being yourself takes practice

• Related to:
  • generative vs discriminative modeling
  • The difficulty of learning new skills

I remember learning to play the violin when I grew up. I knew how I wanted the piece to sound. I could even hear it in my head. But when I tried to play it, it comes out all jumbled. I knew the music that I wanted, but I hadn't trained every piece of myself, parts of my brain that control the infinitesimal muscle movements and my fingers to move in just the right way to produce that music. I didn't even know what that right way would be until I kept trying.

Now, people often give the advice, "just be yourself". I found this frustrating as a kid, because it seemed almost tautological and meaningless. Who else am I going to be, other than myself? And what does it even mean to be myself? As a kid I didn't really even feel that I knew what my self was. And as a kid, you don't know. You don't necessarily know what style of music is going to play in your head for the rest of your life yet. But as you keep growing you will develop a self.

Here by 'self' I mean what distinguishes you from other people. It's all of the little patterns, and how you think, the conceptual scaffolding that you use. What little things in your life you like and don't like; all of your preferences. All of the things that you've learned from your own experiences. Since your experiences are incredibly different from anybody else's you will have learned things that are in some ways incredibly different from everybody else. And those things that you've learned will give you opinions about the world that you don't see out in public. They might help entrench some sort of deep seated values. There might be some logic by which some set of principles seems really consistent with some other set of principles in your mind, in a way that hasn't really coalesced out in the world yet. And so to whatever extent, you develop a different inner self than other people. And that will only happen over time.

But to whatever extent you develop a different inner self will require you to put on a novel performance in the world. Now when I say performance I don't mean something artificial, necessarily. You don't want to play somebody else's music. I mean that other people only see your actions in the world: what you say, what you do, so your appearance in the world is just your pattern of action, which we can call a 'performance'. And the best performances are the ones where the player puts their whole heart into it, where there's an immediate connection and immediate relation between your inner self, your thoughts, and the way that you behave in the world. If that's not true---if your behavior in the world is not in line with your inner self, if you feel like you're you're actually being artificial for your public, if you're always calculating reasoning about what your outer face should do---then you are in some ways, isolated and alienated. You weren't showing anybody else your true self. And no one else will be able to love you for your true self.

So there's a concept, there's a distinction in machine learning, between generative vs discriminative modeling. A discriminative model can recognize a particular phenomenon. It looks at that phenomenon. It takes some observation of the world, and analyzes it reduces it to a simpler explanation: a single label, number, or even something more structured like a segmentation or a parse tree. If you've listened to music, and you know that the Chicago Symphony is better than your high school symphony, that's a discriminative model.

Discriminative models can be very useful; we often have great need for them, we need to understand what's out there in the world. I found that in my life I've produced discriminative models of many things that to some extent makeup myself. I have a discriminative model of political opinions: if I hear someone making an argument, I can usually tell you whether I agree with it. I have a discriminative model of workplaces: if I'm at a company. I can usually tell you whether I enjoy it. And I have many discriminative models of music, of art, of literature. As a student, I especially have discriminative bottles of ideas. When I read a research paper, I can generally tell whether that I have an opinion on whether that research paper is good, whether it contains interesting ideas.

Now one of the common pieces of advice given to PhD students is that your first work will be bad. It won't be as good as the work that you respect and the work that you imagine doing in your mind. That is very good advice, but it applies more broadly than just research for any work. You will learn to evaluate to criticize much more quickly and much more easily than you will learn to generate: multiple choice tests are a lot easier than free response. Judging figure skaters is a lot easier than being a figure skater.

Now, this generative and discriminative distinction doesn't come up just in machine learning. It takes practice to be yourself, to really generate yourself, separate from being able to judge yourself. It's a lot harder to actually come up with the actions that express yourself in the world than it is to recognize when it happens. Sometimes you see someone saying something and you think "they're saying exactly what I thought, but they're saying it much better than I could have". That's a recognition that what you have is a discriminative model while that person has a good generative model.

There are some ideas that are easy to articulate, because you see people out there, articulating them all the time (most learning is by demonstration). Articulating a new idea isn't automatic. And it's not easy. It's like learning to play the music you hear in your head. It is genuinely computationally hard to solve for the thousands of muscle movements that will satisfy your inner critic.