A neuromorphic SLAM architecture using gated-memristive synapses
A neuromorphic SLAM architecture using gated-memristive synapses
复制标题
DOI：
10.1016/j.neucom.2019.09.098
复制
发表时间：
2020-03
期刊：
影响因子：
6
通讯作者：
Alexander Jones;A. Rush;Cory E. Merkel;Eric Herrmann;A. Jacob;Clare D. Thiem;R. Jha
中科院分区：
文献类型：
--
作者：
Alexander Jones;A. Rush;Cory E. Merkel;Eric Herrmann;A. Jacob;Clare D. Thiem;R. Jha
在GPS拒绝的环境中导航是无人机等自主移动的平台的关键挑战。同时定位和地图（SLAM）的概念通过在平台探索其环境时实时绘制平台周围环境来解决这一挑战。传统SLAM实现所需的计算资源（例如，图形处理单元）需要大的尺寸、重量和功率开销;使得在资源受限的应用中采用它们是不可行的。这项工作提出了一种自学习的硬件架构，利用一种新的门控忆阻器件，以解决实现SLAM的节能方式。门控忆阻器件被实现为与新型低能量尖峰神经元串联的电子突触，以创建尖峰神经网络（SNN）。这项工作显示了SNN如何允许通过地标关联在不需要GPS的情况下通过环境进行导航。在网络存在的简单环境中，它可以成功地确定导航的方向，同时仅消耗36 µW的功率，并且仅需要暴露于环境中的每个地标1- 2 ms以记住该位置。
Navigation in GPS-denied environments is a critical challenge for autonomous mobile platforms such as drones. The concept of simultaneous localization and mapping (SLAM) addresses this challenge through real-time mapping of the platform's surroundings as it explores its environment. The computational resources required for traditional SLAM implementations (e.g. graphical processing units) require large size, weight, and power overheads; making it infeasible to employ them in resource-constrained applications. This work proposes a self-learning hardware architecture utilizing a novel gated-memristive device to address the implementation of SLAM in an energy-efficient manner. The gated-memristive devices are implemented as electronic synapses in tandem with novel low-energy spiking neurons to create a spiking neural network (SNN). This work shows how the SNN allows for navigation through an environment via landmark association without needing GPS. In the simple environment in which the network exists, it can successfully determine a direction in which to navigate while only consuming 36 µW of power and only needing to be exposed to each landmark within the environment for 1-2ms in order to remember that location.