Preserving Pre-trained Representation Space: On Effectiveness of Prefix-tuning for Large Multi-modal Models

TL;DR

A simple two-step PEFT strategy called Prefix-Tuned PEFT (PT-PEFT), which successively performs prefix-tuning and then PEFT (i.e., Adapter, LoRA), combines the benefits of both.

Abstract

Recently, we have observed that Large Multi-modal Models (LMMs) are revolutionizing the way machines interact with the world, unlocking new possibilities across various multi-modal applications. To adapt LMMs for downstream tasks, parameter-efficient fine-tuning (PEFT) which only trains additional prefix tokens or modules, has gained popularity. Nevertheless, there has been little analysis of how PEFT works in LMMs. In this paper, we delve into the strengths and weaknesses of each tuning strategy, shifting the focus from the efficiency typically associated with these approaches. We first discover that model parameter tuning methods such as LoRA and Adapters distort the feature representation space learned during pre-training and limit the full utilization of pre-trained knowledge. We also demonstrate that prefix-tuning excels at preserving the representation space, despite its lower performance on downstream tasks. These findings suggest a simple two-step PEFT strategy called Prefix-Tuned PEFT (PT-PEFT), which successively performs prefix-tuning and then PEFT (i.e., Adapter, LoRA), combines the benefits of both. Experimental results show that PT-PEFT not only improves performance in image captioning and visual question answering compared to vanilla PEFT methods but also helps preserve the representation space of the four pre-trained models.

Figures (12)

• Figure 1: Advantages of the proposed PT-PEFT, which performs 1) prefix-tuning and 2) fine-tuning (i.e., parameter-efficient or full fine-tuning) sequentially.
• Figure 2: Performance of different task adaptation methods on COCO image captioning dataset. The proposed method (PT-) consistently improves performance when combined with other methods.
• Figure 3: Qualitative image captioning results of zero-shot learning, prefix-tuned, and fine-tuned models. Although fine-tuning provides accurate answers, its results often ignore visual details compared to the other two.
• Figure 4: Accumulated and normalized singular values of features extracted from the last layer of BLIP-2. A more concave graph indicates that the singular values are more concentrated, implying the narrower representation space.
• Figure 5: Visualization of where the prefixes are inserted for different LMMs. The proposed method can be applied for general Transformer-based architectures.