Error Backpropagation Learning Algorithm

Error Backpropagation Learning Algorithm

Error backpropagation, often simply referred to as backpropagation, is a widely used algorithm in the training of feedforward neural networks for supervised learning. Backpropagation efficiently computes the gradient of the loss function with respect to the weights of the network. This gradient is then used by an optimization algorithm, such as stochastic gradient descent, to adjust the weights to minimize the loss.

Understanding Backpropagation

At the heart of backpropagation is the chain rule from calculus, which is used to calculate the partial derivatives of the loss function with respect to each weight in the network. The term "backpropagation" reflects the way these derivatives are computed backward, from the output layer to the input layer.

The process involves two main phases: a forward pass, where the input data is passed through the network to compute the output, and a backward pass, where the gradients are computed by propagating the error backward through the network.

Forward Pass

In the forward pass, the input data is fed into the network, and operations defined by the network architecture are performed layer by layer to compute the output. This output is then used to calculate the loss, which measures the difference between the network's prediction and the true target values.

Backward Pass

The backward pass starts with the computation of the gradient of the loss function with respect to the output of the network. Then, by applying the chain rule, the algorithm calculates the gradient of the loss with respect to each weight by propagating the error information back from the output layer to the input layer. This process involves the following steps:

  1. Compute the derivative of the loss function with respect to the activations of the output layer.
  2. For each layer, starting from the last hidden layer and moving to the first, compute the gradients of the loss with respect to the layer's inputs, which are the activations of the previous layer.
  3. Compute the gradients of the loss with respect to the weights by considering the gradients with respect to the layer's inputs and the derivatives of the layer's activations with respect to its weights.

These gradients tell us how much a change in each weight would affect the loss, allowing the optimization algorithm to adjust the weights in a direction that minimizes the loss.

Updating Weights

Once the gradients are computed, the weights are updated typically using a gradient descent optimization algorithm. The weight update rule is generally of the form:

new_weight = old_weight - learning_rate * gradient

where the learning rate is a hyperparameter that controls the size of the weight updates.

Challenges with Backpropagation

While backpropagation is a powerful tool for training neural networks, it comes with several challenges:

  • Vanishing Gradients: In deep networks, the gradients can become very small, effectively preventing the weights in the earlier layers from changing significantly. This can slow down or even halt training.
  • Exploding Gradients: Conversely, gradients can also grow exponentially, causing large weight updates that can destabilize the learning process.
  • Non-Convex Loss Functions: Neural networks typically have non-convex loss functions, which means there can be multiple local minima. Gradient descent methods can get stuck in these local minima instead of finding the global minimum.

Improvements and Variants

To address these challenges, various improvements and variants of the backpropagation algorithm have been developed:

  • Activation Functions:

    Functions like ReLU and its variants help mitigate the vanishing gradient problem.

  • Weight Initialization: Techniques like Xavier and He initialization set the initial weights to values that help prevent vanishing or exploding gradients.
  • Gradient Clipping:

    This technique prevents gradients from becoming too large, addressing the exploding gradient problem.

  • Optimization Algorithms:

    Advanced optimizers like Adam, RMSprop, and AdaGrad adapt the learning rate during training to improve convergence.


Backpropagation is a cornerstone of modern neural network training. Its efficient computation of gradients has enabled the training of complex networks that can learn from vast amounts of data. Despite its challenges, backpropagation remains a fundamental technique in machine learning, and ongoing research continues to refine and improve upon this powerful algorithm.


Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323(6088), 533–536.

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.

LeCun, Y., Bottou, L., Orr, G. B., & Müller, K. R. (2012). Efficient BackProp. In Neural Networks: Tricks of the Trade (pp. 9–48). Springer, Berlin, Heidelberg.

Please sign up or login with your details

Forgot password? Click here to reset