Domain adaptation setups were not all born equal, and some domains are easier to adapt to and from than others.
This talk will show and attempt to estimate the difficulty (or ease) of adapting between different domains, based
on the causal effect of certain features in the data on the adapting model’s predictions. This question is
relevant in many real-life scenarios where computational resources exist in relative abundance, while labeling
and data-gathering is time-consuming, expensive or otherwise problematic.
We will discuss approaches to inspecting NLP models, leveraging existing labeled and unlabeled data which might
not seem immediately relevant, and trying to reason about the factors which affect NLP model performance.