tinyML Asia 2021 Partner Session – Latent AI
System Engineering Aspects of End-to-End tinyML
Jan ERNST, Director of AI, Latent AI
As the tinyML domain is growing and maturing, new and more complex applications become feasible and desirable. Their requirements will be less formulaic and demand flexible ways of expressing new tasks on new (or old) data. In building such systems, how does one choose the abstractions on the components of an ML pipeline from data to deployment, while being open to the outside and agnostic to the constituent parts? How can one set the stage for scaling the number of tasks, models and data without being in the way of going deep into tinyML technologies (quantize, prune, throttle, NAS, etc.)? What are caveats in composing a system from disparate components (data, model, evaluation) of varying origin? This talk
will briefly describe one approach to some of these questions in the context of building tiny models across edge devices.