Hacker News new | past | comments | ask | show | jobs | submit login
Overlearning hyperstabilizes a skill by making processing inhibitory-dominant (nature.com)
232 points by alfozan on Feb 6, 2017 | hide | past | web | favorite | 52 comments



Practice doesn't make perfect. Rather, practice makes Permanent - locks it in. And perfect practice makes perfect.

Practicing the wrong way is like having a bad habit. The habit (practice) is bad but it locks in nonetheless.

Perfect Practice

To reach perfection, care has to be taking when practicing to encode things the right way. Good coaches accomplish this by:

1. Slowing down the action e.g. playing a piano or a violin at a painfully slow rate to make errors more noticeable

2. Chunking - breaking down the skill into smaller components and practicing these chunks separately. E.g. tearing music notes into pieces and practicing the notes separately

Books: Little Book of Talent, The Talent Code


1. Slowing down the action e.g. playing a piano or a violin at a painfully slow rate to make errors more noticeable

You have to be careful with this as well. There are some musicians who do this, but can only play at the practice speed. (So be sure to speed things up, and vary things. You'll have true mastery if you can morph what you've practiced.) It's also good to listen to recordings of yourself and of model musicians as a means of feedback.

2. Chunking - breaking down the skill into smaller components

Care has to be taken with this as well. I've encountered musicians who have chunked ornaments in traditional music, but have practiced so much, that they consistently replay the ornament at their practice tempo, regardless of the tempo being played at the moment.

Music is a game of continual contextual awareness. There is a powerful analogy with programming here.


One of the tricks Ive worked with, is practicing something slow and chunked three times in a row perfectly. If you fail you start over, I.e. you may play it perfectly twice, but third time you fail, you start over.

Then once you do a small chunked section slowly perfectly (say 8 measures), you then try at regular speed until you do it three times perfectly.

Then you try playing two chunks (say 16 measures) the same way, constantly building.

For reference, I had a full ride to multiple universities for my music. So at least for me it worked well this way.


This is exactly the way I learned to practice riffs in jazz solos. You start at half speed, then slowly increase the speed, but only increase the speed if you can play the passage perfectly 3 times in a row at the current speed.

It works amazingly well. It also requires persistence and determination, but so does any type of skill that differentiates you from the average person.


I've never been in a university music program, but I've gotten lessons from various people who have, including the director of my college early music group, so I've been introduced to and have implemented quite a bit of that kind of practice.

For traditional music, you need to add variation by rhythmic emphasis as well. Just because something is called a "hornpipe" or a "fling," this doesn't entirely specify the rhythm, beyond the time signature that has been most often used to notate that particular kind of dance. (And it doesn't always even get you that far, actually.) The first note is probably emphasized by giving it a bit more time to varying degrees, and there might be other subtle changes to give it more or less or a different feeling of "swing." To be a very good traditional musician, one needs to listen to and adapt to these changes/variations.


'Chunking down' a piece is pretty much the only way I've managed to learn some harder pieces, at least by my regular-guy standards e.g. Rach Prelude in C sharp minor. That last one was particularly hard as I'm a terrible sight reader, and there's about two pages where you're playing 10 notes every half-beat.

But I think GP has a point about the dangers of this approach (even though it's unavoidable I think). It sometimes results in noticeable 'seams' in an otherwise fluent performance. You can occasionally spot these 'seams' even in performances by world-class concert pianists (though this is rare). I find it weirdly reassuring: no matter how talented you are, those first few days of practising a new piece will always be a tedious grind.


there's about two pages where you're playing 10 notes every half-beat.

There's a bit of stuff in classical music that strikes me as being there so that musicians can show off their speed. It's one of the things that even the lowest common denominator audience can relate to, so of course this happens!

It sometimes results in noticeable 'seams' in an otherwise fluent performance.

The way to avoid seams is to truly make the chunk yours, by being able to morph it in rhythm and time and introducing variation. By the time you can freely do that -- by the time it has become a plaything -- it is truly yours, and you can play it without seams. The problem for me with this and classical music, is that often I really don't care so much for the difficult segments. I'd rather play something beautiful, than something that shows me off.


