Autograd Visualizer

See how automatic differentiation works by watching a computation graph evaluate forward (computing values) and backward (computing gradients). This is the same algorithm that powers backpropagation in neural networks.

Computation is a graph. Autodiff is just two traversals: forward for values, backward for gradients.

Expression

Examples

Operators: + - * / ** Functions: tanh() relu() Variables: a-z

Variable values

a =

b =

c =

Computation Graph

Ready

Forward pass

Backward pass

data —

grad —

data —

grad —

data —

grad —

data —

grad —

out

data —

grad —

React Flow

Ready to start

Press play or step forward to begin the forward pass

Node Inspector

Click a node to inspect

How It Works

Forward pass: Evaluate each node in topological order (leaves first, then operations). Each node computes its value from its inputs.

Backward pass: Start from the output with gradient = 1. Traverse in reverse order, accumulating gradients using the chain rule: each node multiplies the incoming gradient by its local derivative.

The key insight: The gradient at any node tells you how much the output would change if you slightly changed that node's value. This is exactly what we need for gradient descent!