Suppose we take the plasticity principle as implying that the brain is an automatic learning machine acquiring information pretty much on it's own over time (i.e., what if we can't control much about learning/memory rate by meta-cognitive control processes). Maybe that's too strong, but the following analysis ought to hold even just applied to the set of things learning implicitly.
At any given instant, t, there is a set of information active in the brain. This includes the state of the perceptual systems, processing of relationships among items in association cortex (including semantic memory), the contents of working memory and other control processes. Some fraction of this information changes the state of the brain in a way that is detectable later, that's the acquisition of memory. We could make it a formula:
M(t) = f(cognitive state)
I don't know if it's useful, but it's cute to observe that M(t) is technically the derivative of a function that describes the total amount of information currently held in the brain. Perhaps it is useful as it reminds us that we can also forget or lose things we know. So the instantaneous change in memory state in the brain is more properly:
M(t) = f(cognitive state) - g(Information State)
where g is a forgetting function. One way we can describe the process of getting smarter is to figure out how to maximize M (find the second derivative? heh). If all information is equal, then it's a matter of doing things that maximize f() and minimize g(). You could make a pretty good argument that some things are more important to learn than others so simply maximizing f() isn't as important as what the content of the learning is. This is probably correct, but one shouldn't be too sure -- how do we really know what is going to be important in the future? That line of inquiry will lead us tangentially off in the direction of Everything Bad is Good For You.
Let's consider some simpler things about f(). One issue is repetition vs novelty. Exact repetition of a prior cognitive state can be stored pretty compactly (e.g., you could count how often it occurs and store the whole episode as just increasing a counter). A completely novel experience adds all sorts of rich new data and connections among prior ideas and should add a lot more information.
But it's not that simple. Repetition sometimes allows you to see deeper into a domain, e.g., playing over a chess game is a great training exercise in that it lets you see new connections and new ideas. And novel episodes are going to be primarily learned via the MTL memory system, which probably has bandwidth limits (repetition-based, statistical, implicit learning less so). The degree to which novel information "makes sense" based on what you already know strongly increases p(storage), which suggests there's an optimal level of novelty (as an aside, when you hit that optimal level, I suspect you get a humor response). Repetition may also decrease g() and let you tend to hold onto prior knowledge.
It'd be nice if there were biological ways to boost f() and/or decrease g() (memory drugs). Good sleep hygiene probably does, although it also limits your perceptual input.
Leaving that there for now, how do we consider the possibility of getting stupider under this model? We don't know of anything that increases g() dramatically (short of neuropathology or brain damage). Our model is one of constant acquisition. By this model, the main way one can become stupider is to accidentally acquire false information.
I was reviewing a paper the other day that had several major conceptual flaws. It occurred to me that if somebody who didn't know the area read this paper, they would actually be stupider after reading the paper than before (no, my review was not positive). This is a general concern, but many examples occur in the political arena (e.g., WMD were found in Iraq, actionable intelligence was obtained from torture, Obama was born in Kenya).
If you want to protect yourself from accidentally becoming stupider, it's worth noting that the real risk comes from areas you don't know very well. If somebody says, "the moon is made of green cheese." Well, you're probably not stupider because you don't believe it. In fact, you have learned that this person either things the moon is made of green cheese or for some reason wants you to think that it is.
But if you want to read outside your area of expertise, you need to check your information stream carefully. Can you corroborate (spot check) a few facts? Does the source provide corrections to inevitable errors? Sometimes people say to me, you read a lot of information coming from one side of the political spectrum, don't you worry about the "echo chamber" effect? No, because the echo is boringly repetitive and I select a subset of sites that provide generally accurate information. Note that this doesn't include NYTimes columnists like Friedman, Brooks or Broder. They get some things right, but I've seen too many documented instances of them doing things that would have made me stupider. I might have accidentally learned that reconciliation is unconstitutional or only used by Democrats or something.
I deduce two obvious things about learning from this:
1. Manage your input stream (experiences) to optimally balance novelty and repetition to maximize your information storage rate.
