Prof. Dr. Radu Grosu, TU Wien

Abstract:

We introduce a new type of recurrent neural networks (RNNs) which we call WormNets, as they were inspired by a biophysical model for neurons and synapses in the C. Elegans worm. WormNets are interpretable, smaller in size, and more robust to noise attacks when compared to classic RNNs. They can also take advantage of the rich trove of neural policies developed by nature through billions of years of evolution. We show how to model with WormNets and learn their parameters, or even learn the WormNets from scratch, without considerable penalty, by using state-of-the-art RNN learning techniques. We also discuss how to verify WormNets.