What Backpropagation Actually Does
Follow prediction error backward to see how each trainable parameter receives a useful correction.
What Backpropagation Actually Does explained visually
Use a delivery robot to connect the forward pass, loss, gradients, and optimizer updates.
What you will learn
Trace derivatives, gradients, and parameter responsibility through a neural network.
Complete the forward pass
Examples move through weighted layers until the network produces a prediction.
Measure the prediction error
A loss function converts prediction error across a mini-batch into a value the network can minimize.
Assign parameter responsibility
Backpropagation calculates how a small change in each parameter would affect the loss.
Update and repeat
The optimizer combines each gradient with the learning rate and repeats the cycle over many mini-batches.
Quick check: test your understanding
Where does backpropagation begin?
At the output, where the loss is measured.
What does a gradient describe?
How a small parameter change would affect the loss.
What controls update size?
The learning rate.
Accessible referenceLesson 33 transcript374 words · Matches the published narration and captionsOpen transcript
In Lesson 32, we used learning curves to decide whether a model has high bias, high variance, or a useful balance. Now we will look inside the training process and see how a neural network corrects its mistakes. Imagine a delivery robot approaching a road intersection.
Its camera sees several visual clues, and the network predicts that the path ahead is clear. The correct answer says that an obstacle is present. The network made a mistake, but knowing that it was wrong is not enough. It must determine which internal connections contributed to that prediction and
how each one should change. Training begins with a forward pass. A mini-batch of examples enters the network. Each layer combines inputs with weights and biases, applies an activation function, and sends its outputs to the next layer. At the final layer, the network produces predictions.
The loss function compares those predictions with the correct targets. It converts the errors from the mini-batch into one value. A larger loss indicates that the current parameters produced poorer predictions. Next comes the backward pass, commonly called backpropagation. The process begins at the output, where the loss was measured,
and moves backward through the layers. At every connection, it computes a gradient. That gradient estimates how a small change in the weight would affect the loss. The optimizer then uses those gradients to update the parameters. A common update follows a simple rule. The new weight equals the current weight minus the learning rate
multiplied by the gradient. The gradient provides the direction and strength of the correction. The learning rate controls the size of the step. If the step is too large, training may jump past a useful solution. If it is too small, learning may become extremely slow.
The network repeats this cycle across many mini-batches. It performs a forward pass, measures the loss, sends responsibility backward, and updates the parameters. Over time, useful connections become stronger and misleading connections change. Backpropagation does not directly tell the network what feature to learn. It provides a systematic way to assign responsibility for an error
across many connected parameters. To understand how that responsibility is measured, we first need to understand gradients and derivatives. That is the focus of Lesson 34.