SHF: Small: A Design Automation Methodology for Flexible Real-Time Computing based on Split and Early Exit Neural Models

该项目的总体目标是为实时应用开发高度适应性的深度学习(DL)框架，如自动驾驶汽车和移动医疗。为此，提出的设计自动化方法将通过利用拆分计算(SC)和提前退出计算(EEC)技术，将运行时系统优化与高级DL模型体系结构联系起来。在传统方法中，DL模型的设计过程是孤立执行的，其结构相对于数据集和固定的系统条件进行优化。取而代之的是，新的框架将联合使用SC和EEC来构建专门设计的DL模型，以使实时数据分析适应系统的时变特征(例如，可用能量、计算能力、信道容量、计算任务等)。和信息流。为了实现这一目标，该团队将使用深度强化学习和神经结构搜索等工具。该项目将包括一个布局良好的教育和推广计划，其中本科生和研究生的参与尤其有希望。该项目还提出了一系列大学倡议，他们将在执行该项目期间利用这些倡议来增强多样性。研究努力将产生框架，预计将显著提高关键应用的性能，如自动驾驶汽车和人工智能支持的移动健康监测，同时降低它们的能源消耗和无线信道使用。软件和模拟文物将作为开源平台发布，以加强这一重要领域的研究。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The overarching goal of this project is to develop highly adaptable deep-learning (DL) frameworks for real-time applications, such as autonomous vehicles and mobile health. To this aim, the proposed design-automation methodology will bridge runtime system optimization with advanced DL model architectures, through leveraging the techniques of split computing (SC) and early-exit computation (EEC). In traditional methodologies, the design process of a DL model is performed in isolation, where its structure is optimized with respect to a dataset and fixed system conditions. Instead, the new frameworks will jointly use SC and EEC to build DL models specifically designed to adapt real-time data analysis to time-varying characteristics of the system (e.g., available energy, computing power, channel capacity, computing task, etc.) and the information stream. To accomplish this objective, the team will use tools such as deep reinforcement learning and neural architecture search. The project will include a well-laid out educational and outreach plan, in which the involvement of undergraduate and graduate students is particularly promising. The project also proposes a suite of university initiatives which they will leverage to enhance diversity during the execution of this project.The research endeavor will produce frameworks that are expected to considerably boost the performance of critical applications such as autonomous vehicles and AI-empowered monitoring for mobile health while reducing their energy consumption and wireless channel usage. Software and simulation artifacts will be released as open-source platforms to enhance research in this important area.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.