NeuronCanvas
Learn
Module 01 · No prior knowledge required

Neural Networks

This page assumes nothing, not that you know what a neuron is, not that you remember any math from school. Every idea gets explained in plain language and an animation first, then in numbers, then you try it yourself. Everything below is live - touch it.

01

Foundations

3
  1. 01

    What Is a Neural Network?

    A first, no-math mental model for what a neural network is and why it can learn from examples.

  2. 02

    The Single Neuron

    The building block: how one neuron computes a weighted sum, a bias, and an activation.

  3. 03

    Layers and the Forward Pass

    Why one neuron isn't enough, how stacking neurons into layers builds real capacity, and what a forward pass is.

02

Activation Functions

2
  1. 01

    Activation Functions

    The classic activation functions - sigmoid, tanh, ReLU, Leaky ReLU, and PReLU - and the tradeoffs behind each.

  2. 02

    Modern Activation Functions

    The modern activation family - ELU, SELU, GELU, Swish/SiLU, Softplus, and Maxout - and why they exist.

03

Training a Network

7
  1. 01

    Loss Functions

    How a network measures how wrong a prediction is, covering squared error and cross-entropy loss.

  2. 02

    Gradients and Gradient Descent

    What a gradient is and how repeatedly stepping downhill trains every weight in a network.

  3. 03

    Optimizers

    How optimizers like momentum, RMSProp, and Adam improve on plain gradient descent.

  4. 04

    Weight Initialization

    Why weights can't all start identical, and what breaks when naive random initialization is used at scale.

  5. 05

    Xavier and He Initialization

    Xavier and He initialization, the fixes that keep signal variance stable as a network gets deeper.

  6. 06

    Regularization

    Why models overfit, the bias-variance tradeoff, and how L1/L2 penalties pull a network back toward simplicity.

  7. 07

    Training-Time Regularization

    Training-procedure techniques - dropout, early stopping, batch normalization, and data augmentation - that fight overfitting without changing the loss formula.

04

Backpropagation

2
  1. 01

    Backpropagation

    The chain rule in action: a full worked forward and backward pass, weight by weight.

  2. 02

    Vanishing Gradients

    What happens when backpropagation's chained derivatives shrink to nearly nothing in early layers, and why it matters.

05

Output Layers

2
  1. 01

    Output Activations

    Matching a network's final activation to the job its output has to do - regression, binary, or multi-class.

  2. 02

    Softmax for Multi-Class Output

    Softmax turns raw scores into a probability distribution for multi-class classification, and its clean gradient with cross-entropy.

06

Practice

1
  1. 01

    Playground

    A full interactive playground to train, break, and share a real network of your own design.