We used backward masking to investigate the influence of pattern complexity on recognition accuracy in tachistoscopic experiments on human subjects. A rectangular frame of twelve line segments of equal length was used as a masking pattern (MP). The test patterns (TPs), which were nonverbal figures, were the parts of MPs composed by joining four, five, six, seven, and eight line segments. The TP was presented for 10 ms, and then the MP for 500 ms. The interstimulus interval was established for each subject individually as the shortest interval permitting a recognition accuracy of 50%-90%. Each subject took part in six to eight test trials. This approach enabled us to evaluate the role of learning. In the first test trial the smaller the number of line segments the pattern consisted of, the more accurately it was recognised by the subject. In the second and third test trials the subject recognised the figures which differed by number of line segments with more or less the same accuracy. In the subsequent trials the figures composed of six line segments were recognised more accurately then ones composed of four, five, seven, or eight line segments. The different results of the initial test trial and the subsequent ones led to the conclusion that the role of the features by which the figures were recognised had changed in the training process. The inexperienced subjects recognised the TPs by simple features (the number of line segments seemed to be the most important one) and the experienced subjects recognised them by more complex features.