Few-shot Adaptation to Distribution Shifts By Mixing Source and Target Embeddings

TL;DR

MixPro is a data-efficient method for few-shot adaptation to distribution shifts that combines source and target embeddings to improve classifier performance with minimal target data.

Contribution

We introduce MixPro, a novel approach that mixes source and target embeddings for effective few-shot adaptation, supported by theoretical analysis and extensive experiments.

Findings

MixPro outperforms baselines by up to 7% on various datasets.

It requires only 2-4 target examples for effective adaptation.

MixPro effectively preserves source and target features while reducing noise.

Abstract

Pretrained machine learning models need to be adapted to distribution shifts when deployed in new target environments. When obtaining labeled data from the target distribution is expensive, few-shot adaptation with only a few examples from the target distribution becomes essential. In this work, we propose MixPro, a lightweight and highly data-efficient approach for few-shot adaptation. MixPro first generates a relatively large dataset by mixing (linearly combining) pre-trained embeddings of large source data with those of the few target examples. This process preserves important features of both source and target distributions, while mitigating the specific noise in the small target data. Then, it trains a linear classifier on the mixed embeddings to effectively adapts the model to the target distribution without overfitting the small target data. Theoretically, we demonstrate the advantages of MixPro over previous methods. Our experiments, conducted across various model architectures on 8 datasets featuring different types of distribution shifts, reveal that MixPro can outperform baselines by up to 7\%, with only 2-4 target examples.