I’m slowly getting the hang of real-world machine learning
I have spent 6 years reading and learning about ML. And I have spent thousands of hours practicing real-world applications of ML in my own life.
Here's how I first got interested in real-world ML:
Machine learning gives computers the capability to do, what they couldn’t do before.
In 2016, at the very beginning of my PhD, I took a course on advanced image analysis. The professor had us program a neural network from scratch to classify MNIST data. Matlab. No imports. I hyperfocused for two weeks and made 2nd place out of the entire course.
I was hooked.
I saw how machine learning could solve some of the specific issues we were unable to solve in numerics.
The potential to speed up slow iterative solutions, abstract noise, and interpolate over complicated data relations is something that machine learning is uniquely poised to do. It can step in where existing theory often falls short, like signal in the noise that we have not identified as signal yet.
So my goal over the next 3 years is:
- Dive deeper into transformers and how exactly they relate to numerical methods.
- Understand probabilistic methods and forecasting better.
- Better understand ML operations in existing businesses.
- Promote and encourage better research practices applying ML in science.
And some more, it’s 3 years, can’t wait to see where the field develops towards.