Micro feedback loops and learning

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). 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:

  • A to­tal score and break­down by % of notes cor­rect af­ter each ex­er­cise.

  • Let­ting me play my­self back to see what be­ing on vs slightly off pitch sounds like.

  • Helpful text prompts such as “keep your tongue be­hind your teeth.”

  • A built-in pro­gres­sion from eas­ier to more com­pli­cated ex­er­cises.

  • Ex­er­cises for a spe­cific skill, like switch­ing from head to chest voice.

  • The abil­ity to add challenge by turn­ing down/​off the pi­ano play­ing my notes as I sing them.

  • Not a built-in part of the app, but I can glance away from the screen and sing with­out the vi­sual feed­back, and then im­me­di­ately check how I did.


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.