CRII: CNS: A Systematic Multi-Task Learning Framework for Improving Deep Learning Efficiency on Edge Platforms

多任务学习是机器学习的一个子领域，其中数据使用共享模型进行训练，以同时解决不同的任务。多任务学习大大减少了机器学习模型中的参数数量，从而降低了计算和存储要求。例如，在自动驾驶汽车中有多个任务需要实时完成，包括物体检测和深度估计。如果这些任务可以在具有共享参数的单个模型上训练，则可以大大减少模型大小和推理时间。该项目旨在进一步压缩用于多任务学习的模型，因为单个深度神经网络的模型大小仍然是许多计算系统的关键挑战，特别是对于边缘平台。该项目提出了一种方法，可以在压缩多任务学习模型时学习每个任务的难度并保持最困难任务的性能。它增加了压缩率的潜力，并且对于所有任务都具有可接受的性能，因为需要保证最困难任务的性能以提供令人满意的用户体验。该项目还为边缘平台设计了一种高效的多任务联合学习方法。它提高了多任务联邦学习的收敛速度，减少了每次迭代的通信开销。最后，该项目提出解决算法-硬件协同设计问题，以最大化压缩多路复用的实现效率，在边缘平台上的DNN模型的任务。压缩DNN模型的文件和高效DNN训练和实现的想法可能对专注于提高DNN模型在边缘平台和其他硬件平台上的计算效率的研究人员有用。本项目将涉及本科生和研究生学生在研究。该项目的研究成果将被纳入目前的高级本科课程，一个新的计划高级研究生课程，以及本科生和研究生的研讨会。 在为K-12学生举办的研讨会和夏令营期间，还计划进行研究演示。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Multi-task learning is a subfield of machine learning in which the data is trained with a shared model to solve different tasks simultaneously. Multi-task learning highly reduces the number of parameters in the machine learning models and thus reduces the computational and storage requirements. For example, there are multiple tasks to be done in real-time in self-driving cars, including object detection and depth estimation. If these tasks can be trained on a single model with shared parameters, the model size and the inference time can be highly reduced. This project aims to further compress the model used for multi-task learning as the model size of a single deep neural network is still a critical challenge to many computation systems, especially for edge platforms. This project proposes an approach to learn the difficulty of every task and maintain the performance of the most difficult task when compressing a multi-task learning model. It increases the potential in the compression rate with acceptable performance for all the tasks as the performance of the most difficult task needs to be guaranteed to provide a satisfactory user experience. This project also designs an efficient multi-task federated learning approach for edge platforms. It improves the convergence rate of multi-task federated learning and reduces the communication costs in every iteration. Finally, this project proposes to solve an algorithm-hardware co-design problem to maximize the implementation efficiency of the compressed multi-task DNN models on edge platforms.The files of compressed DNN models and the ideas on efficient DNN training and implementation may be useful to researchers who focus on improving the computation efficiency of DNN models on edge platforms and other hardware platforms.This project will involve undergraduate and graduate students in the research. The research achievements of this project will be incorporated into a current senior-level undergraduate course, a new planned advanced-level graduate course, and seminars for both undergraduate and graduate students. There are also planned research demonstrations during the workshops and summer camps for the K-12 students.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.