脉冲（尖峰）神经网络是否优于循环神经网络？
脉冲神经网络和循环神经网络都可以对时变信息进行建模。但我不确定哪种模型相对于计算成本更好。使用更复杂的脉冲神经网络是否值得，或者循环神经网络是否可以在所需计算量少得多的情况下工作？脉冲网络收敛得更快吗？
谢谢
Both pulsed and recurrent neural networks can model time varying information. But I am not sure which model is better relative to the computational cost. Does it pay to use the more complex pusled neural network or will the recurrent neural net work just as well with a lot less calculation required? Does a pulsed net converge more quickly?
Thanks
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（2）
我不相信你以正确的方式看待这个问题。神经网络（或任何预测模型）最重要的属性是它的准确性。如果模型的准确度（和预测性）明显提高，我宁愿花 10 倍的时间来构建模型。
有许多标准技术可用于评估模型的预测能力，例如
* 留一法交叉验证
* 留多交叉验证
* Fisher 随机化 (http://en.wikipedia.org/wiki/Ronald_Fisher)
构建预测模型也有许多指导原则，例如
* 奥卡姆剃刀
*避免过度拟合（http://web.engr.oregonstate.edu/~tgd/classes/534/slides/part10.pdf）
* 过度拟合的惩罚 (http://en.wikipedia.org/wiki/Regularization_(mathematics))
以下是一些可以查找更多信息的地方
http://predictivemodelingblog.blogspot.com/
http://www.statsoft.com/textbook/data-mining-techniques/
底线：选择可以解释数据的最简单模型
I don't believe you are looking at the problem the right way. The most important property of a neural network (or any predictive model) is it's accuracy. I would rather spend 10 times longer building the model, if it were significantly more accurate (and predictive).
There are many standard techniques for assessing the predictive power of your model, such as
* leave-one-out cross validation
* leave-many-out cross validation
* Fisher randomization (http://en.wikipedia.org/wiki/Ronald_Fisher)
There are also many guiding principles for building a predictive model, such as
* occam's razor
* avoid overfitting (http://web.engr.oregonstate.edu/~tgd/classes/534/slides/part10.pdf)
* penalties for overfitting (http://en.wikipedia.org/wiki/Regularization_(mathematics))
Here are a few places to look for more information
http://predictivemodelingblog.blogspot.com/
http://www.statsoft.com/textbook/data-mining-techniques/
Bottom line: go for the simplest model that can explain your data
让我尝试回答最初的问题，该问题特定于尖峰神经网络的性能。我最近才开始熟悉尖峰网络，但迄今为止，我发现的对其性能的最佳评论是 Hélène Paugam-Moisy 和 Sander Bohte 的专着“使用尖峰神经元网络进行计算”，无需付费即可获取< a href="http://homepages.cwi.nl/~sbohte/publication/paugam_moisy_bohte_SNNChapter.pdf" rel="nofollow">此处和其他网站。
从这个来源看来，总的来说，他们在与生俱来的能力方面表现良好；尖峰网络是通用逼近器，就像普通神经网络中的一些同类一样，它们的“计算能力优于图灵机”。 （第 5 页，Paugam-Moisy 和 Bohte）。事实上，对于 B 型尖峰神经元，Wolfgang Maass“得出结论，尖峰神经元网络在计算上比第一代和第二代神经网络（即感知器家族和具有连续激活的网络类别）更强大”（第 17 页）。 21，Paugam-Moisy 和 Bohte）。另一方面，它们很容易出现加载问题，在本文讨论的某些条件下，该问题可能是 NP 完全的。
就计算性能而言，尖峰网络的各种子类型的资源需求差异很大。 Paugam-Moisy 和 Bohte 在第 12、14-15 页上对处理操作中的一些差异进行了精彩、快速的总结：
，计算能力和这些计算成本之间存在权衡；LIF 可能需要很少的 FLOPS，但在文章后面，他们详细说明了它如何无法达到其他尖峰所享有的准确性。计算要求也将是受到尖峰网络编码方式的极大影响（我打算自己做，但还没有机会尝试），在第 38-39 页，Paugam-Moisy 和 Bohte 强烈建议使用事件驱动。我还建议您查看他们广泛的参考书目，我相信这可以带来许多其他实用的性能技巧，希望对您有所帮助。
Let me take a stab at answering the original question, which was specific to the performance of spiking neural nets. I've only recently begun to familiarize myself with spiking nets, but to date the best commentary I've found on their performance is Hélène Paugam-Moisy and Sander Bohte's monograph "Computing with Spiking Neuron Networks," which is available without a paywall here and other sites.
From this source, it appears that in general, they perform well in the sense of their innate capabilities; spiking networks are universal approximators and like some of their kin among ordinary neural nets, their "computational power outperforms a Turing machine." (p. 5, Paugam-Moisy and Bohte). In fact, for Type-B spiking neurons, Wolfgang Maass "concludes that spiking neuron networks are computationally more powerful than both the 1st and the 2nd generations of neural networks," i.e. the perceptron family and the class of nets with continuous activations (p. 21, Paugam-Moisy and Bohte). On the other hand, they're prone to the Loading Problem, which can be NP-Complete under certain conditions discussed in the paper.
In terms of computational performance, the sundry subtypes of spiking nets vary wildly in their resources demands. On pp. 12, 14-15 Paugam-Moisy and Bohte give a nice, quick summary of some of the differences in processing operations:
As expected, there's a tradeoff between computational capabilities and these computing costs; LIF may require few FLOPS, but later in the article they detail how it is not capable of achieving the accuracy that other spiking enjoy. The computational requirements will also be dramatically affected by the manner in which spiking nets are coded (which I intend to do myself, but have not yet had a chance to experiment with yet). On pp. 38-39 Paugam-Moisy and Bohte strongly recommend using an event-driven architecture that can leverage parallelism to the utmost. I'd also recommend checking out their extensive bibliography, which I'm sure can lead to dozens of other practical performance tips. I hope that helps.