2. Guard your information stream carefully to avoid falsehoods from getting into your brain and making you stupider.
Thursday, March 25, 2010
Tuesday, March 16, 2010
The Plasticity Principle
The Plasticity Principle describes my conceptual model of what implicit learning is. The core idea is that there are a lot of damn neurons in the brain and even more synapses, and if we start with the assumption that every one of them is plastic (i.e., changeable), what are the consequences of that?
The first is that it is not correct to think of the brain as a bunch of static processing systems like the visual system, the motor system, control systems, the memory system. Instead, we think of every one of those systems as capable of being shaped by experience. The goal of this learning is to continually refine and optimize processing. We don't know exactly how this learning process will go because we don't yet know either the capacity and limits of this inherent plasticity nor exactly what optimal processing looks like.
Chasing those operating characteristics questions leads to psychological and cognitive neuroscience research on learning and memory processes. But I have a growing sense there are other more accessible implications.
1. The world shapes your brain (physically). On the upside, this is the basis for why Everything Bad is Good For You (in some cases). Games and other entertainments that create knowledge or strengthen cognitive skills actually enhance brain function. This happens because those systems strengthen through incidental practice.
2. On the downside, the world can create bad habits in your brain as well. Implicit racism is a good example of this. Regular co-occurence of negative ideas and minorities creates bias that affects your behavior that you don't even realize is there.
3. Cognitive training works. Practice with mental exercise improves cognitive function. Brain training helps with aging and probably in degenerative disorders as well. It will be cool to see how video games can be integrated with this idea.
4. The trick will be figuring out the right things to practice. You can certainly get good at specific skills that don't help much elsewhere. I'd say chess is a good example of this. Chess players are freakishly good at chess and chess memory. It doesn't mean they are very good at anything else, though.
Most of everything I've jotted down in this blog format is strongly influenced by the underlying idea:
5. Thinking about habits and perfection is motivated by marveling at how close to perfection habits can get. The upper bound on optimal performance is pretty amazing and that argues that this type of plasticity is pretty effective.
6. Teaching should incorporate skill development. Not all cognitive skills can be easily taught in games. The classroom is an environment where students will commit to acquiring some other types of skills and teaching should take advantage of that.
7. The Butterfly effect and Nature/Nurture. Most nature/nurture discussions do not take sufficient account of feedback effects through learning. If you assume there's a lot of plasticity in the brain, you should be very sensitive to feedback spirals. That is, a small push/advantage in a domain can cause you (or others who influence your environment) to direct more attention to that domain, which will push you further away from the mean as learning contributes. This will exacerbate the effect of small genetic differences, but only when the feedback loop isn't externally constrained. In the IQ world, things like race and gender have big impacts on getting the feedback loop started and keeping it going.
I think that humans' expertise in face processing might be an excellent example for #7. Mark Johnson's model of a small genetic/pre-wired push towards looking at faces is a good start. Ken Nakayama's examination of individual differences in face recognition ability contributes. His lifespan/ability graph showing the ability peaks in one's early 30s makes me intuitively confident that there's a big learning component. I should dig in and figure out why.
8. Statistical models of language processing. These capture more language than you'd think. And the concept is showing up in technology like Google. It ought to be more prevalent in recommendation algorithms like Netflix/Amazon, I think (but the argument isn't trivial).
Broadly, my thinking here is strongly influenced by John Anderson's Adaptive Character of Thought (his rational analysis of cognition). And I think also by Herb Simon's economic theory of Bounded Rationality (maybe less obviously). The idea of looking for the "operating characteristics" of implicit learning was developed with Larry Squire. I remember he was a fan of the phrase, although I think his model of nondeclarative memory was as a finite set of separate systems rather than a broad "every neuron" principle.
Anyway, is this a useful/interesting collection of implications? I can haz popular science book? Too much real science work to do right now anyway, but I'll keep collecting related thoughts here in the meantime.
The first is that it is not correct to think of the brain as a bunch of static processing systems like the visual system, the motor system, control systems, the memory system. Instead, we think of every one of those systems as capable of being shaped by experience. The goal of this learning is to continually refine and optimize processing. We don't know exactly how this learning process will go because we don't yet know either the capacity and limits of this inherent plasticity nor exactly what optimal processing looks like.
