Autonomous Learning Paradigm for Spiking Neural Networks
Autonomous Learning Paradigm for Spiking Neural Networks
复制标题
DOI：
10.1007/978-3-030-30487-4_57
复制
发表时间：
2019-09
期刊：
影响因子：
--
通讯作者：
Junxiu Liu;L. McDaid;J. Harkin;Shvan Karim;Anju P. Johnson;D. Halliday;A. Tyrrell;J. Timmis;Alan G. Millard;J. Hilder
中科院分区：
文献类型：
--
作者：
Junxiu Liu;L. McDaid;J. Harkin;Shvan Karim;Anju P. Johnson;D. Halliday;A. Tyrrell;J. Timmis;Alan G. Millard;J. Hilder
与生物系统相比，现有的学习系统缺乏自主学习的能力，尤其是在变化和动态的环境中。本文通过开发自学习尖峰神经网络 (SNN) 并使用简单的机器人控制器应用程序展示其自主学习能力来解决自主学习问题。我们提出的学习规则利用了现有尖峰时序相关可塑性 (STDP) 规则的继承属性，即如果瞬时突触前频率降低，则对于传统的赫布窗口，STDP 规则会增强。相反，如果瞬时频率增加，STDP 规则就会降低：反赫布窗口的情况正好相反。本文还将表明，使用传统的赫布学习窗口可以实现避障，而使用反赫布学习窗口可以学习目标跟踪。因此，所提出的学习范式是新颖的，因为它不需要对这些任务进行外部监督。所提出的学习范式还使用了先前探索的星形胶质细胞神经元相互作用，其中来自星形胶质细胞的周期性慢内向电流（SIC）可以在一段时间内增强突触后神经元：该时间窗口可用于加强/削弱突触通路。使用避障任务进行性能分析，结果表明基于 SNN 的机器人控制器具有动态条件下的自主学习能力。
Compared to biological systems, existing learning systems lack the ability to learn autonomously, especially in changing and dynamic environments. This paper addresses the issue of autonomous learning by developing a self-learning spiking neural network (SNN) and demonstrating its autonomous learning capability using a simple robot controller application. Our proposed learning rule exploits an inherit property of the existing Spike-Timing-Dependent Plasticity (STDP) rule in that if the instantaneous presynaptic frequency decreases, then for a conventional Hebbian window the STDP rule potentiates. Conversely if the instantaneous frequency increases the STDP rule depresses: the opposite is true for anti-Hebbian window. This paper will also show that obstacle avoidance is achievable using a conventional Hebbian learning window while object tracking can be learned using an anti-Hebbian learning window. Hence the proposed learning paradigm is novel in that it does not require external supervisions for either these tasks. The proposed learning paradigm also uses a previously explored astrocyte neuron interaction where a periodic Slow Inward Current (SIC) from an astrocyte can potentiate a postsynaptic neuron for a period of time: this time window can be used to strengthen/weaken synaptic pathways. An obstacle avoidance task is used for the performance analysis and results show that the SNN based robot controller has autonomous learning capabilities under the dynamic conditions.