Why bother with geometry? On the relevance of linear decompositions of Transformer embeddings
This work addresses the relevance of geometric interpretations in Transformer models for researchers in NLP and machine learning, highlighting incremental insights into model-specific characteristics.
The study investigated whether linear decompositions of Transformer embeddings are empirically meaningful by analyzing machine-translation decoders, finding that decomposition-derived indicators correlate with model performance but show high variability across runs.
A recent body of work has demonstrated that Transformer embeddings can be linearly decomposed into well-defined sums of factors, that can in turn be related to specific network inputs or components. There is however still a dearth of work studying whether these mathematical reformulations are empirically meaningful. In the present work, we study representations from machine-translation decoders using two of such embedding decomposition methods. Our results indicate that, while decomposition-derived indicators effectively correlate with model performance, variation across different runs suggests a more nuanced take on this question. The high variability of our measurements indicate that geometry reflects model-specific characteristics more than it does sentence-specific computations, and that similar training conditions do not guarantee similar vector spaces.