Understanding the relationship between energy cost, cognitive activity, and performance is a key problem in psychophysiology. Prior studies have shown that solving cognitive tasks is accompanied by changes in the fractal dimension of EEG signals. In the present study, EEG was recorded from 68 participants solving a task in which participants matched graphs and tables. Participants identified which of three tables matched the numerical data presented in a graph. EEG signals were analyzed using the Higuchi fractal dimension algorithm. The results showed a short-term increase followed by a gradual decrease in EEG fractal dimension (F(4, 7656) = 36.594, p < 0.0001). We modeled this effect in the dynamics of neural network training. The first model was trained to classify 2 input values in the range from -π/2 to π/2. The target function was sinusoidal with further classification into two states: 0 or 1. At each training step, we summed the activations across all network neurons and computed the fractal dimension, which briefly increased and then gradually declined. The second model, based on ResNet18 and modified to solve the task of matching graphs and tables, showed a monotonic decrease in fractal dimension.