![](https://csdnimg.cn/release/download_crawler_static/16681971/bg4.jpg)
with diversive curiosity (DCPROP) in their previously
published paper, by integrating the mechanism of diversive
curiosity with BP learning algorithm for solving its ‘‘flat
spot’’ problem and for helping it escape from local minima.
Moreover, Ioannou et al. [6] have realized emotion
recognition through facial expression analysis based on a
neurofuzzy network. Shi et al. [7] have proposed a hier-
archical processed frame construction of artificial emotion
model for intelligent system according to the basis con-
clusion of emotional psychology. Bhotti et al. [8] have
presented a language-independent emotion recognition
system in their paper for the identification of human
affective state in the speech signal. Kishore et al. [9] have
described a novel hybrid emotional neural network for
classification of emotions from facial expressions.
Besides, Khanchandani et al. [10] have explored per-
formance analysis of multilayer perceptron neural network
and generalized feedfor ward neural network for detection
of seven human emotions, i.e., neutral, anger, boredom,
disgust, fear, happiness, and sadness, using speech signals.
Dai et al. [11] have used landmark and other acoustic
features to recognize different emotional states in speech.
Me
´
riau et al. [12] have investigated how changes in the
functional magnetic resonance imaging signal during per-
ceptual decision making on facial stimuli covary with
individual differences in the ability to identify and com-
municate one’s emotional state.
In one of their published wor ks [3], Khashman presented
a modified backpropagation (BP) learning algorithm,
namely the emotional backpropagation (EmBP) learning
algorithm. Their new algorithm has additional emotional
weights that are updated using two additional emotional
parameters: anxiety and confidence. Their EmBP neural
network was firstly implemented to a facial recognition
problem. And experimental results have show n that the
addition of the two novel emotional parameters improves
the performance of the neural networ k yielding higher
recognition rates and faster recognition speed.
Khashman et al. [13] have also applied their EmBP
algorithm to a blood cell-type identification problem.
Experimental results showed that the additional emotional
parameters and weights improved the identification rate as
well as the classification speed. And they further applied
EmBP algorithm to a credit risk evaluation problem [14].
Their experimental results have suggested that both emo-
tional and conventional neural networks can be used
effectively for credit risk evaluations; however, the emo-
tional models outperform their conventional counterp arts
in decision-making speed and accuracy, thus making them
ideal for implementation in fast automatic processing of
credit applications.
However, in their work, the input values to the emo-
tional neurons of both hidden and output layers of the
model are identically assigned as the global average value
of the specific input image pattern under processing. Their
reasons for such an evaluation are mainly accor ding to a
statement of Baumgartner et al., ‘‘Most of the publishe d
neuroimaging papers examining emotional processes have
used visual stimuli in order to evoke emotions [15].’’
Therefore, the global average of the input pattern is fed into
the emotional neurons of hidden and output layers of the
model, and the resulting total potentials are used to cal-
culate the outputs of the two layers, respectively. Through
this pattern averaging, the author attempts to mimic the
tendency of human emotional judgments and preferences
to be based on general impressions rather than precise
details of the objects being perceived [3, 13, 14].
However, during our theoretical and experimental
studies to the EmBP algorithm [3, 13, 14], we found that
this assignment of emotional input values is not reasonable
and may not be enough desirable for the successful
implementation of the whole algorithm. Fo r a simple
counterexample, there might exist two input images, which
are observably different from each other, and belong to two
different persons, but their global average values may be
rather close to each other. In this kind of cases, the
implementation of EmBP might abate.
Enlightened by the above observations, we develop a
novel self-organizing map (SOM) [
16]-based emotional
(EmSOM) neural network learning algorithm of our own.
In the EmSOM algorithm, the input values to the emotional
neurons are not simply assigned as the global average value
of the input pattern. Instead, the unsupervised SOM
learning algorithm [16] is employed to compute these input
values. Specifically, first , a SOM network associated with
the hidden layer of the model is generated taking all the
input patterns as its training data. Then, for a specific input
pattern under processing, the winni ng element of this
generated SOM is found out, and its average value of
weights with respect to its dimension is fed into the emo-
tional neuron of the hidden layer of the EmSO M model,
and the resulting total potential is used to compute the
output of the hidden layer.
Correspondingly, another SOM network associated with
the outpu t layer of EmSOM is generated taking all the
activation outputs from the hidden layer of EmSOM as its
training data. And then, accordingly, for a specific input
pattern under processing, the winni ng element of this
generated SOM is selected which is the closest to the
specific activation output vector of the hidden layer of
EmSOM. And then, its average value of weights with
respect to its dimension is fed into the emotional neuron of
the output layer of EmSOM, and the resulting total
potential is used to compute the final output.
Our motivations behind the design of EmSOM algo-
rithm are as follows: Just as what we have mentioned
Neural Comput & Applic
123