Daniel Situnayake: AI on the Edge
The Gradient: Perspectives on AI
00:00
The Tradeoff Between DSP and Model to Get the Best Possible Output
There's a duality where you've got two algorithms, there's a signal processing algorithm which often takes quite a lot of time to run. And you've got your model side by side with that, which is learning to make sense of those outputs. So if you do a better job of feature extraction with the DSP, then maybe you don't need as big a model. However, there's also heterogeneous hardware. On a given microcontroller, you might have acceleration for certain configurations of signal processing, or acceleration for certain kind of configurations of model. We're trying to build tools that allow practitioners to solve all of these pieces at once and kind of take a view over the entire
Transcript
Play full episode
