Gradients for dummies

What Are Gradients?

Gradients are simply the mathematical representation of how much a given parameter, like weights in a neural network, affects the output of the model.

In other words, gradients tell us how to adjust the parameters to minimize the error (or maximize the performance) of the model.

The Backpropagation Algorithm

Backpropagation is the most widely used algorithm to train neural networks. It works by calculating the gradients of the loss function with respect to each weight by using the chain rule, computing the gradient one layer at a time, iterating backward from the last layer to the first.

Vanishing Gradients

Vanishing gradients occur when the gradients become extremely small as they propagate through the layers. This happens primarily in deep networks with many layers. When the gradients vanish, they make it very difficult for the weights to update in the early layers, resulting in slow or no learning at all.

A common cause of vanishing gradients is the activation functions used in neural networks, like the sigmoid or hyperbolic tangent (tanh) functions. These functions squash the input into a small range (between 0 and 1 for sigmoid, and -1 and 1 for tanh), causing the gradients to become very small and sometimes vanish.

Exploding Gradients

On the other hand, exploding gradients occur when the gradients become excessively large, causing the weights to update too rapidly. This can lead to an unstable learning process, with the model’s performance oscillating wildly, making it difficult to converge to an optimal solution.

Exploding gradients often happen in deep networks with large weights and long sequences, especially when using activation functions like the Rectified Linear Unit (ReLU) that do not constrain the output range.

Solutions to Vanishing and Exploding Gradients

To combat vanishing gradients:

  1. Use activation functions that alleviate the problem, like ReLU, Leaky ReLU, or Parametric ReLU.
  2. Implement batch normalization to normalize layer inputs and improve the flow of gradients.
  3. Apply techniques like skip connections or residual connections, as seen in ResNet architecture, to allow gradients to bypass certain layers.

To address exploding gradients:

  1. Use gradient clipping, which limits the maximum value of the gradient during backpropagation.
  2. Initialize weights carefully, using techniques like Glorot or He initialization, to prevent large gradients at the beginning of training.


Vanishing and exploding gradients are common challenges faced in deep learning. By understanding the causes behind these issues and applying the appropriate solutions, you can improve your model’s training process and overall performance. With this newfound knowledge, you’re one step closer to mastering the art of deep learning.

Leave a Comment