This site uses cookies, so that our service can work better. I understand

Charl Linssen

Session:
AoN Brainhack 2017 - Sattelite Event


Institute:
The Tiesinga group, Department of Neuroinformatics, Donders Institute for Brain, Cognition and Behaviour, Nijmegen, the Netherlands

Website:
http://www.ru.nl/english/people/donders/linssen-c/
Biography: 
Charl has a background in engineering, and is particularly interested in systems that actively respond to their environment. After studying Embedded Systems at TU Eindhoven, he realised that brains, or even something as comparatively simple as the nervous system of an insect, are ultimate embedded systems. Following this, he went on to pursue a Master’s program in Cognitive Neuroscience at Radboud University, where he is currently pursuing a PhD on the topic of active sensing in the rodent whisker system.
 

Abstract:

Convergence between artificial intelligence and simulation of the brain: from theory to GitHub

 

Artificial neural networks have received a recent spur in attention after notable successes in diverse areas: computer vision, motor control, natural language processing, as agents in computer and board games, and many others. Computer simulation of neural networks has been around since the late 1950s, but recent successes rely both on better knowledge of how to design and train these networks, as well as increases in scale made possible by increased computational power and the availability of large training datasets (“big data”).

For a neuroscientist, neural network simulation can be of interest from two different points of view: in the theoretical sense, as a model of how real brains work, or in the empirical sense, as a tool that analyses a dataset or performs a certain task (as in the game playing agent). These objectives may not be mutually exclusive, but depending on the application or goal, a researcher has to make concrete decisions about what model to use to approach it.

In this talk, we will review modern neural network architectures and consider them from the perspective of both theory and application. For example, if a certain network model requires a certain training paradigm, this could, on the application side, inform the allocation of CPU time, while on the theoretical side, lead to empirical predictions on neuronal biophysics involved in plasticity. Taken together, the goal is to give the audience (that’s you!) the knowledge needed to critically assess neural network models, and subsequently to download and run the chosen network on your own dataset and with your own selection of parameters.