Addressing Some Limitations of Transformers with Feedback Memory
This addresses a problem for researchers and practitioners in NLP and AI by improving model efficiency and performance, though it appears incremental as it builds on existing Transformer architectures.
The paper tackled the limitation of Transformers in exploiting sequential input due to their feedforward nature, proposing a Feedback Transformer that allows all past representations to influence future ones, resulting in small, shallow models with stronger performance on benchmarks like language modeling, machine translation, and reinforcement learning.
Transformers have been successfully applied to sequential, auto-regressive tasks despite being feedforward networks. Unlike recurrent neural networks, Transformers use attention to capture temporal relations while processing input tokens in parallel. While this parallelization makes them computationally efficient, it restricts the model from fully exploiting the sequential nature of the input. The representation at a given layer can only access representations from lower layers, rather than the higher level representations already available. In this work, we propose the Feedback Transformer architecture that exposes all previous representations to all future representations, meaning the lowest representation of the current timestep is formed from the highest-level abstract representation of the past. We demonstrate on a variety of benchmarks in language modeling, machine translation, and reinforcement learning that the increased representation capacity can create small, shallow models with much stronger performance than comparable Transformers.