Two category-learning experiments were conducted to examine the role of category structure and learning regime in category learning. We particularly focused on effects of these factors on selective attention, which was measured by eye-tracking methods. Results show that even though supervision was weaker than in previous studies, attention optimization and cost of attention were observed during category learning (Experiment 1). Moreover, there were faster learning and stronger attention optimization when statistically denser categories were learned (Experiment 2). At the same time, there were weaker costs of selective attention when learning denser categories than when learning sparser categories. Results are discussed in relation to theories of category learning.