It got better for me when I started using more systematic approaches. Research papers in your problem area are actually super helpful for hints on hyperparameters. Transfer learning can be a godsend too if you're working with limited data on a problem that's been tackled before.