Micro feedback loops and learning

post by Swimmer963 (Miranda Dixon-Luinenburg) (Swimmer963) · 2019-05-26T00:50:36.202Z · LW · GW · 12 comments

Contents

  A high-tech feedback loop
  Rewardingness
  How does this generalize? 
  Implications
None
12 comments

Tldr: some not-particularly-ordered thoughts about how learning works in humans, via the example of singing.

A high-tech feedback loop

I recently discovered Singer’sStudio, an iPhone app for voice training that is approximately the Best Thing Ever (h/t Raemon [LW · GW]). It’s a work of pedagogical art, and describing what it does right pulls together various nebulous thoughts I’ve had about learning and developing intuitions for a particular domain.

How the app works: it gives me various singing exercises, and then tells me in real-time, via a pretty graph, if I’m singing on pitch. That’s it.

Okay, it does have a few other features, like:


But I’m pretty confident that the pitch contour is the most important part, and an app that only did that thing would be almost as good on the skill-building side.

There are obviously a lot of singing sub-skills other than hitting notes, like vocal timbre, head versus chest voice, breathing from the diaphragm, enunciating consonants clearly, and singing vowels that don’t sound weird. A voice teacher would be able to comment on these directly.

Still, in practice the app ends up causing me to improve on a bunch of these sub-skills as well, to the extent that getting them wrong results in singing off-pitch. The instantaneous feedback is doing a lot of work here – if the app only gave me a score at the end of each exercise, I predict I would end up quite stuck on what to change up in order to do better. However, I’ve ended up e.g. remembering to breathe from my diaphragm, because not doing that reliably resulted in my voice being wobbly, which was obvious on my pitch contour. (It also has instructions for breathing exercises as part of each warm-up, but I only actually did those exercises once or twice; they are boring compared to singing and getting pretty graphs.)

It’s amazing how fast my brain learned to turn a smooth, stair-like pitch contour into dopamine hits, and feel a flicker of pain every time I jumped above or below the note before finding it. I end up very motivated to play around with any variables that make the graph prettier.

Rewardingness

It occurs to me that another feature, which I think is less key than the instantaneous pitch contour but still pretty important, is how it slightly gamifies the entire thing, and thus makes it addictive. There are % scores at the end of the exercise! And points! It logs my all-time high score for each exercise so I can try to beat it! It also logs how many minutes a day I’ve practiced. All of this makes me more likely to use it, and actually putting in the time is a key part of training any skill.

It’s also cool that I can listen to my voice and, in addition to catching mistakes (ouch!), notice when hey, wow, I actually sounded good there. This lets me gradually figure out what correlates with liking how my voice sounds, and it also gives me a warm glow of satisfaction and helps me feel like a Real Singer.

(I’m not sure if the app is cheating by doing some kind of post-processing on these recordings to make them sound pretty. Normally I hate recordings of myself speaking, let alone singing. Still, I’ll take it.)

It also seems relevant that there’s zero embarrassment factor – this isn’t a human watching me and judging me for daring to sing when I kind of suck. I’m pretty shameless, as humans go, but I’ve been slightly nervous with every voice teacher I’ve had – they’re an expert! they’re probably really unimpressed! – and tension is not good for singing well. With this, I can sing my heart out in the privacy of my apartment, and even feel safe experimenting.

How does this generalize?

This is a post about singing, but it’s also a post about learning skills in general.

Learning to sing (or to play the piano, tie one’s shoes, draw, dance, swim; all the things commonly known as procedural memory) isn’t like memorizing a list of dates for a history test. There are some steps that can usefully happen in the explicit verbal loop, like “remember to breathe from the diaphragm”, but the end goal is that basically nothing is being held in working memory, and everything happens on the level of microsecond-to-second intuitions and muscle memory.

I think there are a really large number of skills, physical and mental, that are somewhere on a gradient between this and explicit memorization. Chess seems like a good example; it’s possible to play it from knowledge of the rules and explicit reasoning/strategizing, and there’s always some of this, but analyzing the entire game explicitly is intractable, and from what I’ve heard, experts heavily rely on intuition. ICU nursing and running events, both things I have more personal inside knowledge of, are maybe a bit closer to the intuitive or procedural end; the time pressure and urgency, and the sheer amount of stuff going on, makes explicitly reasoning through each decision less useful. Something like accounting is further towards the explicit end, but even that gets a lot more automatic with practice, and intuition is what makes something jump out as “wrong” to me.

