Hybrid and Compressed DNN Ensemble Models on Edge Devices for Handling Diverse Noises
Abstract: Edge computing usage in many applications, such as transportation and healthcare, has been becoming increasingly prominent. These applications often use deep neural network (DNN) prediction, which is highly dependent on time-series data collected by the sensors in the edge devices. However, the presence of noise in the on-device sensors negatively affects the sensing output of the DNN models. Recently proposed time-series based DNN approaches (e.g., SADeepSense) address this issue by assuming that in the presence of Gaussian noise, the correlation of sensor inputs in an edge device changes. However, there exist different types of noises, such as shot, burst, transient noises, and the combination of these noises. In this paper, we propose an ensemble-based DNN model, namely EnsembleSense, which consists of different expert models for different noises and shows higher prediction accuracy. We further propose EnsembleCompression, a novel searching-based compression scheme using knowledge distillation for EnsembleSense that minimizes the inference time, satisfies the memory constraint of an edge device while meeting the accuracy requirement. Our trace-driven experiments for three real traces and real experiments for live sensor data show that EnsembleSense outperforms other methods in accuracy, and EnsembleCompression significantly reduces inference time and memory demand without sacrificing great accuracy compared with other DNN compression methods.
Committee:
- Andrew Grimshaw (Chair)
- Haiying Shen (Advisor)
- Lu Feng
- Yuan Tian