About
About NeuronCanvas
NeuronCanvas is a free, interactive explanation of deep learning, built so a total beginner and a working ML engineer can open the same page and both walk away with sharper intuition than they came in with.
01 / 07
Origin
It started as a study problem
I was learning deep learning the normal way - lectures, a textbook, a lot of notation - and kept hitting the same wall. I could follow the derivation of backpropagation across a page and still not be able to say, in one sentence, what it was actually doing to the numbers.
The formula was legible. The mechanism wasn't.
What finally moved it was TensorFlow Playground. Forty seconds of dragging one weight and watching a boundary bend taught me more about what a weight is than the chapter had. But Playground covers one small feedforward network, and the moment I needed that same intuition for a convolution or an attention head, there was nothing to drag.
NeuronCanvas is me building the thing I wanted while I was still stuck. Every module gets written twice, once while I learn the topic properly, and again afterwards, once I know which part was actually the hard part rather than which part looked hardest.
02 / 07
Why
Why interactive beats static
Most explanations of neural networks are either a wall of matrix notation or a static diagram with arrows. Neither builds intuition, because neither responds when you push on it.
A textbook can tell you that a weight controls how much an input matters; it can't show you the decision boundary bending in real time as you drag that weight from -1 to 1.
So I'm building a third way: every concept - a single neuron, a loss function, backpropagation, a convolution kernel - gets a small interactive canvas you can push on until it makes sense, before it ever gets a formula.
03 / 07
Principles
The rules these pages are written under
There’s no editorial board here - it’s one person. So the standards have to be written down and publicly checkable, or they aren’t standards.
- 01
Behaviour before notation.
No formula appears before you've watched the thing it describes move. If a symbol turns up first on a page, that page is wrong and gets rewritten. This is the rule most often broken by textbooks and it's the one this site exists to enforce. - 02
Nothing here is a cartoon.
The networks in these playgrounds are real, trained live in your browser by an actual implementation of the forward and backward passes, not a pre-recorded animation dressed up as one. If a diagram shows you a gradient, a gradient was computed to draw it. - 03
Say when it's simplified.
Teaching means leaving things out. Where a page simplifies - a single example standing in for a batch, a two-dimensional picture of a surface that isn't two-dimensional - it says so in the text, rather than letting you discover it later and wonder what else was quietly untrue. - 04
One convention, everywhere.
Blue means a positive weight and orange means a negative one, on every diagram on this site. The convention is borrowed from TensorFlow Playground and kept deliberately unchanged. You should never have to relearn a legend between two pages. - 05
Corrections are public.
The whole site lives in a public repository. When an explanation turns out to be wrong, both the fix and the fact that it was wrong stay visible in the commit history. Nothing gets quietly edited.
What’s built, what’s next
Five architectures are queued to start, and the list is open-ended rather than final. Each one ships complete before the next one starts, rather than five modules built halfway at once.
- 01
Neural Networks
The building block everything else is made of.Live - 02
Convolutional Networks
How machines learn to see.Not built - 03
Recurrent Networks
Giving a network memory.Not built - 04
GANs
Two networks, playing a game against each other.Not built - 05
Transformers
The architecture behind modern AI.Not built
More architectures join the queue as the project grows.
05 / 07
Prior art
Built on the shoulders of prior art
NeuronCanvas didn’t invent interactive ML education - it extends two projects that got there first, to the architectures they don’t cover.
- 01
TensorFlow Playground
Set the convention this site still follows: blue for positive weights, orange for negative, a live loss curve, and a network you shape with your hands instead of reading about.Google - 02
CNN Explainer
Showed that a real, layer-by-layer computation on real input, not a simplified cartoon, is what makes a convolution finally click.Georgia Tech
06 / 07
Scope
What NeuronCanvas isn’t
Being precise about the edges is part of being trustworthy about the middle.
- 01
Not a replacement for a course.
It's built to sit next to your lecture notes or your textbook and make one concept finally click, not to certify you. Nothing here is graded, there's no certificate, and no progress is tracked. - 02
Not a research reference.
The explanations are pitched at intuition. Where you need precision - a proof, an exact algorithm, what a paper actually claimed - go to the primary source. These pages will point you at one rather than pretend to be one. - 03
Not a library or a framework.
The in-browser training engine exists so the pictures can be honest, not so you can import it. If you're building something real, use PyTorch or JAX; come back here when you want to understand why it behaved the way it did.
07 / 07
Colophon
Who’s building this
Anand Sundaramoorthy
Building NeuronCanvas
I build the explanation I wish I'd been handed, and NeuronCanvas is where that habit ended up - a site I work on module by module, in public, alongside working through the material properly myself. Nothing here gets written from memory; if a page exists, I had to understand it well enough to make it move first.
If you find an explanation on this site that's wrong, confusing, or condescending, opening an issue about it is genuinely the most useful thing you can send me, more useful than a code contribution, and much more useful than a compliment.
Neural Networks is finished. Convolutional Networks is next.
Want to help build the next module?
Convolutional networks, RNNs, GANs, and transformers are all still open, and so is whatever comes after them.