Software developer · Founder of NEURON.

Why I started NEURON

The moment that made machine learning click for me was seeing exactly how a wrong prediction changes a model. You can calculate the error, trace it back to the weights, and work out why the next prediction moves in a different direction. Once I could follow that process on paper, the code made much more sense. I started NEURON to give others that same way into the subject.

How I teach

I like to start with a problem small enough to work through by hand. Take a straight line and a few data points. What does the line predict? How far off is it? What happens if you change its slope? A diagram lets you see the effect. The equations explain it. Then the code lets you repeat the experiment and change one thing at a time.

From a straight line to handwritten digits

In Machine Learning from First Principles, you follow that approach all the way to a neural network in PyTorch. You start by fitting a line, work through backpropagation step by step, and end up with a model that reads handwritten digits. Along the way, you can trace how 784 pixel values become ten class scores and see where a prediction goes wrong.

That’s the part I want you to take away: when a model makes a mistake, you know where to look.