RNN Interactive Playground

Sequence Unrolling

See how a single RNN cell is reused across every timestep — the same weights process each input character sequentially while passing hidden state forward.

Key Concepts

Each RNN cell shares identical weights W, U, b
W (recurrent weight): multiplies the previous hidden state h_t-1
U (input weight): multiplies the current input x_t
b (bias): a learned offset added at each step
Hidden state h_t flows left → right as a memory chain
At each step: h_t = tanh(W·h_t-1 + U·x_t + b)

How to Use

Press Play to animate the sequence processing
Use the Step slider to jump to any timestep
Adjust Speed (0.5x, 1x, 2x)
The heatmap below shows hidden state values at each timestep — columns = timesteps, rows = hidden dimensions (d₀…d_n-1)

What to Observe

The active cell pulses and shows internal operations (Embed → tanh → Linear)
Hidden state preview mini-cells show activation patterns
The heatmap reveals which dimensions are most active at each timestep
Blue = negative, Orange = positive activations

Hidden State Lab

A hidden state is a vector of neuron activations — a distributed memory. Watch it change as new inputs arrive.

Controls

Hidden Size 8

Activation

Observe

Each square = one neuron's activation
Blue = negative, Red = positive
Click a neuron to track its value
Try adding noise to see recovery

Selected Neuron

—

Click a neuron

Activation History

Step: 0

Sequence Learning

Watch a toy RNN learn temporal patterns. See how hidden size, learning rate, and task complexity affect learning.

Task

Hyperparameters

Learning Rate 0.010

Hidden Size 8

Max Epochs 0/50

Input Sequence

—

Prediction Probabilities

Training Loss

Ready