Training and understanding artificial neural networks with cognitive neuroscience inspired methods

Click here to watch the video

Deep Learning models are very successful in various fields of application. Performance gains are often achieved by an increase in model complexity in terms of types of architectures or the number of neurons.
At the same time, these models become harder to interpret, which can be risky in critical applications.
Real brains are even more complex systems which have been studied in neuroscience for decades.
This wealth of experience can help us to better understand how deep neural networks solve their tasks.
We research how well-established methods from cognitive neuroscience can be adapted to analyze and train artificial neural networks. In the talk, I will present model analysis techniques, which are inspired by the Event-Related Potentials (ERP) technique and demonstrate them using an exemplary convolutional Automatic Speech Recognition model.

Andreas Krug is PhD student on explainable AI at the Artificial Intelligence Lab at Otto-von-Guericke-University Magdeburg. He has a background in natural sciences and holds a Master's degree in Bioinformatics. After his studies he specialized in Deep Learning during 3 years as researcher and lecturer at the Machine Learning in Cognitive Science Lab at University of Potsdam. Andreas researches techniques to analyze and visualize, how deep neural networks perform their tasks, partly inspired by methods from cognitive neuroscience. His work is driven by the idea to make AI better explainable to also enable non-experts to get insights into inner workings of these systems. To this end, he regularly communicates AI to the general public, for instance through exhibits, science quizzes and public talks or discussions.