Deep Learning - Recurrent Neural Networks with TensorFlow - GRU and LSTM (Part 2)

The video tutorial covers the differences between LSTM and GRU, highlighting their structures, gates, and performance based on recent research. It emphasizes the importance of experimentation over philosophical reasoning in machine learning. The LSTM is explained in detail, including its gates and states, and practical implementation tips are provided for using LSTM in TensorFlow and Keras.

What is the main limitation of simple RNNs that GRUs aim to solve?

• Difficulty in remembering long-term dependencies

• Complexity in implementation

• Inability to process sequential data

• High computational cost

Why were GRUs initially considered more efficient than LSTMs?

• They are more accurate

• They require less data

• They are easier to implement

• They have fewer parameters

What recent finding has changed the perception of LSTMs compared to GRUs?

• LSTMs outperform GRUs in most tasks

• LSTMs are faster to train

• GRUs are more widely used

• GRUs are more flexible

What additional state does an LSTM have compared to a GRU?

• Memory state

• Cell state

• Activation state

• Output state

Which gate in an LSTM controls the information flow from the cell state to the hidden state?

• Update gate

• Output gate

• Input gate

• Forget gate

What is the default activation function used in LSTM calculations?

• Tanh

• Softmax

• Sigmoid

• ReLU

What is the purpose of the 'return sequences' option in RNNs?

• To simplify the model

• To increase computational efficiency

• To return all hidden states

• To return only the final output

What does setting 'return state' to true in an LSTM return?

• Only the cell state

• Only the output

• Output, hidden state, and cell state

• Output and hidden state

How can you make an LSTM GPU compatible in TensorFlow?

• Use a different optimizer

• Follow TensorFlow documentation requirements

• Increase the number of layers

• Use a custom activation function

What is a key similarity between the APIs of GRUs, LSTMs, and simple RNNs?

• They all use the same activation function

• They all output the final hidden state

• They all have the same number of parameters

• They all require the same input format