Hallucination Augmented Contrastive Learning for Multimodal Large Language Model (2312.06968v4)

Abstract: Multi-modal LLMs (MLLMs) have been shown to efficiently integrate natural language with visual information to handle multi-modal tasks. However, MLLMs still face a fundamental limitation of hallucinations, where they tend to generate erroneous or fabricated information. In this paper, we address hallucinations in MLLMs from a novel perspective of representation learning. We first analyzed the representation distribution of textual and visual tokens in MLLM, revealing two important findings: 1) there is a significant gap between textual and visual representations, indicating unsatisfactory cross-modal representation alignment; 2) representations of texts that contain and do not contain hallucinations are entangled, making it challenging to distinguish them. These two observations inspire us with a simple yet effective method to mitigate hallucinations. Specifically, we introduce contrastive learning into MLLMs and use text with hallucination as hard negative examples, naturally bringing representations of non-hallucinative text and visual samples closer while pushing way representations of non-hallucinating and hallucinative text. We evaluate our method quantitatively and qualitatively, showing its effectiveness in reducing hallucination occurrences and improving performance across multiple benchmarks. On the MMhal-Bench benchmark, our method obtains a 34.66% /29.5% improvement over the baseline MiniGPT-4/LLaVA. Our code is available on https://github.com/X-PLUG/mPLUG-HalOwl/tree/main/hacl.

An Analysis of "Hallucination Augmented Contrastive Learning for Multimodal LLMs"

The paper under review addresses a critical challenge in the deployment of Multimodal LLMs (MLLMs): the generation of hallucinations. Hallucinations refer to instances where models generate erroneous or fictitious information, especially when multimodal input, such as text and image data, is involved. This paper introduces a novel approach titled "Hallucination Augmented Contrastive Learning" (HACL), aimed at mitigating hallucinations by improving the alignment of visual and textual representations within MLLMs.

Core Contributions

The authors identify two primary issues within current MLLMs: a significant gap between textual and visual representations and an entanglement of representations between hallucinated and non-hallucinated texts. To address this, the authors propose integrating contrastive learning into MLLMs, where text containing hallucinations serves as hard negative examples. This method aims to bring representations of non-hallucinative texts closer to their related visual samples and distance them from hallucinative texts, effectively reducing the frequency of hallucinations.

Methodology

The proposed methodology employs a two-stage training process:

1. Initial Pre-training Stage: Here, the model learns to generate visual tokens through a vision encoder and aligns these with text tokens using contrastive learning. By appending <EOS> tokens to both inputs, the method obtains global representations of visual and textual sequences.
2. Hallucination Augmented Contrastive Learning (HACL): The authors enhance the alignment between visual and textual representations by incorporating hallucinative captions. These captions are generated using GPT-4 to introduce elements that either deviate from or are entirely absent in the visual data. Through this augmentation, the model is better equipped to distinguish between hallucinated and genuine textual data, utilizing a hard negative sampling strategy.

Empirical Results

• The proposed HACL method demonstrated significant quantitative improvements across several benchmarks such as MMhal-Bench and POPE, with notable performance enhancements in reducing hallucination rates. For instance, LLaVA equipped with HACL improved its overall score on the MMhal-Bench benchmark by 29.5% over previous baselines.
• The authors also observed enhanced performance in non-hallucination-focused tasks (i.e., general visual question answering) across various datasets, suggesting that improved representation alignment positively influences the model's overall comprehension capabilities.

Practical and Theoretical Implications

From a practical standpoint, the reduction in hallucination rates allows for more reliable deployment of MLLMs in real-world applications. Theoretical implications of this work underscore the importance of representation learning in bridging modality gaps and provide a framework for future research aimed at reducing errors in AI-generated content.

Speculations on Future Developments

The methodology presents a scalable approach to enhancing multimodal understanding, suggesting a direction for future research to explore additional modalities (e.g., audio or video) and their alignment with textual and visual data. Furthermore, adaptive learning strategies could be developed to dynamically adjust the alignment process based on evolving model requirements.

Overall, this paper introduces a promising method for mitigating hallucinations in MLLMs, providing a foundation for more effective and reliable AI systems in multimodal applications.