DEV Community: 0byte

The Shape of a Neuron — Weights, Bias and Activation Functions Explained Visually

0byte — Wed, 03 Jun 2026 10:39:39 +0000

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.

I've Been Using This Word My Entire Career

0byte — Sat, 18 Apr 2026 11:32:38 +0000

As a person who went through many phases, 3d graphic, real-time engines, programming and now AI, I noticed that there is a plethora of words that persist through these fields. At the very fundamental level, they might describe the same thing, but they are abstracted into a unique, niche fitting idea.

That's why, depending on your expertise you'll understand the word "model" differently. Mostly because you see it through a different abstraction or you use models in a conceptually different way. A model can be used as a predictive tool but it can also be used as a descriptor. Is some cases it's also prescriptive.

3D MODEL – MY FIRST MODEL!

If you're a 3D Modeller working in Blender, 3Ds Max or other tool, you see a mesh that can be deformed, textured, animated or rendered. That all it was for me. Something I look at and manipulate with instant feedback.

At some point I got involved in interactive applications and worked alongside Unity (a real time engine) engineers. I realised “model” meant something completely else to them. They were asking about vertex orders and weights. Flipped normals, gizmos that weren’t matching Unity coordination system. (Damn you, right-hand coordinate system!). Model wasn’t a visual thing for them. It was a weird combo of constraints and very specific properties.

Then I talked to graphic programmers. The world I knew shattered to pieces. Literally. Those solid-looking shapes I'd been happily pushing and pulling for years? Not solid. Not even connected. Just points in 3D space, grouped into triangles, moving together while remaining fundamentally separate. The mesh - the thing I thought I understood completely - was an illusion the renderer was constructing 30+ times a second.

At this point there wasn’t even a model. It was trigonometry.

Read Mode: https://0byte.io/articles/many_meanings_of_model.html

Many Meanings of the Word Model

0byte — Thu, 19 Mar 2026 22:54:48 +0000

3D MODEL – MY FIRST MODEL!

Read the rest on my blog: https://0byte.io/articles/many_meanings_of_model.html

Visual Introduction to PyTorch

0byte — Mon, 16 Mar 2026 11:21:26 +0000

PyTorch is currently one of the most popular deep learning frameworks. It is an open-source library built upon the Torch Library.

Most tutorials assume you're comfortable jumping straight into code. I made a visual introduction that walks through the core concepts step by step, with animations and diagrams instead of walls of text.

Whether you're completely new to deep learning or just want a clearer mental model of what's happening under the hood, this should help.

What's covered:
Tensors - what they are and how PyTorch thinks about data
Initialisation Functions - how weights get set up before training
The basic ML training loop - forward pass, loss, backward, update
Activation functions - ReLU, Sigmoid, Tanh and when in the training pipeline they are used

...and a few more concepts to tie it all together!