The article discusses the use of Parameter-Efficient Fine-Tuning (PEFT) in speech processing. The authors conducted experiments to compare different PEFT methods and their layer-wise placement using Differentiable Architecture Search (DARTS). They also explored the use of ensemble learning to leverage diverse PEFT strategies. Findings show that DARTS does not outperform the baseline approach, where the same PEFT method is inserted into all layers of a Self-Supervised Learning (SSL) model. Instead, an ensemble learning approach, especially one using majority voting, shows superior performance. This suggests that different PEFT methods learn in varied ways, and their synergistic integration through ensemble learning can harness their unique learning capabilities more effectively.
Publication date: 5 Jan 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2401.02122
