Author(s): DrSwarnenduAI

Originally published on Towards AI.

For a decade, we asked if RNNs can represent what Transformers represent. We proved they can. We forgot to ask how expensively. That omission just cost us ten years.

“Can our architecture represent everything a Transformer can?” The benchmarks run. The perplexity scores appear. The answer, roughly, is yes.

The article discusses the limitations of recurrent neural networks (RNNs) compared to transformers, particularly regarding their ability to represent complex structures succinctly. It reveals that while RNNs can compute functions similar to transformers, they require exponentially more parameters, especially in tasks requiring deep compositional structures. The piece highlights that evaluations of model efficiency often overlook the underlying parameter costs, which become apparent at higher nesting depths in tasks. Ultimately, it advocates for hybrid architectures that leverage the strengths of both RNNs and transformers to optimize performance in various computational contexts.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI