Designing Bayesian based Adaptive Resource Constrained Hardware Algorithms for Next Generation of Embedded Systems

从深度学习社区引入了新的学习方法，使深度学习代理能够在相同或新的环境中适应并持续学习，以模仿人类的学习，如迁移学习、少机会学习和持续学习。尽管有这些方法，但仍有一些限制必须解决，以便在资源受限的传感平台上实现无缝和自适应的人工智能。首先，现有的解决方案关注的焦点往往是准确性，而缺乏对其他性能指标的关注，如内存、CPU和能耗，这对低功耗、资源受限的设备至关重要。其次，目前的工作没有考虑学习过程中的不确定性，而专注于对设备的推理。我们希望超越推理，在设备上进行更具挑战性的学习。该研究项目的总体目标将是为资源受限的设备，如微控制器单元(MCU)和现场可编程门阵列(FGA)，研究高效和可靠的设备学习解决方案。具体地说，该项目的目标将包括探索不同的贝叶斯连续学习方法在各种传感任务中的性能，以发现存在的问题，通过解决这些问题来研究和设计用于贝叶斯连续学习的新的最大似然算法，并提出基于系统的优化技术。

Novel learning approaches from deep learning (DL) community are introduced to allow DL agents to adapt and continually learn in the same or new environments to mimic human like learning such as transfer learning, few shot learning, and continual learning. Despite these approaches there are still limitations that must be addressed to enable seamless and adaptable AI on resource constrained sensing platforms. Firstly, the existing solutions focus is often dominated by accuracy and lacks attention to other performance measures such as memory, CPU, and energy consumption which is critical to low-power resource-constrained devices. Secondly, the current work does not account for the uncertainties in the learning process and focus on inference on the device. We would like to go beyond inference and take on more challenging pursuit of learning on the device. The overall goal of this research project would be to investigate efficient and reliable on-device learning solutions for resource-constrained devices such as MCUs (microcontroller units) and FPGAs. Specifically, the aims of the project would include exploring the performance of different Bayesian continual learning approaches for various sensing tasks to find existing issues, investigate and design novel ML algorithms for Bayesian continual learning by solving these issues and propose system-based optimization techniques.