![[PukiWiki]](image/pukiwiki.png "[PukiWiki]")
Conventional processors, like CPUs and GPUs, are efficient in processing dense, synchronously delivered data structures, not sparse, asynchronous event streams. Throughput is maintained by keeping the instruction and data pipelines as full as possible, even at the cost of executing redundant computations on unchanging data [Ref].
Recently, a new generation of natively event-based neuromorphic processors has appeared that can operate on sensor event streams directly [Ref.]. These many-core systems instantiate large populations of spiking neurons in massively parallel, low-power hardware, and inherit all of the advantages of event-based sensors, such as efficient, data-driven resource consumption[ Ref.].
Processing latency can be as fast as the event propagation time through the longest chain of neurons, so a neural network running on these systems can react to a stimulus in tens of milliseconds, fast enough to identify quick motions like hand gestures in real-time.
https://adaptive.u-aizu.ac.jp/aslint/index.php?Hand-gesture%20Recognition
| Date | Task |
