Multi-Modal Classifier-Fusion for the Recognition of Emotions - Coverbal Synchrony in Human-Machine Interaction - page 77

Figure 2. Layered (mid-level) classifi cation architecture to recognize dispositions in human-
companion interaction. The level of abstraction increases in each layer to obtain high-level
symbolic information.
such as activity recognition (Glodek et al., 2012), EEG analysis (Schels
et al., 2011) and classification of bio-acoustic signals (Dietrich et al.,
2003) to mention just a few examples. There are different techniques
in the literature to attain diverse ensemble classifiers. The individual
classifiers can, for example, be trained on different subsets of the
training data (Breiman, 1996). Another way is to conduct multiple
training runs on the data using different base models or different
configurations of a model (model averaging). Furthermore, different
subsets of the available feature space (so-called feature views) are
often used to construct individual classifiers, which are then treated
as independent data streams.
In order to formally reflect that the accuracies of individual
classifiers in real-world scenarios and especially in non-acted affective
computing are generally low, it is useful to implement mechanisms
to increase the robustness and to assess the quality of a decision for a
sample. While the robustness can be achieved using the aforementioned
ensembles of classifiers, the self-assessment of classifiers can be
obtained by defining an appropriate uncertainty measure. Common
ways to establish uncertainty measures are to use probabilistic or
fuzzy classifiers, or to use the degree of agreement of an ensemble of
classifiers, i.e. the more the individual classifiers agree on a specific
value or label, the more confident this decision can be seen as.
When combining multiple decisions, the uncertainty can be used as
a weight in the fusion (Glodek et al., 2012).
Especially in real-world scenarios, it has been proven to be
successful to stabilize weak decisions by integrating individual results
over time (Glodek et al., 2012). Hereby, the confidence of the classifier
can also help to assess weak decisions. This integration could also
slow down the sample rates to match the sample rates of the sensory
channels.