[Paper study] Delay learning and polychronization for reservoir computing

最新推荐文章于 2019-09-08 22:39:56 发布

原创最新推荐文章于 2019-09-08 22:39:56 发布 · 1.5k 阅读

1 ·

本内容遵循CC 4.0 BY-SA版权协议

标签

#polychronization #reservoir computing #delay learning

Paper Study 专栏收录该内容

4 篇文章

订阅专栏

本文探讨了将脉冲神经网络(SNN)作为卷积神经网络(CNN)的基础架构，并介绍了一种名为水库(reservoir)的SNN结构及其学习算法——水库计算(RC)。通过多时间尺度学习算法，包括毫秒级的突触可塑性(STDP)及监督学习调整轴突传输延迟，实现了对输出神经元编码的优化。

CNNs use spiking neural networks (SNNs) as its backbone. A SNN structure called reservoir and its learning algorithm “reservoir computing” (RC) is described in Paugam-Moisy et al. 2008. She adopted a multi-timescale learning algorithm, which use STDP in millisecond level (synapse level) and a supervised adaptation of axonal transmission delays toward readout neurons coding.

Here is the summary after studying this paper.

This is the network structure,

This is the STDP weight updating rule,

A multiplicative approach is adopted for weight updating where weights are updated as Wij = Wij + α * (Wij – Wmin) * ΔW, if ΔW <=0. And Wij = Wij + α * (Wmax – Wij) * ΔW, if ΔW >=0.

This is the SRM0 spiking neuron model,

This is the pseudo-code for supervised adaptive delays (the multi class version is analogical).

repeat
for each example X=(p, class) of the database
{
    present the input pattern p;
    define the target output neuron according to class;
    if (the target output neuron fires less than μms before the non-target one, or fires after it ) then
    {
        select one triggering connection of the target
        output neuron and decrement its delay (-1ms),
        except if dmin is reached already;
        select one triggering connection of the non-target
        output neuron and increment its delay (+1ms),
        except if dmax is reached already;
    }
}
until a given maximum learning time is over.

Note that the recognized pattern is represented by the first firing output neuron.

Simulations and visualization of the reservoir will follow in the future posts. Enjoy.