Which transformer architecture fits my data? A vocabulary bottleneck in self-attention
This addresses the challenge of configuring Transformer architectures for new data types, providing insights for researchers and practitioners in machine learning, though it is incremental as it builds on existing Transformer theory.
The paper identifies an embedding rank bottleneck in Transformer self-attention that limits the benefit of width, linking optimal depth-to-width ratios to vocabulary size and rank, and demonstrates this with empirical evidence while revealing 25%-50% size redundancies in models like ALBERT and T5.
After their successful debut in natural language processing, Transformer architectures are now becoming the de-facto standard in many domains. An obstacle for their deployment over new modalities is the architectural configuration: the optimal depth-to-width ratio has been shown to dramatically vary across data types (e.g., $10$x larger over images than over language). We theoretically predict the existence of an embedding rank bottleneck that limits the contribution of self-attention width to the Transformer expressivity. We thus directly tie the input vocabulary size and rank to the optimal depth-to-width ratio, since a small vocabulary size or rank dictates an added advantage of depth over width. We empirically demonstrate the existence of this bottleneck and its implications on the depth-to-width interplay of Transformer architectures, linking the architecture variability across domains to the often glossed-over usage of different vocabulary sizes or embedding ranks in different domains. As an additional benefit, our rank bottlenecking framework allows us to identify size redundancies of $25\%-50\%$ in leading NLP models such as ALBERT and T5.