The following summaries are my takeaways from attending the TinyML Asia Summit in the week of November 16-19 2020.
First off, it was great to have an option to attend the virtual conference for free. I only recently became interested in the topic and having this low-barrier to gaining a tonne of new information was just amazing! Winning a $30 Starbucks gift card didn't hurt either :P
The conference was a week of thematically linked talks. You can jump to different sections, organized by day, here:
Hardware
Algorithms
Systems and Applications
I summarize the talks for each day below, so you can skip ahead to whatever interests you.
I think if I had to peg one thing as a huge trend in TinyML, I would say it is the strong emphasis on hardware-software co-design. This is natural, really, given that the layers of abstraction we see today came with computers having increasing resources.
TinyML applications have to be as lite as possible because power and energy are precious resources. This means simpler hardware, with most of the complexity in the compilers/software region. Many of the talks above mention