AdaFlow: Non-blocking Inference with Heterogeneous Multi-modal Mobile Sensor Data

Multi-modal deep learning offer conceptual advantages by integrating information from different vantage points. However, data flow blocking and corruption due to modalities asymmetric which make exist methods can hardly balance latency and accuracy still pose many challenges. To address these challenges, this paper proposes AdaFlow to establish a multi-modal mapping mechanism using an affinity matrix, achieving non-blocking data flow in multi-modal systems.