Meet-up: Practical applications of deep learning at WUR

Published on
June 25, 2019

Deep Learning is a technique from artificial intelligence. It is a class of machine learning algorithms that use multiple layers to extract higher level features from raw input. Deep learning is very well applicable to the research domain of WUR.

During this second meetup on the 12th of June, we learned and discussed about current applications of deep learning. We invited different guest speakers to illustrate questions and possibilities related to deep learning in their research domains. What are the new possibilities offered by deep learning? What challenges still lay ahead? How does it work?

Willem Jan Knibbe, head of the Wageningen Data Competence Center (WDCC), opened the meeting and introduced the audience in the world of deep learning. Deep learning works by assembling an end-to-end differentiable model from base components that can be fed data and labels. By iteratively refining the components the model can then be taught to find relationships between the data and labels we supply. The deep neural network works with multiple layers between the input and the output. The network moves through the layers calculating the probability of each output. Deep learning can be applied in domains such as image recognition, speech recognition, machine translation, bioinformatics, composition of medicines, board game programs, etc. They can sometimes achieve results comparable to or sometimes better than human specialists.

Four guest speakers of the WUR, talked about their experiences with deep learning in the various domains:

Erwin Mollenhorst, Changes with deep learning in animal sciences (Lifestock Research)

Ronald de Jongh, Extracting Biology from deep DNA classification (Bioinformatics)

Manya Afonso, Deep learning for agricultural machine vision (Biometris)

Devis Tuia, Deep learning in Geoinformation science: beyond classifying pixels (Geoinformatics)

After the presentations there was time for networking and sharing experiences.

Points for attention: A lot is happening in the field of deep learning within WUR. A new field offered by deep learning is for example the investigating DNA with Convolutional Neural Networks. A lot of development is happening here and where the surface of what is possible has barely been scratched. Ronald de Jongh foresees multi-modal networks like the one presented by professor Tuia coming to biology as well, combining very different streams of data about the same biological sample into a coherent prediction.

For deep learning on biological data, and probably also for the other domains, one of the biggest challenges is still data interpretation; how do we know for sure the model has truly learned the underlying biology and hasn't just made spurious connections that will vanish when more data comes available?

This meeting
was organized by Campus Connect and Wageningen Data Competence Center (WDCC) of
Wageningen University and Research.