. By investigating a theoretical prediction which is comparable to looking for a dog on a crowded photo: find the nose and the tail and then the rest of the dog. With enough examples, a computer can learn to do something like that. The other method is to use a lot of data to teach a computer what is normal and then to compare the real data with that. These techniques are far faster than traditional analyses. The calculations can be performed in many seconds instead of days or weeks. However, computers cannot replace scientists. They are merely tools. Ultimately, only we can produce physical explanations.’

Machine Learning computer scientist Sander Bohté, Centrum Wiskunde & Informatica

Efficient local network like in our brain

‘The brain is an inspiration for neural networks: metaphorically, as a network of neurons that pass on signals to each other, and literally, because the brain is so efficient. It is about one million times more efficient than what we can currently achieve using neural network chips. That is because brain neurons only pass on a signal to each other once per second in the form of a zero or a one that are simple to add up. It is therefore a so-called spiking neural network. In traditional artificial neural networks, 16- or 32-bit numbers proceed step by step through the network and multiplications are continuously needed. That complex communication process requires a lot of energy, especially because the entire network is always on. When you use a system like Google Assistant, everything goes to the cloud and back again. A simple question easily requires 100,000 times as much energy than a local processing on the device would require. However, a traditional neural network would quickly empty the battery of your mobile phone or smartwatch. Our work here is aimed at algorithms that make optimum use of the spiking neural networks. An advantage of this is that not all of your data are continually sent to the cloud, which is better from a privacy viewpoint.’

Financial economist Albert J. Menkveld, VU Amsterdam

Blessing and curse of high-frequency trading

‘As a scientist, I do not really have a clear view of how the securities market makes use of artificial intelligence, but we all know that this happens. High-frequency trading in shares depends on recognising patterns and responding to these. I investigate what that fast technology triggers and whether that is efficient. The need for your technology to be always faster than that of your competitor is expensive. What is the benefit of being a microsecond ahead of other market parties? Conversely, it is interesting to see how superfast responses can help or disrupt the market. High-frequency traders provide liquidity for the smaller short-running orders but seem to predate upon large long-running orders. Economists in Chicago have already proposed exchanging the continuous trade for high-frequency auctions, for example ten per second. This would bring the expensive arms race in microseconds to an end.

Artificial intelligence is mainly good in recognising patterns in a lot of data. Is a current movement in the market similar to what we have previously seen and what happened back then? That helps give shape to the investments you make now. However, these remain stochastic processes. As they say, past performances are never a guarantee for the future.’