Probabilistic Embeddings for Frozen Vision-Language Models: Uncertainty Quantification with Gaussian Process Latent Variable Models
This addresses the need for uncertainty quantification in VLMs for applications like retrieval and active learning, but it is incremental as it builds on existing frozen models rather than introducing a new paradigm.
The paper tackles the problem of deterministic embeddings in vision-language models (VLMs) failing to capture uncertainties from ambiguities in visual and textual data, proposing GroVE, a post-hoc method that uses Gaussian Process Latent Variable Models to generate probabilistic embeddings from frozen VLMs, achieving state-of-the-art uncertainty calibration in tasks like cross-modal retrieval and visual question answering.
Vision-Language Models (VLMs) learn joint representations by mapping images and text into a shared latent space. However, recent research highlights that deterministic embeddings from standard VLMs often struggle to capture the uncertainties arising from the ambiguities in visual and textual descriptions and the multiple possible correspondences between images and texts. Existing approaches tackle this by learning probabilistic embeddings during VLM training, which demands large datasets and does not leverage the powerful representations already learned by large-scale VLMs like CLIP. In this paper, we propose GroVE, a post-hoc approach to obtaining probabilistic embeddings from frozen VLMs. GroVE builds on Gaussian Process Latent Variable Model (GPLVM) to learn a shared low-dimensional latent space where image and text inputs are mapped to a unified representation, optimized through single-modal embedding reconstruction and cross-modal alignment objectives. Once trained, the Gaussian Process model generates uncertainty-aware probabilistic embeddings. Evaluation shows that GroVE achieves state-of-the-art uncertainty calibration across multiple downstream tasks, including cross-modal retrieval, visual question answering, and active learning.