Addressing Some Limitations of Transformers with Feedback Memory

TL;DR

This paper introduces the Feedback Transformer, a novel architecture that enhances sequential modeling by allowing all previous representations to inform future ones, leading to improved performance in language, translation, and reinforcement learning tasks.

Contribution

The paper proposes the Feedback Transformer architecture, which exposes all previous representations to future ones, overcoming limitations of standard Transformers.

Findings

Feedback Transformers outperform comparable models on multiple benchmarks.

Small, shallow Feedback Transformers achieve strong performance.

Increased representation capacity improves efficiency and effectiveness.

Abstract

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.