Source-linked AI summary

Align before Fuse: Vision and Language Representation Learning with Momentum Distillation

Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, Steven Hoi

arXiv:2107.07651v2cs.CVcs.AI

TL;DR

Existing vision-language pretraining struggles with misaligned image and text representations, costly detector-based features, and noisy web supervision. ALBEF aligns modalities before fusion and adds momentum distillation, achieving strong performance across downstream tasks, including 2.37% and 3.84% absolute gains on VQA and NLVR2 versus VILLA.

  • Problem

    Existing VLP methods face misaligned modalities, detector and high-resolution image costs, and noisy web labels that can penalize valid predictions.

  • Method

    ALBEF aligns independently encoded image and text representations with contrastive learning before multimodal fusion, while momentum distillation supplies pseudo-targets from a moving-average model.

  • Results

    2.37% and 3.84% absolute improvements are reported on VQA and NLVR2 versus VILLA, with faster inference and strong performance across multiple downstream tasks.

  • Takeaways & Limitations

    ALBEF provides a detector-free framework that combines aligned unimodal and multimodal representations for retrieval and reasoning tasks.

  • Takeaways & Limitations

    The authors state that further data and model analysis is needed before deployment because web data may contain private information, unsuitable images, or harmful texts.

Abstract

from arXiv · show

Large-scale vision and language representation learning has shown promising improvements on various vision-language tasks. Most existing methods employ a transformer-based multimodal encoder to jointly model visual tokens (region-based image features) and word tokens. Because the visual tokens and word tokens are unaligned, it is challenging for the multimodal encoder to learn image-text interactions. In this paper, we introduce a contrastive loss to ALign the image and text representations BEfore Fusing (ALBEF) them through cross-modal attention, which enables more grounded vision and language representation learning. Unlike most existing methods, our method does not require bounding box annotations nor high-resolution images. In order to improve learning from noisy web data, we propose momentum distillation, a self-training method which learns from pseudo-targets produced by a momentum model. We provide a theoretical analysis of ALBEF from a mutual information maximization perspective, showing that different training tasks can be interpreted as different ways to generate views for an image-text pair. ALBEF achieves state-of-the-art performance on multiple downstream vision-language tasks. On image-text retrieval, ALBEF outperforms methods that are pre-trained on orders of magnitude larger datasets. On VQA and NLVR$^2$, ALBEF achieves absolute improvements of 2.37% and 3.84% compared to the state-of-the-art, while enjoying faster inference speed. Code and pre-trained models are available at https://github.com/salesforce/ALBEF/.

1 Introduction

ALBEF addresses misaligned image-text representations, detector and resolution costs, and noisy web supervision by aligning modalities before fusion and using momentum distillation. It achieves strong results across downstream vision-language tasks, including retrieval, VQA, and NLVR2.

  • Existing VLP methods fuse detector-based image features with word tokens, but their representations occupy separate spaces, hindering image-text interaction learning.These methods commonly train multimodal encoders with masked language modeling and image-text matching.
  • ALBEF independently encodes images and text, applies image-text contrastive learning before cross-modal fusion, and uses a detector-free image encoder.The contrastive objective aligns modalities and supports a shared low-dimensional embedding space for harder negative mining.
  • Momentum distillation uses a moving-average model to generate pseudo-targets, reducing penalties for reasonable outputs that differ from noisy web annotations.The method improves pre-training and downstream tasks with clean annotations.
  • ALBEF interprets ITC and MLM as mutual-information lower-bound objectives over different views of image-text pairs, with momentum distillation generating semantically similar views.The resulting representations are described as invariant to semantic-preserving transformations.
  • 2.37% and 3.84% absolute improvements are reported on VQA and NLVR2 versus VILLA, alongside faster inference speed and strong retrieval performance.On image-text retrieval, ALBEF outperforms methods pretrained on orders of magnitude larger datasets.

2 Related Work

ALBEF combines strong unimodal and multimodal representation learning, targeting both retrieval and reasoning while avoiding object detectors. Its momentum distillation is framed as online self-distillation and as a broadly applicable learning algorithm for vision-language tasks.

  • Multimodal representation learning: Transformer-based multimodal methods model image-text interactions effectively but commonly require high-resolution images and pretrained object detectors.
  • ALBEF: ALBEF unifies unimodal and multimodal representation learning, achieving strong performance on both retrieval and reasoning tasks without object detectors.This removes a major computational bottleneck in many existing methods.
  • Knowledge distillation: Momentum distillation is interpreted as online self-distillation, using a temporal ensemble of the student model as its teacher.The paper distinguishes this approach from methods that distill from a fixed pretrained teacher and reports benefits across many vision-language tasks.

