[Paper Review] DeepWear: Adaptive Local Offloading for On-Wearable Deep Learning

Mengwei Xu, Feng Qian+4 more|arXiv (Cornell University)|Dec 1, 2017

IoT and Edge/Fog Computing51 references4 citations

TL;DR

DeepWear is a lightweight, context-aware framework that enables adaptive local offloading of deep learning (DL) inference from wearable devices to paired handheld devices via Bluetooth, improving performance and reducing energy consumption. It achieves up to 5.08× speedup and 85.5% energy savings compared to local-only execution by using intelligent model partitioning, pipelining, and real-time workload awareness.

ABSTRACT

Due to their on-body and ubiquitous nature, wearables can generate a wide range of unique sensor data creating countless opportunities for deep learning tasks. We propose DeepWear, a deep learning (DL) framework for wearable devices to improve the performance and reduce the energy footprint. DeepWear strategically offloads DL tasks from a wearable device to its paired handheld device through local network. Compared to the remote-cloud-based offloading, DeepWear requires no Internet connectivity, consumes less energy, and is robust to privacy breach. DeepWear provides various novel techniques such as context-aware offloading, strategic model partition, and pipelining support to efficiently utilize the processing capacity from nearby paired handhelds. Deployed as a user-space library, DeepWear offers developer-friendly APIs that are as simple as those in traditional DL libraries such as TensorFlow. We have implemented DeepWear on the Android OS and evaluated it on COTS smartphones and smartwatches with real DL models. DeepWear brings up to 5.08X and 23.0X execution speedup, as well as 53.5% and 85.5% energy saving compared to wearable-only and handheld-only strategies, respectively.

Motivation & Objective

• Address the challenge of running computationally intensive deep learning workloads on resource-constrained commercial-off-the-shelf (COTS) wearable devices.
• Overcome limitations of cloud-based offloading by eliminating reliance on Internet connectivity and reducing privacy risks.
• Leverage the proximity and computational power of paired handheld devices (e.g., smartphones) to improve performance and energy efficiency.
• Develop a practical, developer-friendly framework that integrates seamlessly with existing deep learning workflows on Android.
• Enable transparent, adaptive offloading decisions based on real-time system and network conditions to balance latency, energy, and resource usage.

Proposed method

• Introduces context-aware offloading scheduling that dynamically evaluates device load, network conditions, and application latency requirements to make real-time offloading decisions.
• Employs a heuristic-based model partitioning algorithm that splits deep learning models into sub-models executed on the wearable and paired handheld, minimizing data transfer overhead.
• Supports partial offloading by identifying optimal split points within a model to reduce intermediate tensor size and communication delay.
• Implements pipelining to overlap computation and data transfer between wearable and handheld, improving end-to-end throughput.
• Uses lightweight online prediction models for latency and energy consumption to guide offloading decisions without significant runtime overhead.
• Deploys as a user-space library with TensorFlow-like APIs, enabling seamless integration into existing DL applications on Android.

Experimental results

Research questions

• RQ1 How can deep learning inference be efficiently offloaded from wearables to nearby handheld devices without relying on cloud connectivity?
• RQ2 What are the performance and energy trade-offs of offloading DL tasks to a paired handheld versus executing them locally on the wearable?
• RQ3 How can model partitioning be intelligently optimized to minimize communication overhead while balancing computational load across devices?
• RQ4 To what extent can context-aware scheduling improve end-to-end latency and energy efficiency in real-world wearable scenarios?
• RQ5 What is the runtime overhead of the offloading framework, and how can it be minimized without compromising functionality?

Key findings

• DeepWear achieves up to 5.08× speedup and 53.5% energy savings compared to running inference solely on the wearable device.
• It delivers up to 23.0× speedup and 85.5% energy savings compared to executing the entire model on the handheld device alone.
• The runtime overhead of DeepWear is minimal, with system profiling adding less than 5% energy overhead when using active Bluetooth probing.
• The model partitioning heuristic reduces computation complexity to nearly O(n), where n is the number of model nodes, ensuring scalability.
• Bluetooth-based communication incurs low overhead (0.49% to 4.21% of total computation time) due to efficient data transfer and pipelining.
• The framework demonstrates robustness across diverse devices (LG Urbane, Galaxy S2, Nexus 6) and popular DL models, including GoogLeNet and TextRNN.