Practicing a procedural skill without the benefit of a clever app involves two steps – trying it out (controlling my vocal chords to sing a set of notes), and being able to evaluate how good or bad a given attempt was (my “ear”). Becoming an expert involves improving both of these functions; at any given moment, in order for an instance of practicing to lead in the direction of progress, the evaluation process needs to be more accurate than the execution process.

I’ve heard anecdotes about the phenomenon of having discernment for “good X” that far exceeds one’s ability to perform X, and I’ve occasionally experienced this for other things, like writing. Therein lies frustration and embarrassment and self-judgement. In my experience, this often causes people to avoid practicing a skill, even though really it’s a good thing if your taste for X is refined enough to notice your own mistakes – that's how you know what to do differently the next time!

With singing, past a certain level I struggle to improve my ear enough to notice mistakes, which limits how much progress I can make just singing by myself at home. Voice teachers and other sources of feedback can help, but this is slower and thus less useful for shaping, in the same way that giving a dog a treat with a 5-min delay is going to be much less effective for training purposes.

This app creates a feedback loop on the sub-second level, using a different channel (my visual cortex) that doesn’t interfere with singing or listening, and which is a lot more accurate than my own internal sense for whether I’m on pitch. My hypothesis is that some lucky people do have this sort of high-quality discernment for pitch already developed in their brains, or develop it young, and they’re the ones who learn how to sing without much effort – and, because singing badly is tied up in embarrassment and shame, they’re probably the people who tend to become excellent singers at all. I expect most people with my level of innate talent, or lack thereof, just give up.

The app also lets me test and calibrate my sense of pitch more directly, so that maybe, someday, I’ll be able to tell in real-time, on my own, if I’m landing the notes in a song. (Currently I cheat and use VocalPitchMonitor for this, since SingerStudio isn't quite cool enough to let me upload random sheet music and use that to get feedback on real songs.) I play myself back and stare at the pitch contour, and try to hear the slight wrongness when the line indicates I’m just barely sharp or flat; it’s better than previous “ear training” I’ve done with a teacher or out of a book, and I’m hopeful.

Implications

This app is a really cool category of thing, that’s only possible at all due to fairly recent technological advances, and there are probably a ton more instances that I don’t know about.

I’m curious where else this has been explored. Singing may be an easy case, because measuring a single straightforward variable, pitch, gets you so far. I can imagine an app that trains, say, krav maga fighting techniques, via video analysis and/or accelerometer data, but I’m not sure that’s possible yet given current tech.

It has me thinking about other pedagogical techniques, though. Martial arts teachers will shout real-time feedback at you ("turn your hips more! get your knee higher!") I’ve taught swimming, and one issue is that waiting until a swimmer finishes a lap before giving any feedback introduces a huge delay, but grabbing onto them every time they do something slightly wrong is incredibly irritating and disruptive. Now I’m imagining giving them waterproof headphones and narrating the feedback in real time (“elbow higher please”, “roll your shoulder deeper into the water”, “keep your head back when you breathe”, etc etc.) This would be so cool.

My friend, when I brought this up, recommended I look up Tagteach. It seems to be largely based on clicker training, but lays out a bunch of thoughts I’ve had on pedagogy, like the importance of breaking a skill down into really small increments where success is easily measurable. Their site claims this protocol has been used in dance and sports coaching, but also business skills training and medical school; I’d be curious to know what this actually means in practice, but it does hint that it could be useful beyond purely physical skills (though presumably it requires success to be easily visible to the trainer; clicker-training accounting, or anything where a lot of the process is mental, seems hard!)

Of course, one of the awesome things about my app is that it removes the need for costly one-on-one time with an instructor. For swimming, could an accelerometer measure forward speed and narrate that? Speed relies on getting an absurdly large number of muscle movements just right, but the same thing is true of singing on pitch, and the single-input feedback seems to be enough to guide me towards progress. As a bonus, it could play generic reminders and prompts (I’m assuming the Singer’sStudio prompts aren’t responsive to specific mistakes I make, but they’re still useful.)

It also occurs to me that if getting direct and immediate feedback on your evaluation process is key to improving it, it could be useful to focus on that directly, separately from executing the skill itself – if you’re better at catching your own mistakes, later practice will be more valuable. I’m imagining a gymnast watching videos of Olympic gymnastics that have real-time commentary analyzing what the athletes are doing, noting successes and mistakes.