3 ALBEF Pre-training

ALBEF pre-trains separate image and text encoders before multimodal fusion, combining contrastive, matching, and masked-language objectives with momentum-generated pseudo-targets to handle noisy web supervision.

  • Model architecture: ALBEF uses image, text, and multimodal encoders, with cross-modal attention fusing image features and text features.The image and text encoders independently produce representations before the multimodal encoder models their interactions.
  • Pre-training objectives: ALBEF pre-trains with ITC on unimodal encoders and MLM and ITM on the multimodal encoder.The full objective is L = L_itc + L_mlm + L_itm, with ITM enhanced by online contrastive hard-negative mining.
  • Image-text contrastive learning: Image-text contrastive learning aligns unimodal representations before fusion, learns a shared 256-d space, and supports contrastive hard-negative mining for image-text matching.The similarity function compares normalized projected [CLS] embeddings, while in-batch negatives are sampled according to contrastive similarity.
  • Pre-training objectives: MLM masks 15% of input tokens and predicts them using both the image and contextual text, while ITM classifies image-text pairs as matched or unmatched.ITM uses the multimodal [CLS] representation and a two-class softmax prediction.
  • Momentum distillation: Momentum distillation uses a continuously evolving moving-average teacher to generate pseudo-targets that supervise the base model alongside standard losses.The method addresses noisy web pairs whose one-hot labels can penalize alternative correct matches or words, and applies a weighted KL-divergence term with α = 0.4.

4 A Mutual Information Maximization Perspective

ALBEF interprets its training objectives as maximizing mutual information between different views of an image-text pair. ITC and MLM use partial views, while momentum distillation generates semantically similar alternative views.

  • ITC, MLM, and momentum distillation are framed as different ways to generate views of an image-text pair for mutual-information maximization.The goal is to learn representations invariant to changes in view.
  • ITC treats the image and text modalities as two views and maximizes their mutual information for positive pairs through a symmetric InfoNCE objective.The loss trains unimodal encoders to associate matching image-text representations.
  • MLM generates views by pairing a randomly selected word token with the image and text context in which that word is masked.This corresponds to maximizing mutual information between the masked token and its masked context.
  • Momentum distillation generates alternative views from the proposal distribution, including semantically similar texts or images and alternative masked-word targets.These views act as data augmentation beyond the original image-text pairs.

5 Downstream V+L Tasks

The paper adapts ALBEF to retrieval, visual entailment, VQA, NLVR2, and weakly supervised visual grounding with task-specific fine-tuning strategies. VQA uses answer generation, while NLVR2 modifies the multimodal encoder for two-image reasoning.

  • Downstream Tasks: ALBEF is adapted to five downstream vision-language tasks using task-specific fine-tuning strategies.The tasks include image-text retrieval, visual entailment, VQA, NLVR2, and weakly supervised visual grounding.
  • Image-Text Retrieval: Image-text retrieval evaluates image-to-text and text-to-image retrieval on Flickr30K and COCO while jointly optimizing ITC and ITM losses.ITC provides an image-text scoring function based on unimodal representation similarity.
  • Visual Entailment: Visual entailment predicts whether an image-text relationship is entailment, neutral, or contradiction using the multimodal encoder’s [CLS] representation.The task is formulated as three-way classification.
  • Visual Question Answering: VQA formulates answer prediction as generation with a six-layer autoregressive transformer decoder receiving multimodal embeddings through cross-attention.The decoder is initialized from the multimodal encoder and trained with conditional language modeling.
  • Caveat: SNLI-VE results should be interpreted cautiously because the test data has been reported to be noisy.
  • Natural Language for Visual Reasoning: NLVR2 extends the multimodal encoder with replicated transformer blocks so it can reason over a pair of images.An additional text-assignment pre-training task assigns text to the first image, second image, or neither.

6 Experiments

ALBEF's ablations and benchmark evaluations show that ITC, hard-negative mining, and momentum distillation improve performance across vision-language tasks. The detector-free model achieves state-of-the-art retrieval, reasoning, and weakly supervised grounding results, with faster inference.

