Understanding Domain Learning in Language Models Through Subpopulation Analysis
This work provides incremental insights into domain learning mechanisms in language models, which is relevant for researchers in natural language processing and neural network interpretability.
The paper tackled the problem of understanding how domain information is encoded in language models by analyzing latent representations using subpopulation analysis with SVCCA on Transformer-based models, finding that model capacity affects domain storage differently across layers and that larger models embed domain-specific information similarly to conjoined single-domain models.
We investigate how different domains are encoded in modern neural network architectures. We analyze the relationship between natural language domains, model size, and the amount of training data used. The primary analysis tool we develop is based on subpopulation analysis with Singular Vector Canonical Correlation Analysis (SVCCA), which we apply to Transformer-based language models (LMs). We compare the latent representations of such a language model at its different layers from a pair of models: a model trained on multiple domains (an experimental model) and a model trained on a single domain (a control model). Through our method, we find that increasing the model capacity impacts how domain information is stored in upper and lower layers differently. In addition, we show that larger experimental models simultaneously embed domain-specific information as if they were conjoined control models. These findings are confirmed qualitatively, demonstrating the validity of our method.