Collaborative Research: CSR: Small: Expediting Continual Online Learning on Edge Platforms through Software-Hardware Co-designs

深度神经网络（DNN）在机器人辅助老年人护理、移动的诊断和野生动物监测等新兴应用领域获得了极大的普及。这些应用程序通常（i）采用持续的在线学习，使用流输入训练数据来微调DNN模型，以服务超时推理请求，以及（ii）在能量受限的边缘设备上部署DNN模型。因此，模型自适应性和设备能效对于用户满意度至关重要。这项研究揭示了微调中的冗余，并提高了计算效率，以实现边缘设备上实用，高效和自适应的持续在线学习。该项目的教育和推广部分包括（i）深度学习和边缘计算的课程和课程项目扩展。(ii)通过高级课程项目和PI研究所的外展计划，让本科生参与研究活动。(iii)提高女性和少数民族学生在计算机科学和工程领域的参与度和知名度。这项研究旨在同时实现边缘设备上持续在线学习的适应性和能源效率。(i)它开发了一种注意力引导的智能层冻结，以减少计算成本，自动和动态冻结收敛层。(ii)它为边缘设备设计了一个有效的情境学习框架。该框架选择性地延迟和合并微调迭代，以减少微调频率和处理场景更改。(iii)该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Deep neural networks (DNNs) have gained significant popularity in emerging application domains such as robot-assisted eldercare, mobile diagnosis, and wildlife surveillance. These applications commonly (i) employ continual online learning that fine-tunes the DNN model based using streaming-in training data to serve overtime inference requests and (ii) deploy the DNN models on energy-constrained edge devices. As such, both model adaptiveness and device energy efficiency are critical for user satisfaction. This research uncovers redundancy in fine-tuning and enhances the computation efficiency to achieve practical, efficient, and adaptive continual online learning on edge devices. This project's educational and outreach components include (i) curriculum and course project expansion on deep learning and edge computing. (ii) Engaging undergraduate students in research activities through senior course projects and outreach programs at PIs’ institute. (iii) Increasing the participation and visibility of female and minority students in computer science and engineering.This research aims to simultaneously achieve adaptiveness and energy efficiency for continual online learning on edge devices. (i) It develops an attention-guided smart layer freezing to reduce computation costs by automatically and dynamically freezing converged layers. (ii) It designs an efficient in-situation learning framework for edge devices. The framework selectively delays and merges fine-tuning iterations to reduce the fine-tuning frequency and handles scenario changes. (iii) It designs hardware-support memorization to reduce the amount of fine-tuning computation and memory accesses.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.