Reproduce “ProbVLM: Probabilistic Adapter for Frozen Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Uddeshya Upadhyay, Shyamgopal Karthik, Massimiliano Mancini, Zeynep Akata
ProbVLM introduces a post-hoc probabilistic adapter that converts deterministic embeddings from frozen large-scale Vision-Language Models (VLMs) into probabilistic distributions. This method allows VLMs to quantify the inherent ambiguity in cross-modal concepts, providing calibrated uncertainty estimates that enhance active learning and model selection without requiring retraining of the base VLM.
Key insights
- Inherent cross-modal ambiguity can be efficiently captured by transforming deterministic VLM embeddings into probabilistic distributions using a lightweight, post-hoc adapter.
- Well-calibrated uncertainty estimates, derived from these probabilistic embeddings, directly correlate with retrieval performance, providing a reliable signal for model confidence.
- These uncertainty estimates are practical for downstream tasks, enabling data-efficient active learning and effective, label-free pre-trained model selection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.