Covariant spatio-temporal receptive fields for spiking neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Biological nervous systems offer a blueprint for more efficient computing.
• Neuromorphic engineering seeks to optimize hardware and algorithms simultaneously.
• Current neuromorphic implementations lack guiding theories for efficient design.

Purpose of the Study:

• To present a principled computational model for neuromorphic systems.
• To develop a theory for efficient spatio-temporal signal processing in neuromorphic hardware.
• To improve the training of spiking neural networks for event-based vision.

Main Methods:

• Developed a model based on spatio-temporal receptive fields using affine Gaussian kernels and leaky-integrator/leaky integrate-and-fire models.
• Ensured the model's covariance to spatial affine and temporal scaling transformations.
• Applied the model as a prior in an event-based vision task using spiking networks.

Main Results:

• Demonstrated theoretical covariance to spatial and temporal scaling, mirroring mammalian visual processing.
• Showed significant improvement in training spiking networks for event-based vision tasks.
• Established a theoretically grounded approach for processing spatio-temporal signals.

Conclusions:

• The proposed model provides a theoretically sound framework for neuromorphic systems.
• This work bridges scale-space theory and computational neuroscience for efficient signal processing.
• Findings are directly applicable to event-based vision and extendable to other spatio-temporal tasks.