Related Experiment Videos
A regenerating spiking neural network.
1Complex Adaptive Organically-Inspired Systems Group (CAOS), Department of Computer and Information Science, Norwegian University of Science and Technology, N-7491 Trondheim, Norway. federici@idi.ntnu.no
Summary
This study shows that artificial neural networks can develop fault-tolerant structures, mimicking biological self-healing. Evolved spiking neural networks demonstrated recovery from cell loss, improving robot navigation.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
- Robotics
Background:
- Artificial neural networks exhibit graceful degradation upon unit loss.
- Biological systems possess self-healing mechanisms for fault tolerance.
- Spiking neural networks offer a biologically plausible model for neural computation.
Purpose of the Study:
- To investigate the fault tolerance of spiking neural networks (SNNs) during development and operation.
- To model the self-healing capabilities of biological systems in artificial neural networks.
- To assess the impact of cell loss on SNNs controlling simulated robots.
Main Methods:
- Developed a model for SNNs that undergo simulated cell loss during development.
- Subjected networks to random faults during development and operational mutilation.
- Evolved SNNs to control Khepera robots in a navigation task.
- Compared fault tolerance in plastic and non-plastic SNNs.
Main Results:
- Developed SNNs exhibited fault-tolerant structures capable of recovering normal operation.
- Both plastic and non-plastic networks showed varying degrees of recovery from cell loss.
- The model demonstrated the potential for SNNs to develop resilience similar to biological self-healing.
Conclusions:
- Spiking neural networks can be developed to be inherently fault-tolerant.
- The study provides insights into creating more robust artificial intelligence systems.
- This research highlights the potential of biologically inspired computing for self-healing capabilities.