From Uncertainty to Trust: Enhancing Reliability in Vision-Language Models with Uncertainty-Guided Dropout Decoding

2024-12-09

arXiv.org

AI Summary

LVLMs frequently suffer from hallucinations and unreliable outputs due to misinterpretation of visual inputs. To address this, we propose an uncertainty-guided inference-time visual token dropout method: (1) the first adaptation of dropout to the visual token level during inference; (2) decoupled modeling of epistemic and aleatoric uncertainty, with explicit focus on quantifying perceptual errors; (3) uncertainty estimation via projection of visual tokens into the text embedding space, followed by weighted masking; and (4) robust, training-free correction via multi-context masked decoding and ensemble prediction. Evaluated on CHAIR, THRONE, and MMBench, our method significantly reduces object hallucination (OH) while substantially improving output reliability and cross-scenario generation quality.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Safety and Robustness

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Large vision-language models (LVLMs) demonstrate remarkable capabilities in multimodal tasks but are prone to misinterpreting visual inputs, often resulting in hallucinations and unreliable outputs. To address these challenges, we propose Dropout Decoding, a novel inference-time approach that quantifies the uncertainty of visual tokens and selectively masks uncertain tokens to improve decoding. Our method measures the uncertainty of each visual token by projecting it onto the text space and decomposing it into aleatoric and epistemic components. Specifically, we focus on epistemic uncertainty, which captures perception-related errors more effectively. Inspired by dropout regularization, we introduce uncertainty-guided token dropout, which applies the dropout principle to input visual tokens instead of model parameters, and during inference rather than training. By aggregating predictions from an ensemble of masked decoding contexts, Dropout Decoding robustly mitigates errors arising from visual token misinterpretations. Evaluations on benchmarks including CHAIR, THRONE, and MMBench demonstrate that Dropout Decoding significantly reduces object hallucinations (OH) and enhances both reliability and quality of LVLM outputs across diverse visual contexts.

Problem

• Reduces object hallucinations in vision-language models
• Improves reliability of multimodal model outputs
• Mitigates visual token misinterpretations via uncertainty-guided dropout

Innovation

• Uncertainty-guided dropout masks uncertain visual tokens
• Measures epistemic uncertainty by projecting tokens to text space
• Aggregates predictions from ensemble of masked decoding contexts