TAVARES NETO, A. V.; http://lattes.cnpq.br/9306851808731924; TAVARES NETO, Antonio Victor.
Abstract:
Within the context of the efforts to understand the nervous system, the acquisition and
analysis of information from neurons coexist with the possibility of the occurrence of numerous
interferences, whether environmental, instrumental, or related to the complexity of
neuronal connections, which could raise doubts about the accuracy of the results. In this
dissertation, an analysis was conducted relating the possible impact of annotation errors in
the animal activities for the classification of complex objects in the brain (more specificaly
in the primary cortex and hippocampus). To this end, we simulated annotation errors, which
map a temporal sequence of neuronal activations (obtained by microelectrode implants) to
the contact intervals of an animal with external stimulus (objects), changing the original annotation
values, with the purpose of evaluating the effect of such errors on the quality of
pattern classification, when considering a set of computational techniques. The area under
the ROC (AUROC - Area under the Receiver Operating Characteristic) was employed as a
measure of the classification quality.
As the main contribution of this dissertation, it was shown that the variation of the contact
intervals between animals and objects has a direct influence in the quality of classification.
However, such influence does not happen abruptly, allowing variations of less than two seconds
to retain the properties of classification and quality of the original experiment without
variations or errors in the contact intervals.