Artificial spiking neural network advantages/disadvantages?
Peter Bencsik
accurate model of biological neural networks. For biological systems I
see why SNN are there, they are just conforming better to the real
neurons, that produce spikes instead of issueing a constant output level
as it is assumed for perceptron-based ANNs.
However, what is the motivation to use SNN also for ANNs? Computer
arithmetik is perfectly fine with continouus output levels and the
asynchronous model of SNN is more difficult to emulate on
microprocessors, I think.
Is my assessment correct?
Do artifical SNNs offer some advantages over perceptron-based ANNs?
Are there simple implementations for SNN (like one find them for other
ANNs, see or example http://www.ip-atlas.com/pub/nap/zip/) where one can
play with?
Stephen Wolstenholme
wrote:
>So far I understood that Spiking Neural Networks (SNN) are a more
>accurate model of biological neural networks. For biological systems I
>see why SNN are there, they are just conforming better to the real
>neurons, that produce spikes instead of issueing a constant output level
>as it is assumed for perceptron-based ANNs.
Biological neural networks use both levels and spikes.
>However, what is the motivation to use SNN also for ANNs? Computer
>arithmetik is perfectly fine with continouus output levels and the
>asynchronous model of SNN is more difficult to emulate on
>microprocessors, I think.
>
>Is my assessment correct?
>
Yes, you are correct.
>Do artifical SNNs offer some advantages over perceptron-based ANNs?
>
None that I have noticed.
>Are there simple implementations for SNN (like one find them for other
>ANNs, see or example http://www.ip-atlas.com/pub/nap/zip/) where one can
>play with?
I modified a BP trained feed forward system to also use internode
"assist" spikes. It works but has no real advantages that I can
determine.
Steve
--
Neural Planner Software Ltd http://www.NPSL1.com
EasyNN-plus. Build Neural Networks. http://www.easynn.com
JustNN. Just Simple Neural Networks. http://www.justnn.com
SwingNN. Forecast the Future. http://www.swingnn.com
Marcus Lauster
>[...]
>
> Do artifical SNNs offer some advantages over perceptron-based ANNs?
I've never used pulse coded or spiking neuron models up to now, but
according to [1] some advantages are:
- speed
- real time action
- complexity
- biological fidelity
In [2] the author says that spiking models are nice because "[...] they
can encode temporal information in their signals[...]".
Wolfgang Maass in [3] compares the standard McCulloch Pitts
approach to the spiking model approach and concludes that the latter is
more powerful because fewer Neurons are needed to accomplish the same,
it is more biological plausible and you can approximate any function
with this model.
[1]Pulsed Neural Networks and their Application (Kunkle, Merrigan 2002)
[2]Spiking neural networks, an introduction (Jilles Vreeken)
[3]Networks of Spiking Neurons: The third generation of neural network
models (Wolfgang Maass, 1997)
Additional interesting resources might be:
Which Model to Use for Cortical Spiking Neurons? (Eugene M. Izhikevich,
2004)
Bayesian inference in spiking neurons (Sophie Deneve,2004)
Spiking Neuron Models: Single Neurons, Populations, Plasticity (Wulfram
Gerstner, Werner Kistler, 2002)
wlorenz65
> Do artifical SNNs offer some advantages over perceptron-based ANNs?
With Backpropagation Through Time for timesteps of 40 ms per iteration this would mean having to unfold the network 3000 times.