I would be curious to hear any examples others have of this.

12 comments

Comments sorted by top scores.

comment by Puredoxyk · 2019-05-26T12:48:44.245Z · LW(p) · GW(p)

I have an example of this! I'm also in total agreement with you about what makes the app cool, and it's partly because I immediately related it to the following.

I'm a martial artist, and I teach newbies regularly as part of my training, and a friend I was teaching told me I had to read a book about tennis, which *really* confused me. (I also thought at first that he meant Infinite Jest, but he didn't. :P) The book is The Inner Game of Tennis, and it was a huge hit in the 80's that spawned a lot of useless, ignorable, X-For-Dummies-type spinoffs. But the original is one of the best books on learning I've ever read period, and it both confirmed and revolutionized everything I knew about learning and teaching martial arts (and swimming and singing, which are also interests of mine -- hello, friend!).

The book is by a dude who trains tennis players, who's played his whole life, and who had SILLY successes with total noobs and never-been-athletic people, and finally wrote down how he was doing it. And there's a ton of great detail in the book about specific methods and tiny brain things that, for the most part, it's very easy to translate out of tennis and into a million other things (the tendency for spinoffs to happen makes sense here, at least); but the main crux of it is almost exactly what you're describing with the app: Give the brain "what it feels like to do it right", and then let it constantly compare what you're doing to that, in real-time, and adjust all the little sub-skills however it needs to to get the right result. Everything else is just sauce; Practice is the brain making that in-the-moment comparison and learning to adjust for it, and Practice is everything to developing skill. Thus, by focusing directly on training people to feel the difference, in the now, when hitting the ball, he trained them in the one fundamental of tennis that would give them the fastest possible access to actual expertise: How to Practice.

Anyway, I think I've said enough; you should definitely check out the book if you're interested in further information on this lesson about learning; and I'm super geeked that there's a singing app that uses this type of feedback, it sounds like, to excellent effect! I'm also equally annoyed that it isn't for Android, but that's not your fault. :) Thanks for the awesome post!

Replies from: Swimmer963
comment by Swimmer963 (Miranda Dixon-Luinenburg) (Swimmer963) · 2019-05-26T15:46:25.025Z · LW(p) · GW(p)

I've had The Inner Game of Tennis on my recommended list for several years, as a general rationality-related book; clearly I should just go ahead and read it!

I did a quick Google for you, and it looks like there are a lot of apps for Android that are similar (though I can't speak to which of them are actually good).


Replies from: Jonathan_Graehl
comment by Jonathan_Graehl · 2019-05-28T22:41:31.024Z · LW(p) · GW(p)

It's not a good book, except it says: focus/visualize on the result you want, not on verbal coaching cues. That can be good advice, but I object: when there's time+space to consciously plan some part of your gameplay, verbal (self-)coaching is fine.

comment by Noah Topper (noah-topper) · 2019-05-30T00:54:16.131Z · LW(p) · GW(p)

I don't have a wonderful example of this insta-feedback (which definitely sounds ideal for learning), but I've gotten annoyed lately with any math book that doesn't have exercises. Some of the books on MIRI's Research Guide list are like this, and it really boggles my mind how anyone could learn math from a book when they don't have anything to practice with. So I'm getting more selective.

Even some books with exercises are super hard, and really don't have any kind of walkthrough process. AI: A Modern Approach is a critically-acclaimed textbook, but has little to no build up in the difficulty of its exercises, and little to help you if you get lost. Right now I'm reading How to Prove It, which is *super* good. The whole book is one big walkthrough on mathematical proof, down to how to organize your scratch work. It has tons of exercises with varying difficulty, with some answers and hints. It's much better feedback, and is helping me a lot more, although the material is comparatively simple.

comment by Raemon · 2019-05-26T19:34:39.989Z · LW(p) · GW(p)
But I’m pretty confident that the pitch contour is the most important part, and an app that only did that thing would be almost as good on the skill-building side.

BTW, my girlfriend (who originally recommended this app to me after trying a bunch of singing apps), was actually annoyed at all the bells and whistles that the app came with, and recently build a version that just did the "what pitch are you singing? draw a graph" part.

And... last week she went ahead and made it: It's called Simple Pitch. So you can try it out and see how useful it seems compared to the other one.

Replies from: cousin_it
comment by cousin_it · 2019-05-28T12:18:30.390Z · LW(p) · GW(p)

It's in equal temperament, right? Have you thought about using just intonation?

