DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models

• URL: http://arxiv.org/abs/2405.20985v1
• Date: Fri, 31 May 2024 16:31:38 GMT
• Title: DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models
• Authors: Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, Lu Hou,
• Abstract summary: This study examines the projector module by interpreting the vision-language semantic flow within MLLMs. Our findings reveal that compressive projectors abstract visual patches into a limited set of semantic concepts, such as objects or attributes, resulting in a 'double abstraction' phenomenon. We propose the key insight of 'Decouple Compression from Abstraction (DeCo) that is compressing the visual token number at the patch level by projectors.
• Score: 28.019592576500113
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: The visual projector, which bridges the vision and language modalities and facilitates cross-modal alignment, serves as a crucial component in MLLMs. However, measuring the effectiveness of projectors in vision-language alignment remains under-explored, which currently can only be inferred from the performance of MLLMs on downstream tasks. Motivated by the problem, this study examines the projector module by interpreting the vision-language semantic flow within MLLMs. Specifically, we trace back the semantic relevance flow from generated language tokens to raw visual encoder patches and the intermediate outputs produced by projectors. Our findings reveal that compressive projectors (e.g., QFormer), abstract visual patches into a limited set of semantic concepts, such as objects or attributes, resulting in a 'double abstraction' phenomenon. This involves a first visual semantic abstraction by the projector referring to pre-defined query tokens, and a second extraction by the LLM based on text instructions. The double abstraction is inefficient in training and will result in cumulative vision semantics deficiency. To mitigate this issue, we propose the key insight of 'Decouple Compression from Abstraction (DeCo), that is compressing the visual token number at the patch level by projectors and allowing the LLM to handle visual semantic abstraction entirely. Consequently, we adopt a simple compressor, i.e., 2D Adaptive Pooling, to downsample visual patches in a parameter-free manner. Empirical evaluation demonstrates that DeCo surpasses traditional compressive projectors regarding both performance and efficiency. It achieves performance gains of 0.9%, 7.1%, and 2.9% across the MLLM Benchmarks, Visual Localization, and Open-ended VQA tasks with fewer trainable parameters and faster convergence speed.

