Robots and burial mounds
Neural networks have a wide range of applications. In Leiden, psychologists use them to build robot brains, whereas archaeologists use them to hunt for prehistoric graves.
Building an artificial brain: ‘you learn a lot about the human brain along the way’
‘We will not succeed anytime soon,’ says Roy de Kleijn, Assistant Professor in Cognitive Psychology, ‘but at some point we will be able to simulate the human brain in a computer. There are no fundamental limitations.’
This idea is the cornerstone of De Kleijn’s research, which involves giving robots ever more advanced brains. Such a brain does not physically reside in the robot; it consists of a neural network simulated on a computer, which is in touch with the robot via wifi.
Even a tiny neural network of only 20 neurons (‘brain cells’) can learn to control a robot, for instance to make it ride towards a red light in a room. Such a robot has about the intelligence of a bacterium.
But is human behaviour, which is much more complex, simply the result of more neurons? De Kleijn: ‘If you add a million neurons to such a network, nothing interesting happens. That can’t be the secret of the human brain.’
Complex behaviour only arises if the neural network has a more complicated architecture, in which the signals emitted by the neurons are not just one-way. So not only do signals control the wheels, but some neurons also feed their signals back into the network.
If neural networks are made to evolve in a computer – by making near-identical copies and subjecting them to goal-driven selection – they even start to specialise spontaneously. Like in a human brain, part of the network ‘looks’ while another part drives the robot.
One day, De Kleijne hopes to have such an advanced model of the human brain that it can even simulate an affliction like depression. One could then carry out ‘medical’ experiments that would be unethical when performed on humans. Perhaps this is overambitious, says De Kleijne. ‘But you learn a lot about the human brain along the way.’
Most of our archaeological heritage is still undiscovered
Neural networks can be used in other fields as well. Archaeology has traditionally been limited by data scarcity: only a tiny fraction of our heritage has been preserved in the ground. But now, a deluge of data is coming, mainly because of new ways of exploring.
In Leiden, Karsten Lambers, Associate Professor of Archaeological Computer Sciences, uses Lidar data for this. The whole of the Netherlands has been laser scanned by planes, at a vertical resolution of a few centimeters. This was primarily done to detect surface subsidence and defects in dikes, but these Lidar maps are accessible for free.
In Lamber’s research group, a neural network searches these maps for burial mounds and other archaeological remains that can be spotted from small variations in surface height. As always, this network has to be trained with examples. However, only few Lidar maps are available that show proven archaeological finds. Lambers: ‘But we do have large databases with pictures taken on the ground, with animals, people, cars and so on. A neural network that is trained on these images can also be used to recognise archaeological objects in Lidar images. This is called transfer learning, and this technique is very important to us.’
Lambers believes that only the minority of archaeological objects have been discovered: ‘We expect at least to double the number of objects found in the area that we are going to study thanks to this method.’