I had always bought the story that the two are close enough for most musical purposes, but a few weeks ago I unfretted my guitar and started playing music in JI, and it's night and day. For example, in ET the major third from C to E sounds kind of restless, while in JI it's peaceful with all harmonics overlapping as they should (4:5 frequency ratio). Same for the minor third (5:6), hearing the JI interval makes me feel like "ah, so that's what the ET interval was trying to hint at". And then there's the harmonic seventh chord (4:5:6:7), which sounds very musical but can't be imitated on an ET guitar or piano at all.

comment by Gordon Seidoh Worley (gworley) · 2019-05-28T20:52:28.722Z · LW(p) · GW(p)

To give an additional example of tight feedback loops being helpful, I've been taking Alexander lessons for nearly a year. Each lesson consists of 30 minutes of me doing movements (although sometime the "movement" is holding a posture, like sitting, standing, standing on toes, or crouching) and 30 minutes of "table time", i.e. I lay on a massage table while my teacher users her hands to very subtly suggest changes to my posture. Although I could go on about how great this has been and how much value I get from it, what I mostly want to say about it here is that it depends very much on tight feedback loops to perform a kind of reinforcement learning. As I make a movement she uses her hands and some taught jargon (part of the technique involves associating jargon with postures and movements so you can easily call them up on command by saying or thinking the jargon) to adjust what I do, giving me rapid feedback on how I'm doing. The result was that within the first 10 hours of training I dramatically improved my posture and reduced posture and movement related pain.

For comparison, overlapping with learning Alexander technique I've been more deeply practicing formal meditation, and learning formal meditation has very long feedback loops and requires months to make significant progress. Now, maybe the long feedback cycles are not why it takes months to make progress, and I can think of reasonable stories as to why that would be, I can also imagine finding ways to shorten feedback cycles would have made progress much faster. For example, when I've done biofeedback stuff in the past it only took 4 or 5 hours of sessions before I could make myself fall asleep at will (sadly I've forgotten how to do this), and I think it's quite likely that it was helped a lot by having a computer telling me when I got a little closer to what I needed to do to make that happen and when I got a little farther away, such that I didn't have to spend as much time guessing and waiting for strong evidence that I was doing the right thing before I could reliably train that ability and then go on to the next step.

comment by Matt Reyes (matt-reyes) · 2019-05-26T04:01:00.886Z · LW(p) · GW(p)

Re: Skills that can be learned in fast feedback loops via app: Basketball shooting, especially free throw shooting, depends on a small number of variables, with release angle being the most important(optimal release angle gives the widest range of acceptible velocities). There is an iphone app that tracks shots with compiter vision and gives you all these numbers.

Replies from: Swimmer963
comment by Swimmer963 (Miranda Dixon-Luinenburg) (Swimmer963) · 2019-05-26T17:30:41.843Z · LW(p) · GW(p)

Wow! That...is not a thing that I would have expected to exist. I guess it makes sense, if it in fact depends on relatively few variables.

comment by Jason Gross (jason-gross) · 2019-05-26T02:07:42.471Z · LW(p) · GW(p)

I imagine one thing that's important to learning through this app, which I think may be under-emphasised here, is that the feedback allows for mindful play as a way of engaging. I imagine I can approach the pretty graph with curiosity: "what does it look like if I do this? What about this?" I imagine that an app which replaced the pretty graphs with just the words "GOOD" and "BAD" would neither be as enjoyable nor as effective (though I have no data on this).

comment by CronoDAS · 2019-05-26T03:17:45.368Z · LW(p) · GW(p)

You can get a lot of very tight feedback loops when playing video games, too...

comment by gruyere · 2019-05-31T04:33:35.268Z · LW(p) · GW(p)

The instant feedback concept works for low-tech measurement games too.

I like to guess the time then look at a clock to see if I was accurate. My hope is that it improves my sense of time, but even if it doesn't, it's a fun diversion.

And Shigeru Miyamoto (the Nintendo boss) says: "I've always enjoyed guessing the lengths of objects, which is why I carry a measuring tape around with me... For instance, I might guess that the table in front of us is about 1.2 metres long. Then I'd actually measure it with the measuring tape to check. If I got it right, I'd think: 'I'm on form today!' But if I missed the mark by a long way, I'd think: 'I've been slipping a bit recently!'"