Related papers

• SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs [74.2538340966038]
  We investigate how Multimodal Large Language Models (MLLMs) process visual inputs by analyzing their attention mechanisms.<n>We reveal a surprising sparsity phenomenon: only a small subset of attention heads in LLMs actively contribute to visual understanding.<n>We introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores.
  arXiv  Detail & Related papers  (2025-06-05T17:59:55Z)
• Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models [21.197083685420584]
  We show the role of the projector in compressing vision embeddings and aligning them with word embeddings.<n>We then examine patch-level alignment -- the alignment between each vision patch and its corresponding semantic words.<n>Our experiments show that patch-aligned training achieves stronger compression capability and improved patch-level alignment.
  arXiv  Detail & Related papers  (2025-05-22T22:10:27Z)
• Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings [69.35226485836641]
  Excessive use of visual tokens in existing Multimoal Large Language Models (MLLMs) often exhibits obvious redundancy and brings in prohibitively expensive computation. We propose a simple yet effective method to improve the efficiency of MLLMs, termed dynamic visual-token exit (DyVTE) DyVTE uses lightweight hyper-networks to perceive the text token status and decide the removal of all visual tokens after a certain layer.
  arXiv  Detail & Related papers  (2024-11-29T11:24:23Z)
• Efficient Multi-modal Large Language Models via Visual Token Grouping [55.482198808206284]
  High-resolution images and videos pose a barrier to their broader adoption. compressing vision tokens in MLLMs has emerged as a promising approach to reduce inference costs. We introduce VisToG, a novel grouping mechanism that leverages the capabilities of pre-trained vision encoders to group similar image segments.
  arXiv  Detail & Related papers  (2024-11-26T09:36:02Z)
• Spatial-Aware Efficient Projector for MLLMs via Multi-Layer Feature Aggregation [10.468784974994465]
  The projector plays a crucial role in multi-modal language models (MLLMs) Current explorations on the projector focus on reducing the number of visual tokens to improve efficiency. A Spatial-Aware Efficient Projector (SAEP) is proposed to address this issue.
  arXiv  Detail & Related papers  (2024-10-14T09:25:09Z)
• Treat Visual Tokens as Text? But Your MLLM Only Needs Fewer Efforts to See [37.7015406019386]
  Multimodal Large Language Models (MLLMs) treat visual tokens from visual encoders as text tokens. As token counts grow, the quadratic scaling of computation in LLMs introduces an efficiency bottleneck. In this study, we investigate the redundancy in visual computation at both the parameter and computational pattern levels within LLaVA.
  arXiv  Detail & Related papers  (2024-10-08T16:13:24Z)
• Balancing Performance and Efficiency: A Multimodal Large Language Model Pruning Method based Image Text Interaction [6.467840081978855]
  multimodal large language models (MM-LLMs) have achieved great success in many multimodal tasks, but their high computational costs limit their further promotion and application. We studied the visual tokens of MM-LLMs and designed a dynamic pruning algorithm to address this issue. Our proposed method can achieve performance that competes with the original performance when using an average of 22% of the original token quantity.
  arXiv  Detail & Related papers  (2024-09-02T10:49:10Z)
• VideoLLM-MoD: Efficient Video-Language Streaming with Mixture-of-Depths Vision Computation [66.00245701441547]
  We introduce a novel approach to reduce vision compute by leveraging redundant vision tokens "skipping layers" rather than decreasing the number of vision tokens. Our method, VideoLLM-MoD, is inspired by mixture-of-depths LLMs and addresses the challenge of numerous vision tokens in long-term or streaming video.
  arXiv  Detail & Related papers  (2024-08-29T17:21:58Z)
• Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders [89.38717274524681]
  This study explores the design space for multimodal large language models (MLLMs) using a mixture of vision encoders and resolutions. Our findings reveal several underlying principles common to various existing strategies, leading to a streamlined yet effective design approach. The resulting family of MLLMs, Eagle, surpasses other leading open-source models on major MLLM benchmarks.
  arXiv  Detail & Related papers  (2024-08-28T17:59:31Z)
• TokenPacker: Efficient Visual Projector for Multimodal LLM [37.1071749188282]
  The visual projector serves as an essential bridge between the visual encoder and the Large Language Model (LLM) We propose a novel visual projector, which adopts a coarse-to-fine scheme to inject the enriched characteristics to generate the condensed visual tokens. Our approach compresses the visual tokens by 75%89%, while achieves comparable or even better performance across diverse benchmarks.
  arXiv  Detail & Related papers  (2024-07-02T16:10:55Z)
• Investigating Video Reasoning Capability of Large Language Models with Tropes in Movies [69.28082193942991]
  This paper introduces a novel dataset, Tropes in Movies (TiM), designed as a testbed for exploring two critical yet previously overlooked video reasoning skills. utilizing tropes from movie storytelling, TiM evaluates the reasoning capabilities of state-of-the-art LLM-based approaches. To address these deficiencies, we propose Face-Enhanced Viper of Role Interactions (FEVoRI) and Context Query Reduction (ConQueR)
  arXiv  Detail & Related papers  (2024-06-16T12:58:31Z)
• Towards Semantic Equivalence of Tokenization in Multimodal LLM [149.11720372278273]
  Vision tokenization is essential for semantic alignment between vision and language. This paper proposes a novel dynamic Semantic-Equivalent Vision Tokenizer (SeTok) SeTok groups visual features into semantic units via a dynamic clustering algorithm. The resulting vision tokens effectively preserve semantic integrity and capture both low-frequency and high-frequency visual features.
  arXiv  Detail & Related papers  (2024-06-07T17:55:43Z)
• Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference [59.91176945361035]
  We introduce Visual Tokens Withdrawal (VTW), a plug-and-play module to boost MLLMs for rapid inference. Our approach is inspired by two intriguing phenomena we have observed. Our VTW approach can cut computational overhead by over 40% across diverse multimodal tasks while maintaining performance.
  arXiv  Detail & Related papers  (2024-05-09T14:38:53Z)
• VURF: A General-purpose Reasoning and Self-refinement Framework for Video Understanding [65.12464615430036]
  This paper introduces a Video Understanding and Reasoning Framework (VURF) based on the reasoning power of Large Language Models (LLMs) Ours is a novel approach to extend the utility of LLMs in the context of video tasks. We harness their contextual learning capabilities to generate executable visual programs for video understanding.
  arXiv  Detail & Related papers  (2024-03-21T18:00:00Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.