🚂 Backpropagation Through Time Visualizer

Watch the gradient shrink into dust as it travels back through a long unrolled sequence.

📋 How To Use This Lab
  1. Press ▶ Backpropagate — a real gradient pulse travels RIGHT to LEFT through every time step.
  2. Watch it shrink as it goes — by the time it nears the start, it may be nearly invisible dust.
  3. Drag Sequence Length longer → the earliest blocks clearly never get a meaningful gradient at all.
  4. Switch Activation to ReLU → the pulse survives MUCH further back, barely shrinking.
Sequence Length = 20
Activation = Tanh
Gradient at t=0: —
Unrolled RNN — 20 Time Steps, Left = Start of Sequence t=0 (oldest word) t=T (newest word)
🚂 CONDUCTORPress Backpropagate — I'll send the error signal all the way back down the train, one car at a time.
🧑‍🏫 PROF. TORCHEach step back multiplies the gradient by the activation's derivative — for Tanh, that's always less than 1, so it shrinks every single time.