In the ten years or so that I was taking piano lessons, I caused my teachers a lot of frustration by how slow I was to learn new pieces—one even called me the bane of her existence. It wasn’t that I didn’t practice; I was all too happy to practice a piece once I’d figured out how to play it. No, it was the initial translating of musical notation into finger movements that took me so long. A typical measure of piano music has several notes to be played by multiple fingers on each hand, and the process of working out which finger needed to play which note (and when) never seemed to get any easier over the years, even though I was mastering a growing body of more and more complicated pieces.
For some reason, the pieces I already knew weren’t helping me learn new ones. This reminds me of the machine learning concept of overfitting: a computer training to, say, recognize handwritten letters might do a very good job on the samples it trained with, but not as well with handwriting samples it’s never seen before. Often, the reason is that the computer is paying too much attention to irrelevant details of the training samples in order to tell them apart, and the more complicated a rule the computer invents for itself, the less likely it is to be generally correct. In a way, that’s what was going on with my piano practicing: I’d painstakingly learn to play each piece as it came along, but I wasn’t learning any generalizable principles.
But not every piano player has this difficulty! Sight reading is the skill of being able to pick up some music and just play it; the musical equivalent of touch-typing as opposed to hunt-and-peck. Whenever I see a movie about some would-be singer bringing in her audition piece and being so jaw-droppingly good that she’s cast as the star, I want to stop the movie at the moment where she hands her sheet music to the piano accompanist and say “Look, he doesn’t even need to practice! That’s talent! That’s who I want to be!” But how can I get there?
The paradox is that I need to practice a lot of music, but if I practice any piece too much I’ll be working out how to play it and not music in general. So here’s my attempt at a solution: Every day I get out our hymnal, whose four-part harmonies are just outside my current ability to play impromptu, open a random number generator, and choose a random hymn to play through once. I try to play at an even tempo, not hesitating at tricky bits or going back to fix mistakes. Then, as much as I long to play it again now that I understand its tricky spots better, I choose a new random hymn and keep going. That way, instead of getting better at playing particular note sequences, I’m getting better at the general skills of looking ahead and absorbing as much of the essential information as I can at a glance.
I do feel like I’m already improving, so the worst-case scenario at this point is that I’m only training myself to be able to sight-read hymns, and not more general music. On the one hand, even if that’s true I imagine some of the skill will be transferable, and I can always start a more complex sight-reading-regimen if I feel like it becomes necessary. On the other, being able to sight read hymns is a valuable skill in itself, and one I’m happy to start working on first.
This post is making me wonder what some other things are that I am possibly overfitting with my practice. If I want a better understanding of how food flavors combine, maybe I should work my way through a book of varied recipes instead of just continually trying to optimize my favorite meals. Or if I want to make better impressions when I meet someone new, perhaps instead of stressing about making that particular conversation go as well as possible I should focus on general conversation skills, like remembering the person’s name and finding things we have in common. How about you?
6 thoughts on “Practice and overfitting”
Now there’s an approach to sight reading that’s really cool. I might try that in the fall when I go back to having access to the piano at the school.
It’s been a few weeks, and I’m definitely noticing improvement! I’m also aware now of variants I can do to make it harder: try to play through without looking at my hands, try to sing along with the words as I play (which is in turn easier or harder if I know the melody or not), etc.
Owen, my piano teacher said to play everything very slowly and not speed up until I knew what to do, and not to always start at the beginning. I’d like to play again, and maybe take lessons from another teacher. I’m interested that this technique is so different from what she recommended and yet is working.
This a great comparison of machine learning and human learning. But I don’t think you are applying the analogy far enough. I think great sight readers have internalized the basics of music theory, either intuitively or through studying. Study music theory gives you the basic, general principles that can be applied to most western music. To start with I would suggest learning about chords and scales. Then try analyzing a piece of music before you play it, in terms of chords and scales (starting with hymns like you have already done). You don’t even need super formal music theory knowledge. Even just recognizing that, “Oh, when I play this bass note, then I usually play these other notes in some configuration,” will help you go a long way to sight reading.
Hi JMK, thanks! You’re absolutely right; I’ve noticed that it’s easier to play the hymns whose chords I can recognize more easily. (I suppose it takes more effort for my brain to tell my fingers “Play these G and D keys in the left hand, and the B, D, and G keys in the right hand” than to say “Play the G major chord with the keys you’re already near.”) I’m also picking up on some patterns I don’t have words for, too, so maybe that’ll snap into place when I get there in my music theory book.