[PD] ANN - Gesture Recognition Capabilities

Georg Holzmann grh at mur.at
Sun May 4 13:59:08 CEST 2008


> Hm. You can incorporate changes over time using a standard feedforward
> ANN by wrapping your time-ordered vectors over a given time period into
> a single input vector and increasing the number of inputs to the network
> accordingly. But of course this introduces latency and other problems
> (e.g. it could massively increase the number of training examples
> required).
> Pd has the ann_td external, which provides a 'time delay' neural network
> which I believe incorporates time using a method similar to that
> described above.

Yes of course, one other possibility is to use time delayed neural 
networks ...

>> For example a hidden markov model or echo state network (= special kind 
>> of recurrent neural network) should work.
> I'm intrigued! Presumably these approaches avoid the latency problem by
> maintaining the network's state? Are there other advantages -- easier to
> train?

Hm ... I did not think about latency ... but if you do not process the 
data in blocks there should not be a significant latency (also for the 
time delay NN) ?

However, the advantage of the echo state network is that training is 
linear and you cannot get in a suboptimal solution as with feedforward 
neural networks (where the error surface has multiple local minimas) - 
see for example http://www.scholarpedia.org/article/Echo_state_network 
for a short introduction.
And it is recurrent - so in general more powerful ... from the link above:
"On a number of benchmark tasks, ESNs have starkly outperformed all 
other methods of nonlinear dynamical modelling"

If you are interested, I implemented ESNs (with various extensions) in a 
C++ library with python bindings: http://aureservoir.sourceforge.net/, a 
PD external will hopefully follow in summer ...


More information about the Pd-list mailing list