Questions

1. Write a new configuration for the Drums RNN model that uses an attention length of 64 and that encodes only the snares and bass drums, but inverted.
2. We have a model that underfits: what does this mean and how do we fix the problem?
3. We have a model that overfits: what does this mean and how do we fix the problem?
4. What is a technique that makes sure that, for a given training run, we stop at the optimum?
5. Why might increasing batch size make the model worse in terms of performance? Will it make it worse in terms of efficiency or training time?

6. What is a good network size?
7. Does limiting the value of the error derivative before backpropagation help or worsen the problem of exploding gradients? What is another solution to that problem?
8. Why is using a cloud provider useful to train our models?