See inference.

In general, the joint probability distribution tables are very hard to solve because it requires—for instance for binary variables—requries \(2^{n}\) entires, which is a lot.

1. how do you define very large models?
2. how do you perform inference with very large models
3. what about the data can we use to inform the design process

“If you can tell me a generative story, we can compress our joint probability distribution”. Get ready for…… inference with causality with Baysian Network.

Compositionality

Getting the right contents.

• semantic capacity tested — knowing the propositional content
• operationalization — form => meaning mapping
• measure of success — generalizing to the right meaning representation for novel expressions (this is non-trivial; multiple compatible generalization maybe applicable depending on context)

task

Given that a model can map certain expressions to their meaning representations, can they also do this for new expressions?

results

• lexical generalization (fill in new words/pairs) isn’t too hard for new NN

Its like a transforming distributions procedure, but your \(f\) is not constrained to be differentiable. So you can still sample from it.

we perform a random sample of possible next state (weighted by the action you took, meaning an instantiation of \(s’ \sim T(\cdot | s,a)\)) and reward \(R(s,a)\) from current state