Co-Learning for Missing Arbitrary Modalities in Multi-modal Classification

AI summary

Overview

Research area: Multi-modal machine learning, specifically robustness to missing data modalities at inference time, with experiments in Earth Observation remote sensing (crop-type recognition) and wearable-sensor human activity recognition.

Technical level: Advanced — the paper builds on contrastive representation learning, knowledge distillation and multi-modal co-learning, and its central artifacts are four interacting loss functions.

Scope: The paper proposes two co-learning methods, Co-Miss and FullCo, that combine feature-level collaboration with decision-level distillation so that multi-modal classifiers keep working when any arbitrary subset of their training modalities is unavailable at test time.

What This Paper Is About

Multi-modal classifiers combine several data sources (for example optical satellite imagery, radar, weather variables and topography) to improve predictions, but real deployments rarely have all sources available at once — sensors fail, clouds block optical imagery, missions end, or privacy rules block a data stream. Most prior work handles this only in two-modality settings, or assumes a fixed missing pattern, so the paper instead targets missing arbitrary modalities: any subset of the training modalities may be absent at inference. The goal is a training framework that stays accurate both when a single modality is missing and when all but one are missing, without knowing in advance which case will occur.

Key Contributions

1. A general problem formulation — the paper defines "missing arbitrary modalities" as any subset of the M training modalities being available at inference, spanning minimal (one modality missing), moderate and extreme (all-but-one missing) conditions, rather than assuming predefined missing patterns or bimodal setups.

2. Two co-learning methods built on decision-level fusion — Co-Miss and FullCo both learn modality-shared and modality-specific feature spaces with a volume-based contrastive loss and a modality discriminant classifier, while per-modality predictions are combined by simple averaging (the "

Authors’ abstract

Multi-modal classification leverages complementary information across diverse data sources to enhance predictive performance. However, real-world scenarios subject to operational constraints, such as sensor failures or privacy restrictions, lead to inconsistent modality availability between training and inference times. To handle missing modalities, prior studies have mainly covered bimodal data setups and focused on designing robust fusion processes. Instead, we adopt a multi-modal co-learning framework that prioritizes inter-modal collaboration rather than multi-modal fusion. Specifically, we consider that any subset of modalities may be absent, without assuming predefined missing-modality patterns, an inference scenario we refer to as missing arbitrary modalities. To address this challenge, we introduce two alternative approaches that leverage information at both feature- and decision-level. Experiments on two multi-modal classification benchmarks demonstrate significant robustness gains in various missing modality conditions. The first method shows more robust behavior under minimal missing conditions, where a single modality is absent, whereas the second performs better under extreme missing conditions, where all-but-one modalities are missing. Our code is available at https://github.com/fmenat/Co4Miss.