AdaBridge: Dynamic Data and Computation Reuse for Efficient Multi-task DNN Co-evolution in Edge Systems

TL;DR

This work addresses efficient co-evolution of multi-task DNNs on resource-limited edge devices facing data drift. It introduces AdaBridge, which combines a reuse-friendly data resampling module with an asynchronous multi-task retraining scheduler to enable dynamic data and computation reuse across tasks. The approach employs an adapter-based multi-task learning framework and memory I/O-aware data reorganization to share features and reduce overhead. Empirical results show AdaBridge yields an 11% improvement in the lowest accuracy under data drift, demonstrating enhanced generalization and efficiency for edge-enabled multi-task DNN evolution.

Abstract

Running multi-task DNNs on mobiles is an emerging trend for various applications like autonomous driving and mobile NLP. Mobile DNNs are often compressed to fit the limited resources and thus suffer from degraded accuracy and generalizability due to data drift. DNN evolution, e.g., continuous learning and domain adaptation, has been demonstrated effective in overcoming these issues, mostly for single-task DNN, leaving multi-task DNN evolution an important yet open challenge. To fill up this gap, we propose AdaBridge, which exploits computational redundancies in multi-task DNNs as a unique opportunity for dynamic data and computation reuse, thereby improving training efficacy and resource efficiency among asynchronous multi-task co-evolution in edge systems. Experimental evaluation shows that AdaBridge achieves 11% average accuracy gain upon individual evolution baselines.

Figures (2)

• Figure 1: Illustration of AdaBridge's system workflow.
• Figure 2: Comparison of accuracy fluctuation.