Error-driven learning for self-calibration in a neuromorphic path integration system
Error-driven learning for self-calibration in a neuromorphic path integration system
复制标题
神经形态路径集成系统中用于自校准的误差驱动学习
DOI：
--
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Yulia Sandamirskaya
中科院分区：
文献类型：
--
作者：
Raphaela Kreiser;Alpha Renner;Yulia Sandamirskaya
神经形态硬件提供了一个与基于神经网络的机器人控制器相匹配的计算基底。为了避免神经形态硬件和传统计算机之间的瓶颈，我们的目标是在神经形态设备上实现整个感知-动作控制回路。在这项工作中，我们展示了一个例子，这样的神经形态架构的一个移动的自主代理的基本能力之一：估计当前的姿势的代理的基础上执行的运动，使用路径集成。精确的路径整合需要内部姿态表示中的整合速度与实际姿态变化之间的对准。在本文中，我们提出了一个尖峰神经网络（SNN）的状态估计在1D中，灵感来自动物中发现的航向网络。此外，我们引入了一种机制，自主校准的路径集成系统。如果在环路闭合事件处检测到姿态估计中的错误，则网络使积分速度适应代理的旋转速度。我们实现和验证的网络上的低功耗尖峰神经形态处理器Loihi。在片上SNN中实现自主校准是纯神经形态机器人控制器中有效和自适应状态估计和地图形成的重要组成部分。
Neuromorphic hardware offers a computing substrate that matches the neuronal network-based robotic controllers. To avoid a bottleneck be-tween neuromorphic hardware and a conventional computer, we aim to realize the whole perception-action control loop on the neuromorphic device. In this work, we show an example of such neuromorphic architecture for one of the basic capabilities of a mobile autonomous agent: estimation of the current pose of the agent based on the executed movements using path integration. Accurate path integration requires alignment between the integration speed in the internal pose-representation and the actual pose change. In this paper we propose an Spiking Neural Network (SNN) for state estimation in 1D, inspired by heading direction networks found in animals. Further, we introduce a mechanism for autonomous calibration of the path integration system. The network adapts the integration speed to the rotational speed of the agent if an error in pose estimation is detected at a loop closure event. We implement and validate the network on a low-power spiking neuromorphic processor Loihi. This implementation of autonomous calibration in an on-chip SNN is an important component of efficient and adaptive state estimation and map formation in a purely neuromorphic robotic controller.
DOI：
--
发表时间：
1994
期刊：
Advances in neural information processing systems
影响因子：
--
作者：
W. Skaggs;J. Knierim;H. Kudrimoti;B. McNaughton
通讯作者：
W. Skaggs;J. Knierim;H. Kudrimoti;B. McNaughton