Chasing those operating characteristics questions leads to psychological and cognitive neuroscience research on learning and memory processes. But I have a growing sense there are other more accessible implications.
1. The world shapes your brain (physically). On the upside, this is the basis for why Everything Bad is Good For You (in some cases). Games and other entertainments that create knowledge or strengthen cognitive skills actually enhance brain function. This happens because those systems strengthen through incidental practice.
2. On the downside, the world can create bad habits in your brain as well. Implicit racism is a good example of this. Regular co-occurence of negative ideas and minorities creates bias that affects your behavior that you don't even realize is there.
3. Cognitive training works. Practice with mental exercise improves cognitive function. Brain training helps with aging and probably in degenerative disorders as well. It will be cool to see how video games can be integrated with this idea.
4. The trick will be figuring out the right things to practice. You can certainly get good at specific skills that don't help much elsewhere. I'd say chess is a good example of this. Chess players are freakishly good at chess and chess memory. It doesn't mean they are very good at anything else, though.
Most of everything I've jotted down in this blog format is strongly influenced by the underlying idea:
5. Thinking about habits and perfection is motivated by marveling at how close to perfection habits can get. The upper bound on optimal performance is pretty amazing and that argues that this type of plasticity is pretty effective.
6. Teaching should incorporate skill development. Not all cognitive skills can be easily taught in games. The classroom is an environment where students will commit to acquiring some other types of skills and teaching should take advantage of that.
7. The Butterfly effect and Nature/Nurture. Most nature/nurture discussions do not take sufficient account of feedback effects through learning. If you assume there's a lot of plasticity in the brain, you should be very sensitive to feedback spirals. That is, a small push/advantage in a domain can cause you (or others who influence your environment) to direct more attention to that domain, which will push you further away from the mean as learning contributes. This will exacerbate the effect of small genetic differences, but only when the feedback loop isn't externally constrained. In the IQ world, things like race and gender have big impacts on getting the feedback loop started and keeping it going.
I think that humans' expertise in face processing might be an excellent example for #7. Mark Johnson's model of a small genetic/pre-wired push towards looking at faces is a good start. Ken Nakayama's examination of individual differences in face recognition ability contributes. His lifespan/ability graph showing the ability peaks in one's early 30s makes me intuitively confident that there's a big learning component. I should dig in and figure out why.
8. Statistical models of language processing. These capture more language than you'd think. And the concept is showing up in technology like Google. It ought to be more prevalent in recommendation algorithms like Netflix/Amazon, I think (but the argument isn't trivial).
Broadly, my thinking here is strongly influenced by John Anderson's Adaptive Character of Thought (his rational analysis of cognition). And I think also by Herb Simon's economic theory of Bounded Rationality (maybe less obviously). The idea of looking for the "operating characteristics" of implicit learning was developed with Larry Squire. I remember he was a fan of the phrase, although I think his model of nondeclarative memory was as a finite set of separate systems rather than a broad "every neuron" principle.
Anyway, is this a useful/interesting collection of implications? I can haz popular science book? Too much real science work to do right now anyway, but I'll keep collecting related thoughts here in the meantime.
Monday, March 15, 2010
Perfection
[Some thoughts that seem like they ought to be more related...]
Habits are supposed to be things we do smoothly, seemlessly and generally without error. But we still aren't perfect at them. We stumble on the stairs sometimes or drop our keys when opening the house or car door for no apparent reason. Should we be impressed at how good our habit system is or annoyed about these failures and non-perfections?
Tap Tap Revenge 3 is a sequence learning rhythm game that is structurally like Guitar Hero on the iPhone (or video iPod) except you tap on the screen for your responses. This makes is a lot like our lab task, SISL (for Serial Interception Sequence Learning). So, of course, I'm doing an introspective research experiment on long-term learning in TTR3.
I'm pretty accurate in Medium level difficulty mode. The %hit number at the end of any given song is usually 99% or 100% (often by rounding). I started getting an occasional "Full Clear" (FC) recently. This is an achievement you get if you make zero errors of any type for an entire song of 500-600 notes -- perfection!
