Gradients

The two types of gradients that have been identified are:

• Exploding gradients
• Vanishing gradients

Exploding Gradients

As the name indicates, this happens when gradients explode to much bigger values. This could be one of the problems that RNN architectures could encounter with larger timesteps. This could happen when each of the partial derivatives is larger than 1, and multiplication of these partial derivatives leads to an even larger value. These larger gradient values cause a dramatic shift in the weight values each time they are adjusted using back propagation, leading to a network that doesn't learn well.

There are some techniques used to mitigate this issue, such as gradient clipping, wherein the gradient is normalized once it exceeds a set threshold.

Vanishing Gradients

Whether it is RNNs or CNNs, vanishing gradients could be a problem if calculated loss has to travel back a lot. In CNNs, this problem could occur when there are a lot of layers with activations...