Inventors: Mohit Gupta, Matthew Dutson
UW-Madison researchers have created an event neural network system and associated methods, comprising a family of neural networks in which neurons transmit (thereby triggering downstream computation) only when there is a significant change in their activation. Neurons fire only when they have something “interesting” to say, leading to increased efficiency and accuracy.
Ideal for applications that require processing video data at high speeds with minimal power budget including:
- Robotic navigation
- Autonomous vehicles
- IoT/edge devices
- Improved efficiency for video analysis.
- High accuracy inference.
- Reduced computation costs.