Edge AI Going Beyond Voice and Vision

Find the Latest SoC Solutions
for... Automotive... IoT... Security... Audio... Video

eetimes.com, May. 29, 2019 – 

Widespread public awareness of systems such as the Amazon Alexa and camera-enabled autonomous cars have made voice and vision almost automatically come to mind when discussing the role of AI in edge-device design. But AI technology is applicable well beyond voice and vision interpretation, supporting the implementation of complex system behaviors that can be intractable using conventional algorithmic development. The trick is to move the AI as close as possible to the edge.

The two signature AI applications of voice and vision happen to also illustrate the two architectural alternatives for designing AI into an embedded system. In the case of voice, both of AI's major tasks – learning and inferencing -- are handled in the cloud, where substantial processing power is available. This allows the edge device to get along with much less processing capability. It spends most of its limited capacity capturing and forwarding data to the cloud and implementing any commands coming back. This approach has the advantage of allowing a relatively inexpensive edge device design but suffers from the high bandwidth demands and latency effects of substantial WAN communications activity.

Vision systems, on the other hand, provide considerable local processing power to make real-time inferences and response decisions from live data. The machine learning that informed the inferencing engine's development took place on large computer systems prior to the edge device's implementation. The approach involves little to no WAN traffic and minimizes latency but can requires significant local processing power with an associated inflated cost.

 Back