[44] Hady Elsahar - NLG from Structured Knowledge Bases (& Controlling LMs)
The Thesis Review
How to Maximize Reward Expectations
There are two approaches here that has people have been following so far I would like to maximize certain reward expectation of certain reward on the expense of anything else and then try to make your reward as representative as possible to everything you want to model. So a language model that doesn't generate anything or generates only one sentence is an untoxic model so that's a completely useless and reinforce for example we do in this podcast. You will stumble upon that expectations are the average of the samples but you don't have a global view of what you want to do, he says. He concludes by saying there is no escape from looking into distributional matching framework versus the reward maximization approach.
Remember Everything You Learn from Podcasts
