A Modular-based Strategy for Mitigating Gradient Conflicts in Simultaneous Speech Translation

TL;DR

This paper introduces MGCM, a modular gradient conflict mitigation strategy that improves simultaneous speech translation performance and reduces GPU memory usage by detecting and resolving conflicts at a fine-grained modular level.

Contribution

The paper presents a novel modular gradient conflict mitigation approach that effectively addresses optimization conflicts in multi-task SimulST, outperforming existing methods.

Findings

Achieves 0.68 BLEU score improvement in offline tasks.

Reduces GPU memory consumption by over 95%.

Enhances performance under medium and high latency conditions.

Abstract

Simultaneous Speech Translation (SimulST) involves generating target language text while continuously processing streaming speech input, presenting significant real-time challenges. Multi-task learning is often employed to enhance SimulST performance but introduces optimization conflicts between primary and auxiliary tasks, potentially compromising overall efficiency. The existing model-level conflict resolution methods are not well-suited for this task which exacerbates inefficiencies and leads to high GPU memory consumption. To address these challenges, we propose a Modular Gradient Conflict Mitigation (MGCM) strategy that detects conflicts at a finer-grained modular level and resolves them utilizing gradient projection. Experimental results demonstrate that MGCM significantly improves SimulST performance, particularly under medium and high latency conditions, achieving a 0.68 BLEU score gain in offline tasks. Additionally, MGCM reduces GPU memory consumption by over 95\% compared to other conflict mitigation methods, establishing it as a robust solution for SimulST tasks.