How do LSTM and GRU avoid to overcome the vanishing gradient problem?

Question

I'm watching the video Recurrent Neural Networks (RNN) | RNN LSTM | Deep Learning Tutorial | Tensorflow Tutorial | Edureka where the author says that the LSTM and GRU architecture help to reduce the vanishing gradient problem. How do LSTM and GRU prevent the vanishing gradient problem?

Answer 1

LSTMs solve the problem using a unique additive gradient structure that includes direct access to the forget gate's activations, enabling the network to encourage desired behaviour from the error gradient using frequent gates update on every time step of the learning process.

Answer 2

LSTM passes the previous state's hidden weights to the current state. This simple yet effective solution helps them in minimizing the Vanishing gradient, because all states now have some information about all of the previous states. Consider like you are trading and you have all the numbers from a year ago, which surely helps in making better decisions!

I highly recommend this article, which explains the concept very well.

Answer 3

By controlling the flow of information and gradients, LSTMs mitigate the vanishing gradient problem commonly encountered in traditional RNNs. The gating mechanisms enable LSTMs to learn and retain important information over long sequences. Since the gates are differentiable, their weights are trainable. This means that

• the gates use sigmoid and tanh activation functions, which are differentiable and have bounded output ranges. This property ensures that gradients can flow through the gates without becoming too small or too large.
• the forget gate, input gate, and output gate operations in LSTMs involve element-wise multiplication (Hadamard product) between the gate values and other vectors/matrices. Element-wise multiplication preserves the gradients' magnitude and direction, which allows information to flow without diminishing too quickly.
• the memory cell allows the network to selectively keep or drop information over multiple time steps. This ability to remember or forget information is crucial for capturing long-term dependencies in sequential data.

So, by controlling the flow of the gradient, the gating mechanisms play a key role in determining how information is updated and propagated through the network while preserving the gradients. Thus, the gating mechanisms, along with the LSTM's architecture design (memory cell), contribute to effective gradient flow during training. This gives the network the ability to learn and retain long-term dependencies in sequential data while avoiding the vanishing gradient problem.