Machine learning can be a frustrating endeavour, and I know several ML engineers who have transitioned to SWE roles – the unpredictability of machine learning being at the heart of their frustrations. You never know if or when your model will be good enough for production or if your new algorithm- or data idea will give any improvements. Usually, they don’t. Additionally, the feedback loop in ML is typically slow. These factors combined can lead to a sense of loss of control.

A while ago I came upon this great blog post by Alex Olivier that introduced Google Cloud Run, where you can deploy your webservers for next to nothing. This was perfect timing as I’m currently working on a side project of my own, and had come to the point where I needed to deploy the first version of it.

In our new paper on lightweight segmentation we propose a new architecture utilizing an information blocking decoder and spatial squeeze modules. Check it out here.

In our recent paper ExtremeC3Net: Extreme Lightweight Portrait Segmentation Networks using Advanced C3-modules, we introduce a new extremely lightweight portrait segmentation model consisting of a two-branched architecture based on the concentrated-comprehensive convolutions block. Our method reduces the number of parameters from 2.08M to 37.9K (around 98.2% reduction), while maintaining the accuracy within a 1% margin from the state-of-the-art portrait segmentation method. Check out the full paper here.

I started working at a ML start-up right after graduation and quickly realized that I was not prepared for it. While I had graduated with a bachelor’s degree in engineering physics, a master’s in ML, and had a fair amount of knowledge about algorithms, the skills needed at a company were quite different from what I had learned at university. At the workplace, there was limited value in being able to do complex integrals or prove convergence bounds, while the ability to get things up and running with whatever means possible was crucial.