  • Evaluation of the Proposed Methods: Adding ITC to MLM+ITM substantially improves the pre-trained model across all evaluated tasks.
  • Evaluation of the Proposed Methods: Momentum distillation improves ITC, MLM, and all downstream tasks while enabling effective use of noisier web data.
  • Image-Text Retrieval: ALBEF achieves state-of-the-art fine-tuned and zero-shot image-text retrieval, outperforming CLIP and ALIGN despite their much larger pre-training datasets.
  • VQA, NLVR, and VE: 2.37% on VQA test-std, 3.84% on NLVR2 test-P, and 1.88% on SNLI-VE test are ALBEF's absolute improvements over VILLA.
  • Weakly-supervised Visual Grounding: ALBEF substantially outperforms existing methods on weakly supervised RefCOCO+ visual grounding without bounding-box annotations.
  • Weakly-supervised Visual Grounding: Grad-CAM visualizations indicate grounding of objects, attributes, and relationships, while NLVR2 ranking remains generally insensitive to the top-k candidate count.

7 Conclusion and Social Impacts

The paper concludes that ALBEF combines strong performance with faster inference across multiple vision-language tasks. It also cautions that further analysis is needed before deployment because web data and accuracy-only optimization may create social risks.

  • Conclusion: ALBEF combines image-text contrastive learning and momentum distillation, achieving better performance and faster inference across multiple downstream tasks.
  • Social Impacts: Deployment requires additional data and model analysis because web data may contain private information, unsuitable images, or harmful texts.
  • Social Impacts: Optimizing accuracy alone may have unwanted social implications, limiting the paper's deployment conclusions.

A Downstream Task Details

Downstream evaluations use standardized fine-tuning settings across tasks, with task-specific datasets and procedures. Visual grounding additionally uses Grad-CAM-based patch importance from image or multimodal attention.

  • Fine-tuning setup: All downstream tasks use 384 × 384 input images, resizing without cropping during inference.The implementation reuses RandAugment, AdamW, cosine decay, weight decay, and distillation weight from pre-training.
  • Image-Text Retrieval: Image-text retrieval is evaluated on COCO and Flickr30K using the Karpathy split and 10-epoch fine-tuning.COCO and Flickr30K use batch size 256 and an initial learning rate of 1e−5.
  • Visual Entailment: Visual entailment is evaluated on SNLI-VE with the original split and 5-epoch fine-tuning.The task uses batch size 256 and an initial learning rate of 2e−5.
  • NLVR2: NLVR2 uses an additional text-assignment pre-training stage because each example contains two input images.Text-assignment pre-training runs for 1 epoch on the 4M dataset with 256 × 256 images.
  • Visual Grounding: Visual grounding on RefCOCO+ computes patch importance with Grad-CAM from either visual self-attention or multimodal cross-attention.For ITM, Grad-CAM uses the third multimodal layer and averages across heads and input text tokens.

B Additional Per-word Visualizations

ALBEF’s per-word Grad-CAM visualizations are used to examine fine-grained visual grounding. The accompanying figure presents Grad-CAM visualizations on cross-attention maps for individual words.

  • Additional Per-word Visualizations: Per-word Grad-CAM visualizations demonstrate grounding of objects, actions, attributes, and relationships.The visualizations are presented as additional evidence of the model’s visual-grounding ability.
  • Additional Per-word Visualizations: Figure 10 compares human attention with ALBEF’s Grad-CAM for VQA.

C Comparison with Human Attention

The paper compares ALBEF’s VQA Grad-CAM visualizations with human attention maps and illustrates pseudo-targets generated by momentum models for contrastive and masked-language objectives.

  • Comparison with Human Attention: Human attention maps are compared with ALBEF VQA Grad-CAM on 1,374 validation question-image pairs using rank correlation.The Grad-CAM is computed at the third multimodal cross-attention layer.
  • Momentum-model examples: Pseudo-target examples include semantically related descriptions for demonstrations, rock-star performances, and snowy houses with fences.
  • Momentum-model examples: The examples show top-5 similar texts selected by the momentum model for image-text contrastive learning.
  • Momentum-model examples: The examples show top-5 words generated by the momentum model for masked language modeling.

E Pre-training Dataset Details

The paper reports statistics for the image and text components of its pre-training datasets in Table 8.

  • Pre-training Dataset Details: Table 8 provides statistics for the image data used in pre-training.
  • Pre-training Dataset Details: Table 8 provides statistics for the text data used in pre-training.
  • Pre-training Dataset Details: The pre-training dataset statistics are organized together in Table 8.
Loading 2107.07651v2…