Zero-Shot Embeddings Inform Learning and Forgetting with Vision-Language Encoders

Despite the proliferation of large vision-language foundation models, estimation of the learning and forgetting outcomes following fine-tuning of these models remains largely unexplored. Inspired by work highlighting the significance of the modality gap in contrastive dual-encoders, we propose the Inter-Intra Modal Measure (IIMM). Combining terms quantifying the similarity between image embeddings and the similarity between incorrect image and label embedding pairs, the IIMM functions as a strong predictor of performance changes with fine-tuning. Our extensive empirical analysis across four state-of-the-art vision-language models (CLIP, SigLIP, CoCa, EVA-02-CLIP) and five fine-tuning techniques (full fine-tuning, BitFit, attention-weight tuning, LoRA, CLIP-Adapter) demonstrates a strong, statistically significant linear relationship: fine-tuning on tasks with higher IIMM scores produces greater in-domain performance gains but also induces more severe out-of-domain performance degradation, with some parameter-efficient fine-tuning (PEFT) methods showing extreme forgetting. We compare our measure against transfer scores from state-of-the-art model selection methods and show that the IIMM is significantly more predictive of accuracy gains. With only a single forward pass of the target data, practitioners can leverage this key insight to heuristically evaluate the degree to which a model can be expected to improve following fine-tuning. Given additional knowledge about the model's performance on a few diverse tasks, this heuristic further evolves into a strong predictor of expected performance changes when training for new tasks.

Quantified Task Misalignment to Inform PEFT: An Exploration of Domain Generalization and Catastrophic Forgetting in CLIP

Foundations models are presented as generalists that often perform well over a myriad of tasks. Fine-tuning these models, even on limited data, provides an additional boost in task-specific performance but often at the cost of their wider generalization, an effect termed catastrophic forgetting. In this paper, we analyze the relation between task difficulty in the CLIP model and the performance of several simple parameter-efficient fine-tuning methods through the lens of domain generalization and catastrophic forgetting. We provide evidence that the silhouette score of the zero-shot image and text embeddings is a better measure of task difficulty than the average cosine similarity of correct image/label embeddings, and discuss observable relationships between task difficulty, fine-tuning method, domain generalization, and catastrophic forgetting. Additionally, the averaged results across tasks and performance measures demonstrate that a simplified method that trains only a subset of attention weights, which we call A-CLIP, yields a balance between domain generalization and catastrophic forgetting.

Robust Fine-Tuning of Vision-Language Models for Domain Generalization

Transfer learning enables the sharing of common knowledge among models for a variety of downstream tasks, but traditional methods suffer in limited training data settings and produce narrow models incapable of effectively generalizing under distribution shifts. Foundation models have recently demonstrated impressive zero-shot inference capabilities and robustness under distribution shifts. However, zero-shot evaluation for these models has been predominantly confined to benchmarks with simple distribution shifts, limiting our understanding of their effectiveness under the more realistic shifts found in practice. Moreover, common fine-tuning methods for these models have yet to be evaluated against vision models in few-shot scenarios where training data is limited. To address these gaps, we present a new recipe for few-shot fine-tuning of the popular vision-language foundation model CLIP and evaluate its performance on challenging benchmark datasets with realistic distribution shifts from the WILDS collection. Our experimentation demonstrates that, while zero-shot CLIP fails to match performance of trained vision models on more complex benchmarks, few-shot CLIP fine-tuning outperforms its vision-only counterparts in terms of in-distribution and out-of-distribution accuracy at all levels of training data availability. This provides a strong incentive for adoption of foundation models within few-shot learning applications operating with real-world data. Code is available at https://github.com/mit-ll/robust-vision-language-finetuning