Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces SPICE, a novel framework for multimodal curriculum learning that adapts to model evolution by leveraging Partial Information Decomposition (PID) theory. By dynamically characterizing sample complexity through the decomposition of multimodal interactions into redundant, unique, and synergistic components, SPICE allows for a progressive curriculum that evolves throughout training. Experimental results across various multimodal benchmarks show that SPICE consistently outperforms traditional training methods and state-of-the-art baselines, underscoring the benefits of adaptive sample ordering informed by PID estimates.
SPICE reveals that adapting curriculum learning in real-time based on information decomposition can significantly enhance multimodal model performance.
Multimodal learning exploits complementary information across heterogeneous modalities. The informativeness of each modality can vary widely across samples and training stages. Existing multimodal curriculum learning strategies often assume that the relative complexity of samples remains unchanged throughout training and therefore cannot adapt to model evolution. We propose SPICE (Synergy and Partial Information based Curriculum Evolution), a novel progressive curriculum framework for multimodal interaction learning. Guided by Partial Information Decomposition (PID) theory, our approach decomposes multimodal interactions into redundant, unique, and synergistic information components, enabling an interpretable and dynamic characterization of sample complexity. Building on this decomposition, we design a progressive curriculum that evolves throughout training, allowing the model to transition from learning shared cross-modal cues to modality-specific patterns and, finally, to complex synergistic interactions. Adapting to model evolution, sample ordering is refined in real-time using PID information estimates derived from unimodal and multimodal predictions. Experiments across multiple multimodal benchmarks demonstrate consistent improvements over conventional training and state-of-the-art baselines, highlighting the effectiveness of PID information decomposition and adaptive sample ordering for multimodal curriculum learning.