An official website of the United States government
In multimodal machine learning, effectively addressing the missing modality scenario is crucial for improving performance in downstream tasks such as in medical contexts where data may be incomplete. Although some attempts have been made to retrieve embeddings for missing modalities, two main bottlenecks remain: (1) the need to consider both intra- and inter-modal context, and (2) the cost of embedding selection, where embeddings often lack modality-specific knowledge. To address this, the authors propose MoE-Retriever, a novel framework inspired by Sparse Mixture of Experts (SMoE). MoE-Retriever defines a supporting group for intra-modal inputs—samples that commonly lack the target modality—by selecting samples with complementary modality combinations for the target modality. This group is integrated with inter-modal inputs from different modalities of the same sample, establishing both intra- and inter-modal contexts. These inputs are processed by Multi-Head Attention to generate context-aware embeddings, which serve as inputs to the SMoE Router that automatically selects the most relevant experts (embedding candidates). Comprehensive experiments on both medical and general multimodal datasets demonstrate the robustness and generalizability of MoE-Retriever, marking a significant step forward in embedding retrieval methods for incomplete multimodal data.
more »
« less
This content will become publicly available on February 5, 2026
Sparse MoE as a New Retriever: Addressing Missing Modality Problem in Incomplete Multimodal Data
In multimodal machine learning, effectively addressing the missing modality scenario is crucial for improving performance in downstream tasks such as in medical contexts where data may be incomplete. Although some attempts have been made to effectively retrieve embeddings for missing modalities, two main bottlenecks remain: the consideration of both intra- and inter-modal context, and the cost of embedding selection, where embeddings often lack modality-specific knowledge. In response, we propose MoE-Retriever, a novel framework inspired by the design principles of Sparse Mixture of Experts (SMoE). First, MoE-Retriever samples the relevant data from modality combinations, using a so-called supporting group to construct intra-modal inputs while incorporating inter-modal inputs. These inputs are then processed by Multi-Head Attention, after which the SMoE Router automatically selects the most relevant expert, i.e., the embedding candidate to be retrieved. Comprehensive experiments on both medical and general multimodal datasets demonstrate the robustness and generalizability of MoE-Retriever, marking a significant step forward in embedding retrieval methods for incomplete multimodal data.
more »
« less
- Award ID(s):
- 2505865
- PAR ID:
- 10631052
- Publisher / Repository:
- ICLR 2025 https://openreview.net/forum?id=j9DbobO0mY
- Date Published:
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Learning multimodal representations is a fundamentally complex research problem due to the presence of multiple heterogeneous sources of information. Although the presence of multiple modalities provides additional valuable information, there are two key challenges to address when learning from multimodal data: 1) models must learn the complex intra-modal and cross-modal interactions for prediction and 2) models must be robust to unexpected missing or noisy modalities during testing. In this paper, we propose to optimize for a joint generative-discriminative objective across multimodal data and labels. We introduce a model that factorizes representations into two sets of independent factors: multimodal discriminative and modality-specific generative factors. Multimodal discriminative factors are shared across all modalities and contain joint multimodal features required for discriminative tasks such as sentiment prediction. Modality-specific generative factors are unique for each modality and contain the information required for generating data. Experimental results show that our model is able to learn meaningful multimodal representations that achieve state-of-the-art or competitive performance on six multimodal datasets. Our model demonstrates flexible generative capabilities by conditioning on independent factors and can reconstruct missing modalities without significantly impacting performance. Lastly, we interpret our factorized representations to understand the interactions that influence multimodal learning.more » « less
-
We introduce caption-guided face recognition (CGFR) as a new framework to improve the performance of commercial-off-the-shelf (COTS) face recognition (FR) systems. In contrast to combining soft biometrics (e.g., facial marks, gender, and age) with face images, in this work, we use facial descriptions provided by face examiners as a piece of auxiliary information. However, due to the heterogeneity of the modalities, improving the performance by directly fusing the textual and facial features is very challenging, as both lie in different embedding spaces. In this paper, we propose a contextual feature aggregation module (CFAM) that addresses this issue by effectively exploiting the fine-grained word-region interaction and global image-caption association. Specifically, CFAM adopts a self-attention and a cross-attention scheme for improving the intra-modality and inter-modality relationship between the image and textual features, respectively. Additionally, we design a textual feature refinement module (TFRM) that refines the textual features of the pre-trained BERT encoder by updating the contextual embeddings. This module enhances the discriminative power of textual features with a cross-modal projection loss and realigns the word and caption embeddings with visual features by incorporating a visual-semantic alignment loss. We implemented the proposed CGFR framework on two face recognition models (ArcFace and AdaFace) and evaluated its performance on the Multi-Modal CelebA-HQ dataset. Our framework significantly improves the performance of ArcFace in both 1:1 verification and 1:N identification protocol.more » « less
-
Multimodal medical image synthesis is an important task. Previous efforts mainly focus on the task domain of medical image synthesis using the complete source data and have achieved great success. However, data collection with completeness in real life might be prohibitive due to high expenses or other difficulties, particularly in brain imaging studies. In this paper, we address the challenging and important problem of medical image synthesis from incomplete multimodal data sources. We propose to learn the modal-wise representations and synthesize the targets accordingly. Particularly, a surrogate sampler is derived to generate the target representations from incomplete observations, based on which an interpretable attention-redistribution network is designed. The experimental results synthesizing PET images from MRI images demonstrate that the proposed method can solve different missing data scenarios and outperforms related baselines consistently.more » « less
-
Proteins, often represented as multi-modal data of 1D sequences and 2D/3D structures, provide a motivating example for the communities of machine learning and computational biology to advance multi-modal representation learning. Protein language models over sequences and geometric deep learning over structures learn excellent single-modality representations for downstream tasks. It is thus desirable to fuse the single-modality models for better representation learning, but it remains an open question on how to fuse them effectively into multi-modal representation learning with a modest computational cost yet significant downstream performance gain. To answer the question, we propose to make use of separately pretrained single-modality models, integrate them in parallel connections, and continuously pretrain them end-to-end under the framework of multimodal contrastive learning. The technical challenge is to construct views for both intra- and inter-modality contrasts while addressing the heterogeneity of various modalities, particularly various levels of semantic robustness. We address the challenge by using domain knowledge of protein homology to inform the design of positive views, specifically protein classifications of families (based on similarities in sequences) and superfamilies (based on similarities in structures). We also assess the use of such views compared to, together with, and composed to other positive views such as identity and cropping. Extensive experiments on enzyme classification and protein function prediction benchmarks demonstrate the potential of domain-informed view construction and combination in multi-modal contrastive learningmore » « less
