SARAIVA, E. C.; http://lattes.cnpq.br/0429037888719725; SARAIVA, Eugênio de Carvalho.
Resumo:
There are a growing number of technologies that make use of classification algorithms for automating tasks. In particular, in Neuroscience, classification algorithms were used to
test hypotheses about the functioning of the central nervous system. However, the
relationship between the classes of patterns of neuronal activation in specific brain areas as a result of sensorial experience has received little attention. In the context of Computational Neuroscience , this paper presents an analysis of the level of similarity between classes of patterns of neuronal activation with the use of learning approaches unsupervised and semi - supervised in specific areas of rat brain in contact with objects , obtained during an experiment involving free exploration of objects by animals. The classes were defined according to certain treatments constructed with specific levels with set of 8 factors (Animal, Brain Region, Object or Pair of Objects, Clustering Algorithm, Metric, Bin, Window and Interval Contact). In total 327.680 treatments were analyzed. Hypotheses regarding the relationship of each of the factors with the existing level of similarity between treatments
were defined. The hypotheses were tested through between statistical distributions
representing each class tests. The tests applied where the tests for normality (Shapiro-Wilk,
QQ–plot), analysis of variance and a test for differences in central tendency (Kruskal-Wallis)
were performed. Based on the results found in studies using an unsupervised approach, it
was inferred that the process of acquisition and definition of patterns of activation by an
observer was not subject to a significant amount of noise caused by uncontrollable reasons.
For the semi-supervised approach, it was observed that not all degrees of similarity between
pairs of classes of objects are equal to a given treatment, which indicated that the similarity
between classes of patterns of neuronal activation is sensitive to all the factors analyzed and
provides evidence about the complexity of neuronal coding.