Which technique traces errors backwards to adjust parameters?

Which technique traces errors backwards to adjust parameters?

• A. Backpropagation
• B. Gradient Descent
• C. Regularization
• D. Clustering

Answer: (A)

Backpropagation uses the chain rule to trace the model’s error from the output layer back through every layer of the network, so each parameter’s influence on the loss is quantified. In a typical training cycle you perform a forward pass to compute predictions and the loss, then a backward pass to calculate gradients of the loss with respect to each weight and bias. Those gradients tell you how to adjust the parameters to reduce error, usually via an optimization step like gradient descent, where each parameter is nudged in the direction that lowers the loss. This backward-flow of information is what makes learning in neural networks possible. Gradient descent is the optimization step that uses the gradients, but it doesn’t by itself trace the error through the network’s layers. Regularization adds penalties to the loss to discourage overfitting, not to locate how each parameter contributed to error. Clustering groups data without using labels or error signals, so it isn’t about adjusting model parameters based on prediction error.

Backpropagation uses the chain rule to trace the model’s error from the output layer back through every layer of the network, so each parameter’s influence on the loss is quantified. In a typical training cycle you perform a forward pass to compute predictions and the loss, then a backward pass to calculate gradients of the loss with respect to each weight and bias. Those gradients tell you how to adjust the parameters to reduce error, usually via an optimization step like gradient descent, where each parameter is nudged in the direction that lowers the loss. This backward-flow of information is what makes learning in neural networks possible.

Gradient descent is the optimization step that uses the gradients, but it doesn’t by itself trace the error through the network’s layers. Regularization adds penalties to the loss to discourage overfitting, not to locate how each parameter contributed to error. Clustering groups data without using labels or error signals, so it isn’t about adjusting model parameters based on prediction error.