Talks and Poster Presentations (without Proceedings-Entry):
M. Lechner, R. Hasani, Z. Zimmer, T. Henzinger, R. Grosu:
"Designing Worm-inspired Neural Networks for Interpretable Robotic Control";
Talk: 2019 International Conference on Robotics and Automation (ICRA),
Abstract- In this paper, we design novel liquid time-constantrecurrent neural networks for robotic control, inspired by thebrain of the nematode,C. elegans. In the worm´s nervoussystem, neurons communicate through nonlinear time-varyingsynaptic links established amongst them by their particularwiring structure. This property enables neurons to expressliquid time-constants dynamics and therefore allows the net-work to originate complex behaviors with a small numberof neurons. We identify neuron-pair communication motifs asdesign operatorsand use them to configure compact neuronalnetwork structures to govern sequential robotic tasks. Thenetworks are systematically designed to map the environmentalobservations to motor actions, by their hierarchical topologyfrom sensory neurons, through recurrently-wired interneurons,to motor neurons. The networks are then parametrized ina supervised-learning scheme by a search-based algorithm.We demonstrate that obtained networks realize interpretabledynamics. We evaluate their performance in controlling mobileand arm robots, and compare their attributes to other artificialneural network-based control agents. Finally, we experimentallyshow their superior resilience to environmental noise, comparedto the existing machine learning-based methods.
Created from the Publication Database of the Vienna University of Technology.