Deep learning visualizer

Backpropagation Visualizer

See how prediction error flows backward through layers to update weights.

PhaseBackward PassLossCross EntropyOptimizerSGDUpdateWeights & Biases
Input Layer(3)Hidden Layer 1(3)Hidden Layer 2(3)Output Layer(1)0.80-0.200.600.726-0.100.500.300.70-0.300.201.211-0.500.400.901.2330.40x10.50x2-1.20x30.30h1_1h1_2h1_3h2_1h2_2h2_3y_hatPrediction: 0.84Target: 1.00Loss: 0.17Loss0.17dL/dw = -0.040Step update = +0.002dL/dw = -0.018Step update = +0.001dL/dw = -0.051Step update = +0.003
Forward predictionBackward gradient flow

Weights: Initial vs Current

ConnectionInitialGradientStep updateCurrent
x2 -> h1_20.70-0.040+0.0020.726
h1_2 -> h2_21.20-0.018+0.0011.211
h2_2 -> y_hat1.20-0.051+0.0031.233

Current Update Calculation

error = prediction - target = 0.84 - 1.00 = -0.16gradient = error * activation_derivative = -0.16 * 1.00 = -0.16new_weight = old_weight - learning_rate * gradientstep_update = -0.050 * -0.040 = +0.002current_weight = 0.700 + cumulative updates = 0.726

Backpropagation Steps

  1. Compute output error
  2. Compute output-layer gradient
  3. Propagate gradient to previous layer
  4. Accumulate gradients for weights and biases
  5. Update weights using SGD
What is Backpropagation?

Backpropagation computes how much each weight contributed to the error, then sends that gradient backward so the optimizer can update the network.

OutputHidden 2Hidden 1Input