There's a bit of stuff in classical music that strikes me as being there so that musicians can show off their speed. It's one of the things that even the lowest common denominator audience can relate to, so of course this happens!

Heh, agreed. Though I think the Rach Prelude was more designed to break piano strings :) (they're chords, some of which are marked 'sffff'... also looking at the sheet music it's actually 8 notes)

There seem to be a few etudes like you've described (Liszt comes to mind). I cannot for the life of me figure out what these composers were attempting to teach their students...


Absolutely! I learned to play a lot of fast runs of sixteenth notes by varying the rhythm. Effective and much more fun than starting slowly ticking up the metronome (although I do a lot of that too).


You may want to check out Cortot, who did a series on how to practice Chopin pieces, notably the notorious Etudes. One of the techniques he describes is varying either rhythm or accentuation. I tried especially the latter which works pretty well. But I am a bit unsure whether it is a genuine effect or a consequence that a practice a chunk more often if you do this.


You can avoid seams by learning overlapping (and varied-length) chunks. If you are working on 2-bar sized chunks, then make sure you also work on 4-bar and 8-bar sized chunks, and offset the chunks by half the chunk size.

E.g. if your chunks start at bars 1, 3, 5, 7, ..., then repeat with chunks starting at 2, 4, 6 , 8..., then scale up to larger-size chunks so that you always have multiple chunks than cover up a 'seam'.


"So be sure to speed things up, and vary things"

Yup. It's called taking chances.


I'd like to piggyback on this to also call out, for people interested in this sort of thing, that this hyper-stabilization or mastery training (which is like developing muscle memory, maybe literally in the case of musicians learning a new piece or technique), is very different from general learning.

See "Make It Stick" and the studies referenced in it regarding "desirable difficulties" that help create strong connections; coaches, teachers and learners mostly express preference for controlled, topical, less-confusing practice, but their preference is usually wrong as measured by performance (in sports and classroom contexts, anyway), whereas mixed-up practice that feels harder, and variations on that like trying to solve things before being taught the correct way, show better results.


Any other good literature on this, or learning in general?


Take a look at the wikipedia articles for Desirable Difficulties and Spacing Effect. Spacing and Interleaving was the hot topic in research the last time I looked (about 5 years ago). Note that "spacing" is different than "spaced repetition". There are quite a few papers linked to the Wikipedia articles, though, very unfortunately the "desirable difficulties" page only lists the original papers. If you google around, you should be able to find quite a few more.

Keep in mind that most of the research in this area is psychological. It is often very difficult to build good studies. On top of that, many papers have pretty poor statistics in them, so take whatever you read with a grain of salt. It's very, very easy to suffer from confirmation bias because there are a lot of studies and many of them are not great. You can pick your conclusion and build an impressive looking list of references. Just be a bit careful of that when you are going through the review literature.


I have a few favorites, but by and large I got those favorites from from Make It Stick, which I've reread a few times now. It's the best view-from-10k feet (written by 3 learning/memory researchers) I could have hoped for in a field with so much significant work being done, and the footnotes are a nice map to that world.

I got interested in learning science from reading about the SuperMemo guy about 10 years ago, and there were some HN discussions where people were saying "this is cool, are there any other really solid learning effects, like spaced reps, that we know about?," and I think it was from HN that I learned about some of Bjork's stuff (one of the fathers of desirable difficulty, https://bjorklab.psych.ucla.edu/robert-a-bjork-publications/).

But I got the chance to learn more when researching for some training software I wrote in 2014, which was when Make It Stick was published; it was really helpful.


> Any other good literature on this, or learning in general?

As another commentter mentioned, "The Talent Code" by Daniel Coyle is excellent. Applying the principles it discussed made my musical practice an order of magnitude more effective than it had been previously.


