12. BCN Cognitive Neural Networks
Target Group
Master’s students of the BCN Research Master, PhD candidates
Credits
2 EC for BCN PhD candidates for attending.
Content
This class will examine how to build systems based on recurrent neural networks (RNNs). We will study how biologically motivated RNNs (with excitatory and inhibitory neurons playing different roles) can self-organize to represent the structure of the data they are exposed to, allowing them to perform noise reduction, pattern completion, decision making, and inference. We will then look at how such RNN modules can be interconnected into larger networks that self-organize to represent latent variables that are useful for processing the data. Finally, we will look at how these ideas can be extended to form a distributed hierarchical action selection system that can select the optimal set of actions for the current situation.
Period
June 2027
Form
Variable hours per week. Lectures and self/individual study. Mandatory attendance: lectures.
For much of the material, there is no other source (besides the lectures) to learn it from.
Learning outcomes
At the end of the course, the student is able to:
-
Explain and demonstrate how the combination of competitive dynamics, activity homeostasis, and Hebbian learning results in self organization to represent the manifold from which random input samples are drawn.
-
Explain and demonstrate (in a high-level model) how latent variables can be created by a recurrent neural network.
-
Explain how a distributed hierarchical action selection system can select the optimal set of actions, and how its parameters can be learned.
Time Schedule
Tba
Evaluation
Written assessment (for research master’s students only)
Number of participants
Max. 16
Location
tba
Contact
Diana Koopmans (d.h.koopmans umcg.nl)