1. (25%) Can you think of a few applications for each type of RNN, a sequence-to-sequence RNN, a sequence-to-vector RNN, a vector-to-sequence RNN, and a vector-to-vector RNN?
2. (20%) How many dimensions are required for input of an RNN layer? Can you use some examples to illustrate what are these dimensions and their purposes? How about the dimensions of the RNN output?
3. (25%) Gated Recu
ent Unit(GRU) is another kind of neuron which can perform memory retention as LSTM, could you explain this and compare this with LSTM?
4. (15%, coding assignment) As we demo the bidirectional effect in IMDB dataset (see code: https:
github.com/fchollet/deep-learning-with-python-notebooks
lo
maste
6.3-advanced-usage-of-recu
ent-neural-networks.ipynb), can you try to modify RNN model to improve performance, e.g., improve accuracy to 95%? For example, we may change the number of units in each recu
ent layer, adjust the learning rate or change optimizer, or try using LSTM layers instead of GRU layers.
5. (15%, coding assignment) Can you design your own RNN to predict values of Dow Jones Industrial Average for 5 business days after today? Extra bonus will be given to who has results with the minimum MSE.