Sam Newby
← Learning ML

Me

My first model got worse

I’ve started learning machine learning from the ground up, and one of my first exercises was implementing linear regression using NumPy. Nothing particularly impressive. A few weights, a bias, some predictions, and a loss function.

Then came gradient descent. The idea seemed simple enough: calculate how each weight contributes to the error and adjust it to reduce that error. Except my next update made the model worse.

It was a useful reminder that understanding the maths matters. Learning rates, gradients, and how far you move with each update all affect training. I’m deliberately avoiding jumping straight into high-level ML frameworks. I want to understand what’s happening underneath them, whether that will be valuable long term in the wake of the current frontier models.

It’s early days, but these small discoveries are exactly why I’m enjoying the journey.