430
J. Wu et al.
variability and high inter-class variability in both the BMU activation trajectories and
sequences, which are highly desirable for pattern classification.
We note that the time-warping problem exists in the BMU activation sequences,
whereby the duration of sensory stimuli fluctuates from sample to sample within the
same class. However, the SNN-based classifier is robust to such fluctuations as shown
in the classification results. The decision to fire for a classifying neuron is made based
on a time snippet of the spiking pattern; such is the nature of the single spike-based
temporal classifier. As long as the BMU activation sequence stays similar, duration
fluctuations of input sample will not affect the general trajectory of the membrane
potential in each output neuron; the right classification decision, therefore, can be
guaranteed.
To investigate whether the feature dimension reduction of the SOM is necessary
for the SNN classifier to learn different sound categories, we performed experiments that directly input the spike trains of the latency-encoded (20 neurons) [22]
or population-encoded (144 neurons) [23] mel-scaled filter bank outputs into the
SNN for classification. We find that the SNN classifier is unable to classify such
low-level spatiotemporal spike patterns, and only achieve 10.2% and 46.5% classification accuracy for latency- and population-encoded spike patterns, respectively. For
both latency- and population-encoded spike patterns, as all encoding neurons spike
in every sound frame, albeit with different timing, the synaptic weights therefore
either all strengthen or all weaken in the event of misclassification as defined in the
Tempotron learning rule. Such synchronized weight updates make it challenging for
the SNN classifier to find discriminative features embedded within the spike pattern.
The learning rules for the SNN can be categorized into either membrane-potential
based [25, 48, 49] or spike-time based [50, 51]; the Maximum-Margin Tempotron
learning rule belongs to the former. To study the synergy between the SOM-based
feature representation and spike-time based learning rule, we conducted an experiment using the ReSuMe [50] learning rule to train the SNN classifier. For a fair
comparison with the Maximum-Margin Tempotron learning rule, we use one output
neuron to represent each sound class and each neuron has a single desired output
spike. To determine the desired spike timing for each output neuron, we first present
all training spiking patterns from the corresponding sound class to the randomly
initialized SNN; and monitor the membrane potential trace of the desired output
neuron during the simulation. We note the time instant when the membrane potential
trace reaches its maximum (denoted as T max ) for each sound sample, revealing the
most discriminative local temporal feature. We then use the mean of T max across
all 20 training samples as the desired output spike time. As shown in Table 1, the
SNN trained with ReSuMe rule achieves a classification accuracy of 97.0%, which
is competitive with other models. This, therefore, demonstrates the compatibility
of features extracted by the SOM and spike-time based learning rules, whereby the
intra-class variability of sound samples is circumvented by SOM feature extraction
such that a single desired spike time for each class suffices.
J. Wu et al.
variability and high inter-class variability in both the BMU activation trajectories and
sequences, which are highly desirable for pattern classification.
We note that the time-warping problem exists in the BMU activation sequences,
whereby the duration of sensory stimuli fluctuates from sample to sample within the
same class. However, the SNN-based classifier is robust to such fluctuations as shown
in the classification results. The decision to fire for a classifying neuron is made based
on a time snippet of the spiking pattern; such is the nature of the single spike-based
temporal classifier. As long as the BMU activation sequence stays similar, duration
fluctuations of input sample will not affect the general trajectory of the membrane
potential in each output neuron; the right classification decision, therefore, can be
guaranteed.
To investigate whether the feature dimension reduction of the SOM is necessary
for the SNN classifier to learn different sound categories, we performed experiments that directly input the spike trains of the latency-encoded (20 neurons) [22]
or population-encoded (144 neurons) [23] mel-scaled filter bank outputs into the
SNN for classification. We find that the SNN classifier is unable to classify such
low-level spatiotemporal spike patterns, and only achieve 10.2% and 46.5% classification accuracy for latency- and population-encoded spike patterns, respectively. For
both latency- and population-encoded spike patterns, as all encoding neurons spike
in every sound frame, albeit with different timing, the synaptic weights therefore
either all strengthen or all weaken in the event of misclassification as defined in the
Tempotron learning rule. Such synchronized weight updates make it challenging for
the SNN classifier to find discriminative features embedded within the spike pattern.
The learning rules for the SNN can be categorized into either membrane-potential
based [25, 48, 49] or spike-time based [50, 51]; the Maximum-Margin Tempotron
learning rule belongs to the former. To study the synergy between the SOM-based
feature representation and spike-time based learning rule, we conducted an experiment using the ReSuMe [50] learning rule to train the SNN classifier. For a fair
comparison with the Maximum-Margin Tempotron learning rule, we use one output
neuron to represent each sound class and each neuron has a single desired output
spike. To determine the desired spike timing for each output neuron, we first present
all training spiking patterns from the corresponding sound class to the randomly
initialized SNN; and monitor the membrane potential trace of the desired output
neuron during the simulation. We note the time instant when the membrane potential
trace reaches its maximum (denoted as T max ) for each sound sample, revealing the
most discriminative local temporal feature. We then use the mean of T max across
all 20 training samples as the desired output spike time. As shown in Table 1, the
SNN trained with ReSuMe rule achieves a classification accuracy of 97.0%, which
is competitive with other models. This, therefore, demonstrates the compatibility
of features extracted by the SOM and spike-time based learning rules, whereby the
intra-class variability of sound samples is circumvented by SOM feature extraction
such that a single desired spike time for each class suffices.