Dual Task Vector Merging
A parameter-efficient knowledge injection method that merges task vectors obtained by fine-tuning related tasks (such as emotion recognition and ASR) using layer-wise learnable coefficients and injects them into a frozen pretrained backbone for adaptation to low-resource target tasks.
Timeline (last 90 days)
Dual-task vector merging is a parameter-efficient method that adapts a frozen wav2vec 2.0 backbone by injecting knowledge from sentiment and ASR tasks for Alzheimer's disease detection in low-resource settings.
Speaker-independent experiments on ADReSS and ADReSSo show that dual-task vector merging outperforms frozen probing, LoRA, and full fine-tuning.
Dual-task vector merging works best when combined with LoRA, indicating that the two parameter-efficient methods are complementary.
On the PROCESS dataset, when paired with LoRA, the advantage of this method transfers to the three-class task of HC/MCI/dementia.
In reduced experiments using only 20%, 40%, and 80% of the training data, the performance gains persist stably, confirming the reliability of this method in low-resource settings.
The design of learning merging coefficients per layer brings interpretability, as the differing degrees to which different layers rely on sentiment and ASR knowledge may reveal the hierarchical distribution of AD speech features.
All Facts (6)
Dual-task vector merging is a parameter-efficient method that adapts a frozen wav2vec 2.0 backbone by injecting knowledge from sentiment and ASR tasks for Alzheimer's disease detection in low-resource settings.
50%UnverifiedSpeaker-independent experiments on ADReSS and ADReSSo show that dual-task vector merging outperforms frozen probing, LoRA, and full fine-tuning.
50%UnverifiedDual-task vector merging works best when combined with LoRA, indicating that the two parameter-efficient methods are complementary.
50%UnverifiedOn the PROCESS dataset, when paired with LoRA, the advantage of this method transfers to the three-class task of HC/MCI/dementia.
50%UnverifiedIn reduced experiments using only 20%, 40%, and 80% of the training data, the performance gains persist stably, confirming the reliability of this method in low-resource settings.
50%UnverifiedThe design of learning merging coefficients per layer brings interpretability, as the differing degrees to which different layers rely on sentiment and ASR knowledge may reveal the hierarchical distribution of AD speech features.
50%