Based on many years of martial arts practice and two years of guitar practice I'd also add:

- You need to challenge yourself to move the needle. If it's so slow that it's too easy you're not accomplishing that.

- Sometimes too slow is more difficult.

- There are some aspects of training where it doesn't matter if you make errors. E.g. strength or flexibility can be developed while making mistakes and without those you can't actually execute the proper technique anyways. I think the trick is to try to avoid locking in the error by switching to other practices.

- Some physical skills are speed sensitive, i.e. the slow doesn't transfer to the fast.

- It can help speeding things up to the point where you make mistakes just to get speed up, and then coming back to normal speed.

That F barre chord is still driving me crazy though with any technique I try to throw at it ;)


1. Slowing down the action e.g. playing a piano or a violin at a painfully slow rate to make errors more noticeable

Interesting to contrast this advice with the HN story from the other day: https://news.ycombinator.com/item?id=13570892

In that document, several authorities point out that learning Morse code by listening to characters being sent at a painfully slow rate is actually counterproductive -- as in, the worst possible way to go about it. You never learn what the characters really sound like if you do that. Is there an analogous effect at work when learning a new musical instrument?


I think with Morse code there's something else going on. Listening for the "dah-dits" like beginners do at 5-10 wpm is was impossible for me. The only way I learned was to listen to fast, 20 wpm repetition of letters while I wrote them down. Basically I forced the Morse sounds into the hearing part of my brain as alternative sounds for letters. As an interesting side effect, I can't read Morse code (written like -.-. .--) easily. I'm also terrible at sending it.


"perfect practice makes perfect" - wow this comes as a revelation to me. Never realized how important it is to do it right - something very obvious but struck me by surprise when you put it this way.


It is more complex than that though. Imperfect practice is just fine so long as you recognize in practice what you are doing wrong and try to correct it. I often practice a musical piece wrong, I try again and again, each time trying to correct something, but somehow one correction breaks something else. The next day I come back and I'm much better because even though none of my practices was perfect. Repeat a few times [hundreds in some cases] and I can play perfectly.


The piano is such a good illustration. For decades I didn't give a shit about proper posture and when things started to get really fast as they do in rachmaninoff's Moment Musicaux No.4, I finally realized my bad habit was holding me back.

Now I practice it at a super slow level and slowly build it....

I think it's a great philosophy to have in mind...practice doesn't make perfect but a perfect practice makes perfect.


Interesting. I found on mountain bike riding, after a few crashes one of which broke my wrist I realised my skill level wasn't what it needed to be. Slowed it right down, did some skills courses, focusing on basics, keeping different areas of focus in mind at different times - and before I knew it, was going faster and safer than I was before I started having my crashes.

I've had some breaks away from the riding but even then I find that the techniques come back quickly. So yeah seems like learning how to do things properly certainly has a lot of merit.


Another important aspect of chunking that imoroves the efficiency of learning is to not practice for extended periods without taking a break. This is one of the foundations of the Pomodoro technique


Wax on Wax off.


If you can't get past nature's paywall, and want more detail... https://news.brown.edu/articles/2017/01/overlearn

tldr:

Task: detect which one of the two successively presented images had a patterned orientation and which depicted just unstructured noise.

A first group practiced the task for eight blocks, waited 30 minutes, and then trained for eight blocks on a new similar task. The next day they were tested on both tasks to assess what they learned. The other group did the same thing, except that they overlearned the first task for 16 blocks of training.

On the next day’s tests, the first group performed quite poorly on the first task compared to the pre-test. Meanwhile the overlearning group showed strong performance on the first task, but no significant improvement on the second. Regular learning subjects were vulnerable to interference by the second task (as expected) but overlearners were not.


Curious how overlearning influences political beliefs, and if a subject is less likely to alter their belief system based on living in an "echo chamber" and constant re-enforcement of those beliefs.


There's growing evidence that political beliefs are linked to self-image more than factual evidence, which is why arguing over the facts of an issue often winds up causing both sides to double-down on what they already believe to be true and generally accomplishes nothing.


Source please?


