Machine learning is a promising approach for electroencephalographic (EEG) trials classification. Its efficiency is largely determined by the feature extraction and selection techniques reducing the dimensionality of input data. Dimensionality reduction is usually implemented via the mathematical approaches (e.g., principal component analysis, linear discriminant analysis, etc.) regardless of the origin of analyzed data. We hypothesize that since EEG features are determined by certain neurophysiological processes, they should have distinctive characteristics in spatiotemporal domain. If so, it is possible to specify the set of EEG principal features based on the prior knowledge about underlying neurophysiological processes. To test this hypothesis, we consider the classification of EEG trials related to the perception of ambiguous visual stimuli. We observe that EEG features, underlying the different ambiguous stimuli interpretations, are defined by the network properties of neuronal activity. Having analyzed functional neural interactions, we specify the brain area in which neural network architecture exhibits differences for different classes of EEG trials. We optimize the feedforward multilayer perceptron and develop a strategy for the training set selection to maximize the classification accuracy, being 85% when all channels are used. The revealed localization of the percept-related features allows about 95% accuracy, when the number of channels is reduced up to 90%. Obtained results can be used for classification of EEG trials associated with more complex cognitive tasks. Taking into account that cognitive activity is subserved by a distributed functional cortical network, its topological properties have to be considered when selecting optimal features for EEG trial classification.