Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss
STDP & SNN for real world classification problems
32 views
Skip to first unread message
MuadDib
May 20, 2012, 11:57:47 AM5/20/12
to
Hi there,
I am in search of references/people who have used spiking neural
networks to learn to classify real world signals (sounds/speech,
features in images, EEG waveforms,...).
I am trying to teach to a liquid derived from Izhikevich's spnet c++
code to discriminate between 3 classes (left foot movement, right foot
movement, verbal task) from a 32 channels EEG recording with/without
labels. 3000 neurons (2800 exc./200 inh.) with 240 synapses each,
synaptic delays from 0 to 60 ms.
The network is mostly wired in a forward fashion: input layer (time
coding/conversion to spikes)/700 neur => layer 1 of the liquid/1200
neur => layer 2 of the liquid/600 neur => output layer/3 classes x 100
neur) with recurrences for the layers 1 and 2 and between them (layer
2=>layer1).
The features injected as microcurrents in the 700 input neurons layer
are 22 maximums detected in varied areas of the time-frequency maps
for each of the 32 EEG signals (32x22~700 features each 250 ms).
In supervised lerning phase, the true label/class is also injected:
the spike thresholds are lowered for the output neurons in the true
class and the opposite for the output neurons in the 2 other classes.
The output neurons of a given class are connected to specialized
inhibitory neurons to inhibit the output neurons of the other classes
(mutual inhibition). An output neuron firing in the wrong class is
prevented to inhibit the other classes' output neurons. Standard STDP
is applied to all the synapses but if output neurons fire in the wrong
class => presynaptic antiSTDP is applied to the bad output neuron (and
3 of its presynaptic neurons in the liquid) and its relevant
presynaptic delays are lengthenned. If an output neuron fires in the
true class, it is rewarded by a faster STDP.
I am starting my first simulations with EEG signals and beyond the
computationnal load, the analysis of the neural populations expected
to emerge, of the learning state of the network, or of the optimal
network settings appears difficult (I am coding a routine to extract
the activated populations similar to Izhikevich's but using the
recorded spiking activity of the network)
I am interested in contacts with people doing the same thing or
relevant references !
Take care,
Jean-François B.
I am in search of references/people who have used spiking neural
networks to learn to classify real world signals (sounds/speech,
features in images, EEG waveforms,...).
I am trying to teach to a liquid derived from Izhikevich's spnet c++
code to discriminate between 3 classes (left foot movement, right foot
movement, verbal task) from a 32 channels EEG recording with/without
labels. 3000 neurons (2800 exc./200 inh.) with 240 synapses each,
synaptic delays from 0 to 60 ms.
The network is mostly wired in a forward fashion: input layer (time
coding/conversion to spikes)/700 neur => layer 1 of the liquid/1200
neur => layer 2 of the liquid/600 neur => output layer/3 classes x 100
neur) with recurrences for the layers 1 and 2 and between them (layer
2=>layer1).
The features injected as microcurrents in the 700 input neurons layer
are 22 maximums detected in varied areas of the time-frequency maps
for each of the 32 EEG signals (32x22~700 features each 250 ms).
In supervised lerning phase, the true label/class is also injected:
the spike thresholds are lowered for the output neurons in the true
class and the opposite for the output neurons in the 2 other classes.
The output neurons of a given class are connected to specialized
inhibitory neurons to inhibit the output neurons of the other classes
(mutual inhibition). An output neuron firing in the wrong class is
prevented to inhibit the other classes' output neurons. Standard STDP
is applied to all the synapses but if output neurons fire in the wrong
class => presynaptic antiSTDP is applied to the bad output neuron (and
3 of its presynaptic neurons in the liquid) and its relevant
presynaptic delays are lengthenned. If an output neuron fires in the
true class, it is rewarded by a faster STDP.
I am starting my first simulations with EEG signals and beyond the
computationnal load, the analysis of the neural populations expected
to emerge, of the learning state of the network, or of the optimal
network settings appears difficult (I am coding a routine to extract
the activated populations similar to Izhikevich's but using the
recorded spiking activity of the network)
I am interested in contacts with people doing the same thing or
relevant references !
Take care,
Jean-François B.
amits...@gmail.com
Aug 31, 2015, 6:40:24 AM8/31/15
to
I am trying to implement the spiking neural network but in different domain.Can you please elaborate what learning algorithm you applied.As your implementation is based on Izhikevich's model.If you will elaborate then it can be easy for me to think about.I want to know how you perform your neuron learning strategy for your network.
Thank you,
Amit S. A.
0 new messages