Reducing the memory footprint and improving the speed and portability of deployed models.
As more and more AI models are making their way to production—i.e, they are actually being used in day-to-day business processes—an active discussion has emerged questioning “how AI models should be deployed?” (e.g., using bloated containers? By rebuilding to stand-alone executables? By using “end-to-end” platforms with strong vendor lock in?) and “where AI models should be deployed?” (e.g., on a central cloud? On edge devices? As “close” as possible to the data generating sensors?).
For 2021 we are sharing our vision of AI deployment going forward. You can download the full white-paper here, or read the summary below.
Honestly? I don’t know. But I do think WebAssembly is a good target for ML/AI deployment (in the browser and beyond).
A machinelearning framework isn’t complete without its MNIST number classification demo. Still, we think ours is pretty unique: Once converted to a Scailable WebAssembly task, our MNIST model runs on almost any type of device.