What're the odds? Well, if you are 99.0% accurate, the odds of making 600 errorless responses in a row is... 0.2%, about 1:400. Pretty grim. At 99.5% accurate, you get up to around 5% (1:20). Practically speaking, it's more complicated since your accuracy is higher for songs you know (and I know I tend to play songs more when I like the song and can't promise I like easier songs better because, well, dopamine). Also, it's not totally clear how many error chances there are. Most of the "errors" I make now aren't missing or mis-timing a planned response, but accidentally double responding (or making n+1 responses to an n-response train, particularly for n's > 5) or accidentally dragging a finger on the screen when I don't mean to.
On Hard mode, my %hit score is typically down around 97%-98% rate, which seems pretty good but you can tell intuitively that you have no chance at all of a FC even without doing the odds on a calculator.
What's the point? Well, perfection is hard. And your habit learning system has to be pretty sharp to get you anywhere near perfection in the first place. I didn't download the TTR3 app that long ago, nor do I have that much time to practice, but my cortico-striatal circuits seem to be getting a pretty firm grip on the sequences.
I'm not even sure it's a good idea to dwell on how hard it is to achieve perfection, but the unnerving example I throw out sometimes is "how hard do you think it is to land an airplane?" Cause it seems like it's hard, but 10,000 planes land every day with virtually no incidents. There's a lot of 9's in that reliability rate. How do they do that? Doesn't it ever get boring or anybody lose focus? Either not or there are enough oversight systems to compensate, I guess. Or maybe flying an airplane isn't as hard as playing TTR3?
Habits are supposed to be things we do smoothly, seemlessly and generally without error. But we still aren't perfect at them. We stumble on the stairs sometimes or drop our keys when opening the house or car door for no apparent reason. Should we be impressed at how good our habit system is or annoyed about these failures and non-perfections?
Tap Tap Revenge 3 is a sequence learning rhythm game that is structurally like Guitar Hero on the iPhone (or video iPod) except you tap on the screen for your responses. This makes is a lot like our lab task, SISL (for Serial Interception Sequence Learning). So, of course, I'm doing an introspective research experiment on long-term learning in TTR3.
I'm pretty accurate in Medium level difficulty mode. The %hit number at the end of any given song is usually 99% or 100% (often by rounding). I started getting an occasional "Full Clear" (FC) recently. This is an achievement you get if you make zero errors of any type for an entire song of 500-600 notes -- perfection!
What're the odds? Well, if you are 99.0% accurate, the odds of making 600 errorless responses in a row is... 0.2%, about 1:400. Pretty grim. At 99.5% accurate, you get up to around 5% (1:20). Practically speaking, it's more complicated since your accuracy is higher for songs you know (and I know I tend to play songs more when I like the song and can't promise I like easier songs better because, well, dopamine). Also, it's not totally clear how many error chances there are. Most of the "errors" I make now aren't missing or mis-timing a planned response, but accidentally double responding (or making n+1 responses to an n-response train, particularly for n's > 5) or accidentally dragging a finger on the screen when I don't mean to.
On Hard mode, my %hit score is typically down around 97%-98% rate, which seems pretty good but you can tell intuitively that you have no chance at all of a FC even without doing the odds on a calculator.
What's the point? Well, perfection is hard. And your habit learning system has to be pretty sharp to get you anywhere near perfection in the first place. I didn't download the TTR3 app that long ago, nor do I have that much time to practice, but my cortico-striatal circuits seem to be getting a pretty firm grip on the sequences.
I'm not even sure it's a good idea to dwell on how hard it is to achieve perfection, but the unnerving example I throw out sometimes is "how hard do you think it is to land an airplane?" Cause it seems like it's hard, but 10,000 planes land every day with virtually no incidents. There's a lot of 9's in that reliability rate. How do they do that? Doesn't it ever get boring or anybody lose focus? Either not or there are enough oversight systems to compensate, I guess. Or maybe flying an airplane isn't as hard as playing TTR3?
Subscribe to:
Posts (Atom)