Learning with Enriched Inductive Biases for Vision-Language Models

Abstract

Vision-Language Models, pre-trained on large-scale image-text pairs, serve as strong foundation models for transfer learning across a variety of downstream tasks. For few-shot generalization tasks, i.e., when the model is trained on few-shot samples and then tested on unseen categories or datasets, there is a balance to be struck between generalization and discrimination when tweaking these models. Existing approaches typically rely on one or two strategies during training to learn task-specific knowledge, while preserving as much task-agnostic representation as possible. However, these methods overlook the importance of other useful inductive biases, thereby limiting their generalization capabilities. In this work, we propose a method – Learning with Enriched Inductive Biases (LwEIB) – to explore multiple inductive biases at the text, model, and optimization levels. Specifically, we first propose to enrich the handcrafted text prompt with Large Language Model generated descriptions for each category. To better capture structural cues in both linguistics and vision, we design two new adapters for text and image encoders, respectively. Additionally, we propose a slow-fast optimization method to explore different degrees of adaptation more efficiently, learning task-specific representations while maintaining task-agnostic ones. We empirically validate the effectiveness of LwEIB on three widely used benchmarks. Remarkably, our LwEIB outperforms numerous state-of-the-art methods across all evaluation metrics, demonstrating its efficacy and versatility. Our code is available at https://github.com/ZjjConan/VLM-LwEIB.