CNN variants can reach near-optimal learning capacity by analyzing data variation, saving computational time without validation data.