I can't speak for his source.

But I have encountered this: I dated a bunch of girls from OKC a couple of years ago. Many of them claimed they were 'Buddhist' on their profiles.

I'm very interested in religion, and thought it was odd so many claimed to be this, when clearly it was not of their 'ethnic origin'.

When I talked to them about it, none of them knew anything about Buddhism beyond 'The Dalai Lama' and a few platitudes. They had never read a full book. Knew next to nothing of Buddhist theology. Never been to a Temple. Had no practices, holidays, knew nothing of the various forms of Buddhism (some of them are pretty hard-core and not what we think of in the West) etc.. So in reality, they 'thought positively of the pop-culture Buddhist image' and what that image meant for themselves, and how it projected onto others. And that's about it - but they were not remotely 'Buddhist' in any tangible way.

My point is people often 'identify with some idea' - but those ideas, and even their understanding of the more formalized idea, can be essentially fictional - or rather, they are used to 'paint a picture of their self identity', and possibly nothing more.


Fair question, but the 'learning' here is entirely subconscious - motor skills, gaming skills.

The 'learning' of ideas is quite a different thing. A single experience or exposure to something can fundamentally alter a set of beliefs for ideals however 'learned' they are.

You can change your mind, but you can hardly change 'how you walk'.


it almost sounds like a long winded way of describing "brainwashing" and "propaganda".


Or that those concepts are a direct result of human learning behaviors, rather than an external directed tactic. Anything can be taught, as long as it is reinforced by existing learned behavior.


So it sounds like they're saying that overlearning is what takes us from from "oh, huh, I was wrong" on encountering new evidence, to instead saying "no way, the evidence has to be wrong."


I think that's way too high level. It seems to me that this paper is discussing learning at the level of memorization or motor skills. I would think that behaviors like egoism can't be reliably connected to something at this level.


You beat me by just a bit.

The connection to higher level behaviors is a tempting one, and at least worthy of research, but due to their greater complexity and the way that inhibition and stimulation have less obvious connections to behaviors at higher levels of neural complexity (as the complexity of the neural system goes up you get fun combinations like "inhibition of inhibition" and "stimulation of inhibition" and all kinds of arbitrarily complicated things), I wouldn't leap to conclusions about that.


I wasn't quite trying to communicate egoism; more what the mental process (you could say "subconscious flinch reaction", but that doesn't capture the "physiological-ness" of what's happening) to encountering the kind of contradictory information that the brain "learns" from.

That is, new information that contradicts your knowledge usually triggers focused attention (i.e. NMDA activation) in order to facilitate learning. However, overlearning can make the same information be discounted as irrelevant (i.e. a GABAergic signal of "it's okay, I don't need to figure that out.")

Explaining at a high level exactly what that "it's okay, don't pay attention to that" signal is rationalized as in conscious experience, is a whole different thing.


Whether you call it egoism or something else, I still think what you're describing is too high level for this research to be applied to it.


I think it's more like: I am so right I don't need to know anything else.


Holy shit. It this what over-fit looks like for humans (as compared to neural nets)?


Gerd Gigerenzer is the author you want for looking at overfit in human beings. Sworn life enemy of the Kahneman and Tversky collaboration, and vice versa(Kahneman actually put in a few disses in his Nobel prize ceremony aimed pretty much at him...).


No, that results in false classification.

This is saying that overtraining prevents revision of existing skills, as if the learning process is interrupted.


More like learning rate decay. Learning rate is now zero.


Not really. With over-fitting, performance goes down, not just plateaues.


Is there an free copy of the article not just the abstract?


Try using scihub


There doesn't seem to be any free option (cheapest is 25$)


Sci-hub link for the paywalled paper : http://sci-hub.cc/10.1038/nn.4490


How then do you forget, after this overlearning already occurred?

EDIT: Someone says "If you can't get past nature's paywall,"


Stop practicing wrong, take some time to forget, and start over later.


I was actually thinking about veterins returning to civillian life at the time, but maybe you're right.




Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: