Questions
You're at a strategy meeting with the stakeholders. They want to make sure you have the data required to answer the questions they're most interested in.
Be prepared to answer the following questions:
Network Architecture
Thomas, COO of HackPressIO
Do you think we should be using LSTM layers or GRU layers in this network?
Improving performance
Monika, Senior Developer
Often when we are generating text we see something like this:
"we counter. He stutn co des. His stanted out one ofler that concossions and was to gearang reay Jotrets and with fre colt otf paitt thin wall. Which das stimn"
What would you recommend to improve our results?
Based on your initial analysis of the data, your team feels:
- You should train for more iterations.
- You should include more layers.
- You should try a different kind of cell (for example, consider switching to/from an LSTM, GRU, etc.).
- Your model has reached its peak performance.
Potential Options
Thomas, COO of HackPressIO
We're really trying to impress our investors with your work here.
What would give us the most promise for both a quality model, but also something that could get people excited?
Learning Strategy
Johnny, Data Science Intern
I'm wondering what your views are on using a teacher forcing strategy compared to a curriculum learning strategy?
Model Evaluation
Monika, Senior Developer
Our previous team used logits in the output layer and then used Sparse Categorical Cross Entropy as the loss function. Are you planning to use that approach as well?
Additional Text
Thomas, COO of HackPressIO
We’ve been making good progress with our target author, but we feel we could improve our model's performance by supplementing with additional text, to help it learn basic language constructs better.
Which of the following would you recommend?
- We should not use any additional text, because it will change the style of the generated text.
- We should only use works that are out of copyright and now in the public domain, such as Jane Austen, or other older works.
- We could use all of Wikipedia, or other creative commons works.
- We could use all of the Internet, since we will not be copying any of the text verbatim.