Related Experiment Videos
Sensor fusion by neural networks using spatially represented information
T Boss1, V Diekmann, R Jürgens
1Sektion Neurophysiologie, Universität Ulm, Germany.
Biological Cybernetics
|November 28, 2001
Summary
This study introduces a neural network model for multisensory convergence, demonstrating adaptive signal fusion based on input concordance. The model effectively suppresses sensor noise through averaging and filtering, with performance varying based on input similarity.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
- Signal Processing
Background:
- Multisensory convergence is crucial for robust perception.
- Existing models often lack adaptive mechanisms for varying input reliability.
- Neural networks offer a framework for complex sensory data integration.
Purpose of the Study:
- To analyze a novel neural network model for fusing signals from multiple sensors.
- To investigate the model's adaptive processing modes based on input signal concordance.
- To evaluate the model's noise reduction capabilities in multisensory convergence.
Main Methods:
- Development of a two-stage neural network with a lateral-inhibition feedback layer.
- Simulation of signal fusion under varying degrees of input signal difference.
- Mathematical analysis of network dynamics using first-order low-pass filters.
Main Results:
- The model exhibits weighted averaging for similar inputs and signal suppression for dissimilar inputs.
- Network bandwidth and noise suppression are nonlinearly dependent on input signal concordance.
- Internal neuronal noise reduces but does not eliminate noise suppression for similar inputs.
Conclusions:
- The proposed neural network effectively performs adaptive multisensory convergence.
- The model demonstrates robust noise reduction capabilities, modulated by input signal agreement.
- The architecture shows potential for extension to more than two sensory inputs.