Questions

1. What are the main problems RNN suffers from when learning, and what are the solutions brought by LSTMs?
2. What is a simpler alternative to LSTM memory cells? What are their advantages and disadvantages?
3. You want to configure the lookback encoder-decoder from the Melody RNN to learn structures with a 3/4 time signature. How big is the binary step counter? How are the lookback distances configured for 3 lookback distances?
4. You have the resulting vector, [0.10, 0.50, 0.00, 0.25], from the applied attention mask of [0.1, 0.5], with n = 2, and the previous step 1 of [1, 0, 0, 0] and step 2 of [0, 1, 0, x]. What is the value of x?
5. You have the following the Polyphony RNN encoding: { (START), (NEW_NOTE, 67), (NEW_NOTE, 64), (NEW_NOTE, 60), (STEP_END), (CONTINUED_NOTE, 67), (CONTINUED_NOTE, 64), (CONTINUED_NOTE, 60), (STEP_END), (CONTINUED_NOTE, 67), (CONTINUED_NOTE, 64), (CONTINUED_NOTE...