ML Mule: Mobile-Driven Context-Aware Collaborative Learning

TL;DR

ML Mule introduces a mobility-driven, context-aware distributed learning framework where mobile devices act as mules to transport and update models between fixed devices in different spaces, enabling decoupled-in-time collaboration. The system uses two training phases (in-house and mule) and an adaptive model-freshness filter to govern aggregation, achieving faster convergence and higher accuracy than traditional federated and decentralized baselines on CIFAR-100 and EgoExo4D. Through extensive simulations with varied mobility patterns and a real-world-inspired prototype, ML Mule demonstrates robust performance under intermittent connectivity and non-IID data, while preserving privacy by avoiding centralized data collection. The work highlights potential for scalable, privacy-preserving, context-aware smart environments, with future directions including advanced aggregation, privacy techniques, and integration with FL ecosystems.

Abstract

Artificial intelligence has been integrated into nearly every aspect of daily life, powering applications from object detection with computer vision to large language models for writing emails and compact models for use in smart homes. These machine learning models at times cater to the needs of individual users but are often detached from them, as they are typically stored and processed in centralized data centers. This centralized approach raises privacy concerns, incurs high infrastructure costs, and struggles to provide real time, personalized experiences. Federated and fully decentralized learning methods have been proposed to address these issues, but they still depend on centralized servers or face slow convergence due to communication constraints. We propose ML Mule, an approach that utilizes individual mobile devices as 'mules' to train and transport model snapshots as the mules move through physical spaces, sharing these models with the physical 'spaces' the mules inhabit. This method implicitly forms affinity groups among devices associated with users who share particular spaces, enabling collaborative model evolution and protecting users' privacy. Our approach addresses several major shortcomings of traditional, federated, and fully decentralized learning systems. ML Mule represents a new class of machine learning methods that are more robust, distributed, and personalized, bringing the field closer to realizing the original vision of intelligent, adaptive, and genuinely context-aware smart environments. Our results show that ML Mule converges faster and achieves higher model accuracy compared to other existing methods.

Figures (11)

• Figure 1: Example of ML Mule sharing process
• Figure 2: Illustration of the two main training modes. In (a), the main training occurs on fixed devices $F$; in (b), training takes place on the mobile devices $M$.
• Figure 3: ICA decomposition on a sample of NYC Foursquare mobility data
• Figure 4: Example random-walk trajectories under three different crossing probabilities ($P_{cross} = \{0, 0.1, 0.5\}$). Each subfigure shows device movements in a 2D space partitioned into four spaces within two isolated areas. The black grid lines mark boundaries, while circles and crosses denote start and end points, respectively. Higher $P_{cross}$ values indicate greater likelihood of leaving the current space.
• Figure 5: CIFAR-100 Data distributions across different partitioning methods. The first subplot show IID distribution. Next three subplots illustrate Dirichlet-based distributions with $\alpha= \{0.001, 0.01, 0.1\}$. The last subplot shows our adapted Shards method, wherein super-classes are split between two areas (Area 0 and Area 1), and each space within an area contains exactly one subclass.