I've been participating with the Erlang Ecosystem Foundation Slack channel where we cover Elixir's Nx, Axon, and Bumblebee technologies and help each other.
My main focus right now is language models and multi-modal models. These are the areas with the most interest for a wide range of companies. The emphasis is on learning skills using small datasets and small models. We've now reached the point where one to three billion parameter models are interesting to work with. These smaller models can run on my local, older, GPUs. I'll be writing articles on my experiences with these models.
I presented at ElixirConf US 2024. I wanted to bring some of the Fast.ai image processing techniqies into the Axon capabilities. However, the problem was a little larger than I could complete and cover in the presentation. I still want to bring the image augmentation capabilities, one-cycle learning rate tools, and other capabilities into Nx/Axon. But that may need to wait for a while longer.
Previously, I worked on a self-funded startup in the healthcare domain. With a small set of folks working after hours, we tried to find a product market fit. With lessons learned and shifts in direction, we ended up focusing on bed exit prediction using video streaming and deep learning image recognition. We were using an early version of the Elixir Membrane framework and Fast.ai Pytorch library to perform "real-time" image recognition of bed exit actions. COVID and challenges acquiring representative videos of bed exists, when everyone was locked down, caused some challenges. We finally shut the company down. My Knowfalls experiences working with Elixir and machine learning is the driving factor in my interest in working with the Nx, Axon and Bumblebee libraries.
