Reservoir Computing (RC) was introduced as a methodology to exploit random recurrent neural networks (RNNs) for computation. Key to RC is a linear readout layer which is trained on the activity of the RNN to produce some desired function. The RNN is referred to as the "reservoir" and remains untrained.
Interestingly, the reservoir doesn't need to be a neural network at all. In fact, any system can be used as a reservoir. This insight has spurred a wide range of physical reservoirs, ranging from optical systems, biological neurons, bacteria and even a bucket of water.
An open question with physical reservoirs is how physical limiations affect computational performance. How does the accuracy of our measurement equipment affect the quality of the reservoir? What if we can only partly observe the reservoir state? How does noise affect performance?
The goal of this project is to investigate such fundamental properties related to physical reservoirs.
Within our reserch group, we are working with reservoirs based on biological neural networks and nanomagnetic assemblies.
The project is interdisciplinary and falls within Unconventional computing, Artificial Intelligence, Artificial Life and Physics.