ML is one tough nut to crack. As someone who’s come here after working in Software Engineering, I had to throw the bag out of the window.

I’ve been trying to figure out what’s the secret to get good at ML. In software eng, it’s all about building things. So that would lead to project where you’re trying to build things small or from scratch.

The equivalent for ML is Experiments!

Experimentation Framework

Easiest way is Micro-Experimenting:

1. Take a paper that did something
2. Find something missing or lacking in this ⇒ Create a Hypothesis
   1. Can I mix two concepts?
   2. Did the author forget to do something?
   3. Can you improve the results by tweaking their process?
3. Test it out with a small experiment
4. If the experiment is a success ⇒ Build on it further!

Example

I read the following paper: ‣, which built a framework for 3D models.

Looked pretty cool, what if I could I could use Mixture of Features with multiple models here? (https://arxiv.org/abs/2401.06209)

This led me to take the existing code, try and mix features and see a better performance.

Of course, there’s no where to go from here, but now, I feel more confident of my understanding of 3D models.

Enjoyed my writing? Pls tell me >> [email protected]