The word “neuron” in machine learning can make the whole thing sound unnecessarily complex. But when you build it step by step, the idea becomes surprisingly geometric.
In this video I explore neural networks from a geometric perspective. We start with a single neuron and build an intuition for weights, bias, decision boundaries, and activation functions by visualising them as surfaces moving through an input space.
Instead of treating activation functions as formulas to memorise, we look at what they do to the geometry. Why does a sigmoid produce probabilities? Why does ReLU behave differently? What exactly is a decision boundary, and what is the neuron actually deciding?
Topics covered:
Affine transformations
Input space and decision boundaries
The geometric interpretation of a neuron
Binary Step, Sigmoid, tanh, ReLU, Leaky ReLU, ELU, Swish, and GELU
How training moves and rotates boundaries through space
My experience learning machine learning is that intuition usually comes before terminology. Once you can see the geometry, the equations stop feeling like magic.
Top comments (0)