Hyper-Parallel Calculator-Free Neural Network Accelerator for Edge AI Applications

在这项研究中，我们的目标是为边缘人工智能应用开发一个超并行的无计算器神经网络加速器。我们的框架通过无计算器神经网络和超并行加速器协同加速边缘智能应用。具体来说，在我们提出的超稀疏平分神经网络拓扑的基础上，利用尖峰范式对无计算器神经网络进行了探索，从而基于我们提出的多粒度可重构平台和所提出的拓扑结构开发了超并行神经网络加速器。在2022财年，一种具有梯度隔离记忆机制的新型尖峰神经网络框架在同行评议的国际会议（HPCC 2022）上发表。在此基础上，利用DiaNet实现了无计算器的神经网络拓扑结构，并在Python代码中对MNIST和NMNIST基准进行了评估。此外，计算单元由LUT实现，而不是传统的乘法和加法运算，相应的数字电路设计在Verilog代码上实现。然后，我们在实际硬件（PYNQ-Z1 FPGA）上测试了所提出的无计算器神经网络加速器。目前，我们正在进行实验并撰写论文，将于2023财年向期刊投稿。为了探索加速器应用的多样性，我们还实施了一个初步的问题相关性应用程序，并在同行评议的国际会议（HPCC 2022）上发表。

In this research, we aim to develop a hyper-parallel calculator-free neural network accelerator for edge AI applications. Our framework collaboratively accelerates edge intelligence application through the calculator-free neural network and hyper-parallel accelerator. Specifically, the calculator-free neural network is explored by the spiking paradigm on the basis of our novel ultra-sparse bisection neural network topology, and thus the hyper-parallel neural network accelerator is developed based on our multi-grained reconfigurable platform and the proposed topology. In FY2022, a novel spiking neural network framework with gradient-isolated memorizing mechanism was published at the peer-reviewed international conference (HPCC 2022). Accordingly, the calculator-free NN topology was implemented by DiaNet with the proposed spiking paradigm, which was evaluated on MNIST and NMNIST benchmark in Python code. Furthermore, the computing units are implemented by the LUT rather than the conventional multiplication and addition operations, and corresponding digital circuit designs are implemented on Verilog code. We then test the proposed calculator-free neural network accelerator on real hardware (PYNQ-Z1 FPGA). Currently, we are conducting experiments and writing a paper that will be submitted to a journal in FY2023. To explore the diversity of applications for our accelerator, we have also implemented a preliminary question-relatedness application, which was published at the peer-reviewed international conference (HPCC 2022).