Meta-Adapters: Parameter Efficient Few-shot Fine-tuning through Meta-Learning

Consistent improvements in the representational capacity of large pre-trained transformers has made it increasingly viable to serve these models as shared priors that can be fine-tuned on a large number of downstream tasks. However, fine-tuning the entire model for every task of interest makes a copy of all the model parameters, rendering such scenarios highly impractical. Recently introduced Adapter methods propose a promising alternative, one where only a small number of additional parameters are introduced per task specifically for fine-tuning. However, Adapters often require large amounts of task-specific data for good performance and don't work well in data-scarce few-shot scenarios. In this paper, we approach parameter-efficient fine-tuning in few-shot settings from a meta-learning perspective. We introduce Meta-Adapters, which are small blocks of meta-learned adapter layers inserted in a pre-trained model that re-purpose a frozen pre-trained model into a parameter-efficient few-shot learner. Meta-Adapters perform competitively with state-of-the-art few-shot learning methods that require full fine-tuning, while only fine-tuning 0.6% of the parameters. We evaluate Meta-Adapters along with multiple transfer learning baselines on an evaluation suite of 17 classification tasks and find that they improve few-shot accuracy by a large margin over competitive parameter-efficient methods, while requiring significantly lesser parameters for fine-tuning. Moreover, when comparing few-shot prompting of GPT-3 against few-shot fine-tuning with Meta-Adapters, we find that Meta-Adapters perform competitively while working with pre-trained transformers that are many orders of magnitude (1590x) smaller in size than GPT-3.