Pelt brings theory and practice together in field of Deep Learning

Bringing the theoretical world of mathematics and computer science to more applied research areas. That is what Daniël Pelt, associate professor at the Leiden Institute of Advanced Computer Science (LIACS) is trying to achieve with the methods he is developing as a solution to challenges in image processing.

Failing methods

During his PhD research at the Centrum Wiskunde & Informatica (CWI) in Amsterdam, he focused on tomography: making a three-dimensional image of the internal structure of an object. This technique is used in the medical world, for example in CT scans. In his research, Pelt found that the methods used in tomography often work less well in practice than in theory. ‘Methods that at the time were widely used, were often not accurate enough for advanced applications,’ he says. ‘But more accurate methods were often much too slow to use in practice. In this way, I was triggered to find methods that were fast and accurate enough for practical applications.’

He discovered the factors that cause existing Deep Learning methods to often be difficult to use in science after his PhD, during a year that he spent in America working at the Lawrence Berkeley National Laboratory. Pelt: ‘Often the scientific images are too large for fast processing, there are too few examples to train existing models with, and the settings of the model do not work for every problem, so scientists themselves often spend a lot of time finding the right settings.’

Mixed-Scale Dense Network

He therefore developed the Mixed-Scale Dense Network, a Deep Learning method for scientific images. ‘This method requires choosing far less settings, making it easier to apply,’ explains Pelt. ‘Because all parts of the network are connected, the network learns during its training what is needed to solve a specific problem. As a result, it automatically adapts to the problem, and the scientist does not have to do this himself. The network can be applied to many different types of images: from images of cells to pictures taken by a camera on a car, as well as, for example, CT images of jaws.’

3D images

Currently, Pelt is working on a research project for which he received a Veni. Continuing with the ideas behind the Mixed-Scale Dense Network, he is now developing a method to analyze 3D images. ‘The Mixed-Scale Dense Network is built on 2D images, but many interesting scientific images are in 3D,’ says Pelt. ‘That’s a big challenge, though, because there’s even less training data available for 3D images, and the images are even larger. It is therefore essential that the system works super-efficiently. In addition, we do want the system to continue to adapt to the problem, so that it is easy to use.’

In the future, Pelt hopes to continue along this line and further close the gap between theory and practice. One of the things he wants to work on is the user-friendliness of methods and systems. Pelt: ‘Now, for example, problems often arise if there is an error somewhere in the data, while in practice there are often errors in the data. It is therefore important to develop a more robust method, so that you can also practically continue to use that method.’

/Public Release. This material comes from the originating organization and may be of a point-in-time nature, edited for clarity, style and length. View in full here.