A 65nm 236.5nJ/Classification Neuromorphic Processor with 7.5% Energy Overhead On-Chip Learning Using Direct Spike-Only Feedback

📋 论文概要

该论文提出了一种65nm工艺的神经形态处理器，通过直接尖峰学习方法实现片上学习，能量开销仅为7.5%，在分类任务中每次分类能耗为236.5nJ。解决了传统神经网络处理器在学习过程中计算和能耗过高的问题。

💡 主要创新点

• 提出直接尖峰学习方法，仅使用尖峰信号进行片上学习，大幅降低计算和能耗开销。
• 实现了7.5%能量开销的片上学习，显著提升了能效比。
• 该学习规则可推广至全连接网络，适用于迁移学习等场景。

🏷 关键词

📄 原文摘要

classifying handwritten digits, the learning rule can be directly adopted in general fully-connected networks with different network structures and hence can be extended to other applications. For example, the algorithm could serve a key role in fine tuning large networks (typically based on CNN) through transfer learning and training fully-connected layers in the last stages. Advances in neural network and machine learning algorithms have sparked a wide array of research in specialized hardware, ranging from high-performance convolutional neural network (CNN) accelerators to energy-efficient deep-neural network (DNN) edge computing systems [1]. While most studies have focused on designing inference engines, recent works have shown that on-chip training could serve practical purposes such as compensating for process variations of in-memory computing [2] or adapting to changing environments in real time [3].

👥 作者与机构

Seoul National University, Seoul, Korea, spikes, greatly reducing computation and global interconnects. While the singleshot modified SD algorithm was tested on a relatively small-scale problem of