Overcoming a Theoretical Limitation of Self-Attention
This addresses a fundamental problem in transformer theory for researchers, showing incremental improvements in handling specific language tasks.
The paper tackled the theoretical limitation of transformers in recognizing certain regular languages like PARITY and FIRST, demonstrating three methods to overcome it, including constructing transformers that achieve perfect accuracy and reducing cross-entropy arbitrarily close to zero.
Although transformers are remarkably effective for many tasks, there are some surprisingly easy-looking regular languages that they struggle with. Hahn shows that for languages where acceptance depends on a single input symbol, a transformer's classification decisions become less and less confident (that is, with cross-entropy approaching 1 bit per string) as input strings get longer and longer. We examine this limitation using two languages: PARITY, the language of bit strings with an odd number of 1s, and FIRST, the language of bit strings starting with a 1. We demonstrate three ways of overcoming the limitation suggested by Hahn's lemma. First, we settle an open question by constructing a transformer that recognizes PARITY with perfect accuracy, and similarly for FIRST. Second, we use layer normalization to bring the cross-entropy of both models arbitrarily close to zero. Third, when transformers need to focus on a single position, as for FIRST, we find that they can fail to generalize to longer strings; we offer a simple remedy to this problem that also improves length generalization in machine translation.