EAGER: III: Small: Green Granular Neural Networks with Fast FPGA-based Incremental Transfer Learning

目前，流行的机器学习系统运行在两种主要类型的处理单元上-中央处理单元（CPU）和图形处理单元（GPU）。这些单元产生大量的二氧化碳（CO2）排放，并且还浪费能量，因为传统的训练算法（例如，梯度下降算法和遗传算法）需要大量的时间来优化数十亿个超参数，并且因为CPU和GPU不是节能的。由于估计的CO2排放量与总计算能力成正比，因此有效减少CO2排放的两种主要方法包括开发新型高速非传统训练算法以显着减少训练时间，以及使用具有低CO2排放的新型计算硬件。一个紧迫的挑战是开发一种新型的机器学习系统，在绿色和节能的硬件上运行高速非传统的训练算法，以显着减少CO2排放和能源消耗。一种新型的基于浅层软硬件的绿色颗粒神经网络（GGNN），具有新的快速，现场可编程门阵列（FPGA）。将开发基于FPGA的增量迁移学习算法，以比基于CPU-GPU的训练算法更有效地减少二氧化碳排放和能耗。由于FPGA是一种具有低CO2排放和低能耗的轻量级硬件，因此FPGA用于快速求解线性方程组，以直接计算浅层GGNN超参数的最佳值。具有转移的粒度知识的浅GGNN可以增量地且快速地训练，以针对实时绿色计算应用做出有效的决策。一个新的GGNN树将被开发来解决维数灾难。此外，浅层GGNN是可解释的，因为它可以生成可解释的粒度If-Then规则。该项目对于改善人工智能、计算智能、数据挖掘和操作系统等相关课程的智能绿色计算教育也是有用的。本科生和研究生，包括代表性不足的学生将一起做研究工作。学生将接受良好的培训，成为未来的绿色机器学习劳动力，开发新的基于软硬件的机器学习算法，以大大减少二氧化碳排放和能源消耗。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。

Currently, popular machine learning systems running on the two main types of processing units -- central processing units (CPUs) and graphics processing units (GPUs). These units generate a lot of carbon dioxide (CO2) emissions and also waste energy because traditional training algorithms (e.g., gradient descent algorithms and genetic algorithms) take huge amount of time to optimize billions of hyperparameters, and because CPUs and GPUs are not energy efficient. Since the estimated CO2 emission amount is proportional to the total computational power, two major ways of effectively reducing CO2 emissions include developing novel high-speed non-traditional training algorithms to significantly reduce training times, and use novel computational hardware with low CO2 emissions. An urgent challenge is developing a novel machine learning system with high-speed non-traditional training algorithms running on green and energy efficient hardware to significantly reduce both CO2 emissions and energy consumption.A novel shallow software-hardware-based green granular neural network (GGNN) with new fast, field-programmable gate array (FPGA). FPGA-based incremental transfer learning algorithms will be developed to reduce both CO2 emissions and energy consumption more effectively than the CPU-GPU-based training algorithms. Since FPGA is a light-weight hardware with low CO2 emissions and low energy consumption, the FPGA is used to quickly solve a system of linear equations to directly calculate optimal values of hyperparameters of the shallow GGNN. The shallow GGNN with transferred granular knowledge can be trained incrementally and quickly to make efficient decisions for real-time green computing applications. A new GGNN tree will be developed to resolve the curse of dimensionality. In addition, the shallow GGNN is explainable because it can generate interpretable granular If-Then rules. This project is also useful for improving intelligent green computing education for relevant courses such as artificial intelligence, computational intelligence, data mining, and operating systems. Undergraduate students and graduate students including underrepresented students will do research works together. Students will be well trained to become future green machine learning workforce to develop novel software-hardware-based machine learning algorithms to greatly reduce both CO2 emissions and energy consumption.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.