| dc.description.abstract | With the proliferation of edge devices such as mobile phones, consumer robots, drones, wearables, and IoT devices, the generation of data at the edge of the internet network has been increasing exponentially. Machine Learning (ML) models, particularly Deep Neural Networks (DNNs), have the ability to process this data with remarkable accuracy. However, state-of-the-art ML models require substantial computational resources that edge devices typically lack, necessitating a shift to powerful servers in the cloud as hosts for these models. Running these models at the edge is desirable due to benefits such as low-latency results and adherence to data privacy constraints, but is limited by the available computational power and energy consumption of edge devices. Moreover, lightweight models designed for edge devices often exhibit a significant drop in accuracy. Continuous learning offers a potential solution by improving the accuracy of lightweight models by dynamically adapting them to specific scenes or narrow distributions of inputs, which is especially relevant since in practice, these models do not need to generalize to every possible sample from the distribution.
In this thesis, two key methods are introduced to tackle the challenges in continuous learning systems for edge devices: Model Streaming and Model Reuse. Model Streaming offloads the adaptation process to remote machines with greater computational capacity and updates only a critical subset of model parameters that significantly influence the lightweight model’s performance, reducing the bandwidth needed for model updates. Model Reuse uses an efficient DNN model to dynamically select a suitable lightweight model from a library of historical models designed for similar input distributions, boosting the scalability, responsiveness, and accuracy of continuous learning systems. These methods are applied to practical systems, including MMNet for adaptive neural signal detection in 5G cellular communication systems, AMS for real-time video inference on edge devices, SRVC for efficient video compression, and RECL for responsive, resource-efficient continuous learning for video analytics.
We show how continuous learning can significantly improve lightweight machine learning inference on edge devices. The proposed techniques effectively address the unique challenges posed by resource-constrained edge environments. Practical applications presented in the thesis, such as MMNet, AMS, SRVC, and RECL, demonstrate the real-world effectiveness of these methods. These innovations in continuous learning have the potential to reshape the landscape of edge computing by offering more accurate and adaptable inference capabilities, enabling efficient use of computational resources, reduced latency, and better energy efficiency. | |
