Systems, Man & Cybernetics - April 2017 - 13

of the experiment, a particular volunteer's
Table 1. The number of valid data samples observed
reactions toward the positive and negative
(per person per day and overall).
images and music are highly separable.
Moreover, there was no need to label the two
Number of Valid Data Samples
Individual
emotional states. By forming data clouds
Day 1
Day 2
Day 3
Day 4
Day 5
Overall
Number
(data partitioning) autonomously, the classi1
40/40*
38/40
40/37
40/40
40/40
198/197
fication was made by association with previous observable PV data. In this way, the
2
40/40
40/40
40/40
38/31
40/40
198/191
classifier is truly autonomous and reflects
3
40/40
40/40
38/39
40/39
39/40
197/200
the subconscious nature of the learning taking place.
4
39/40
36/40
38/40
39/36
37/40
191/196
We conducted additional experiments
5
0/0
40/40
38/33
39/39
40/40
157/153
to test and analyze the results to deter6
40/40
0/0
32/40
40/40
40/40
152/160
mine whether we had achieved what we
had hoped, i.e., automatically clustering
7
40/40
40/39
39/40
40/40
40/40
199/199
the observable PVs into emotions and pre8
40/37
39/40
39/40
40/40
40/40
198/197
dicting what emotions new, unseen images
would most likely evoke. As we have stres9
40/40
40/40
40/40
40/40
40/40
200/200
sed in our positive and negative stimuli,
10
40/27
39/40
39/40
40/40
40/38
199/175
we can consider the two PV pairs that are
*number of readings in positive/negative sets.
clustered into different data clouds or
clusters. For this purpose, we built a fully
autonomous AnYa-type self-learning FRB
zero-order classifier per person per day,
and learned to classify each individual's
Negative
Positive
perceptions. The actual human-intelligiImages
Images
ble labels (i.e., positive and negative)
require cognitive feedback (see Figure 1)
and are part of the second stage of the twolayer hierarchical structure/architecture
(see Figure 2).
We started by considering the first pair
of observable PV data points, corresponding to an image and a piece of music from
each positive and negative category, as the Figure 10. Some of the images used in the experiments.
sole training data samples. In this way, we
The focal points of the data clouds were formed by
considered a scenario per person per day whereby the
self-learning. Initially, the first pair or 2-D points of PV
classifier automatically gave every data sample a label of 1
data for each data cloud were used, and then these
or 2, corresponding to the two positive and negative imagpoints were automatically updated. The focal points are
es and music pieces for which the emotional state (and,
critically important and sufficient
consequently, the class label) was
to form data clouds, because they
different. The labels are related to
are formed like a Voronoi tessellawhat we think about these groups.
There are many
tion by assigning each new data
We gathered similar emotional
studies that report
point to the nearest focal point,
reactions into clusters or groups
which means that the system is
and later stated that "the current
that, during
self- learning, reflecting the subemotional reaction is similar to A
sleep, people
conscious nature of human learnor B," where A and B are concepts
ing. Notice that we do not need to
or data clouds or cluster-like
(unconsciously) revisit
know the exact label (positive or
groupings that can be created suband reorganize the
negative, good or bad, and so forth).
consciously and thus autonomousWe simply need to know that they
raw data, and various
ly. As such, the classifier fully
are different. So the scenario is
autonomously predicted into which
outliers come to their
fully autonomous. It is per person,
of the original two classes the perattention.
per day, meaning that it is not a
son's reaction to any future image
universal one-size-fits-all approach.
would fall.
Ap ri l 2017

IEEE SyStEmS, man, & CybErnEtICS magazInE

Table of Contents for the Digital Edition of Systems, Man & Cybernetics - April 2017