Adaptive Multi-Modal Learning Framework for Streaming Big Data: A Dynamic Information- Theoretic and Distributed Optimization Approach

The massive increase in both heterogeneity and velocity of data streams has raised the need for learning systems with adaptivity to evolving distributions at scale. In this paper, we propose the Adaptive Multi-Modal Learning (AMML) framework to provide a one-stop solution for analyzing non-stationary multi-modal big data in real-time. The AMML combines modality-specific representation learning with modelling the drift by an information-theoretic measure based on KL divergence to track distributional changes. By combining a temporally regularized optimization framework with variational information constraints, the model can balance an optimal trade-off between stability and flexibility under continuous updates. The framework is deployed in a distributed Spark–Hadoop environment for large-scale deployment with scalability, low latency and resilient data recovery. For a convex online surrogate of AMML, the theoretical analysis provides a path-variation-dependent dynamic-regret bound. Vanishing average dynamic regret is obtained when the cumulative variation of the time-varying comparator grows sufficiently slowly, which implies it is capable of tracking the changing optimal predictors. Empirical results on simulated and synthetic multimodal streaming data demonstrate that AMML achieves higher accuracy, faster drift recovery, smoother parameter evolution, and near-linear scalability than existing streaming learning algorithms.