Brains Cling to Old Habits When Learning New Tricks(quantamagazine.org)
quantamagazine.org
Brains Cling to Old Habits When Learning New Tricks
https://www.quantamagazine.org/brain-computer-interfaces-show-that-neural-networks-learn-by-recycling-20180327/
28 comments
Who would have thought that a little bit of close-mindedness might be good for something?
Imagine trying to convince an AI of the future to give up deeply held, almost religious beliefs because the designers wanted to mitigate this problem. Sounds like the premise of a sci-fi plot. It might even exist already.
Imagine trying to convince an AI of the future to give up deeply held, almost religious beliefs because the designers wanted to mitigate this problem. Sounds like the premise of a sci-fi plot. It might even exist already.
It reminds me of the adage, "Keep an open mind, but not so open that your brains fall out."
Though I like the more eloquent 1886 speech by Sir Edward Clarke in the U.K. House of Commons:
But what did that speech amount to? It came to this ingenuous confession of an open mind. The mind was indeed so open that it had nothing in it at all.
Though I like the more eloquent 1886 speech by Sir Edward Clarke in the U.K. House of Commons:
But what did that speech amount to? It came to this ingenuous confession of an open mind. The mind was indeed so open that it had nothing in it at all.
> Who would have thought that a little bit of close-mindedness might be good for something?
Who thinks total openness is a good thing? Life does not deal in absolutes.
Who thinks total openness is a good thing? Life does not deal in absolutes.
Degree of (and physio mechanisms regulating) plasticity are brain-region specific.
Motor cortex is using a plasticity framework that supports 'procedural' memory formation. This type of learning is relatively slow and relies on small incremental changes (for ML people, maybe this is similar to having your nn regularization parameter cranked way up).
Other areas of cortex can support 'declarative' memory formation. For example you can learn things like the definition of a new word (lang), features of a novel object (vis), where you parked your car outside your new apartment (spatial/episodic), within just a few seconds.
For the control of prosthetic limbs, there is perhaps no better brain-computer interface than motor cortex. Despite its relatively laborious procedural-type learning framework, limb movement is precisely what PMC has been optimized for. Since we've been digging around PMC for a while and have managed to make these electronic prosthetics work well, it's only natural to see what other tasks we can perform with those PMC-embedded electrodes. We should however keep in mind, though, the human brain has a wondrous variety of subregions optimized via evolution to perform all manner of computations, to support some subtype of behavior or mental capacity, with varying degrees of task-specificity. All I'm really saying is that... we will likely find out that a PMC-computer interface is garbage at doing some of the things we imagine should be possible, like typing words into a text box using thoughts. While it's likely possible (motor cortex could think about moving virtual fingers typing on a keyboard, and the computer interface could pick up on these signal patterns), there is another set of brain regions better suited for this task that could be tapped-into. But yes, "Some Brain Regions Cling to Old Habits When Learning New Tricks".
Motor cortex is using a plasticity framework that supports 'procedural' memory formation. This type of learning is relatively slow and relies on small incremental changes (for ML people, maybe this is similar to having your nn regularization parameter cranked way up).
Other areas of cortex can support 'declarative' memory formation. For example you can learn things like the definition of a new word (lang), features of a novel object (vis), where you parked your car outside your new apartment (spatial/episodic), within just a few seconds.
For the control of prosthetic limbs, there is perhaps no better brain-computer interface than motor cortex. Despite its relatively laborious procedural-type learning framework, limb movement is precisely what PMC has been optimized for. Since we've been digging around PMC for a while and have managed to make these electronic prosthetics work well, it's only natural to see what other tasks we can perform with those PMC-embedded electrodes. We should however keep in mind, though, the human brain has a wondrous variety of subregions optimized via evolution to perform all manner of computations, to support some subtype of behavior or mental capacity, with varying degrees of task-specificity. All I'm really saying is that... we will likely find out that a PMC-computer interface is garbage at doing some of the things we imagine should be possible, like typing words into a text box using thoughts. While it's likely possible (motor cortex could think about moving virtual fingers typing on a keyboard, and the computer interface could pick up on these signal patterns), there is another set of brain regions better suited for this task that could be tapped-into. But yes, "Some Brain Regions Cling to Old Habits When Learning New Tricks".
So essentially, the holy grail of a purely "abstraction of an abstraction of an abstraction ad infinitum" AI is mostly a pipe dream? Everyone loves to hope that everything that matters can be built from [insert object here]. Physicists want strings. Programmers want s-expressions or combinators or [choose your poison]. And AI researchers want the post-TLU artificial neuron model to learn it all..
I have always thought that machine learning is the less likely path to artificial human-level intelligence. It makes more sense to create a universe. I know your laughing now - but please hear me out. A digital universe can be merely a bunch of cells on a grid with specified rules for state-space updates. Cellular automata are the most well known example ie. Game of Life.
If we look at the requirements for evolution and the requirements for information stability and movement, we can then proceed to create a 2d universe, tinker enough until we get life. Information stability and movement is mostly just the entities of matter and energy, and their conversion and total conservation. The life may not be all that remarkable at first but we run the universe more cycles, throw more compute at it, and we can push the clock forward quite a large amount. The trick of course, is not to try to make OUR universe. Too much information. Too much compute power required. Too much space...likely without life -- so wasted resources for our goal. And lastly, too many layers - even going from bosons and fermions we have: Protons/Neutrons/Electrons/Atoms/Molecules/Amino Acids/Proteins/RNA/DNA/Single Cells/Multicellular Life Forms/Mammals/Primates/Humans. The last 3 could be considered as evolution - so not all have to be present at the same time. However, all the other abstractions are a real dimensionality crusher. Are they totally necessary for an intelligent being? I doubt it.
Once we find the local cluster of life we are interested in, in our digital universe, we could even do procedural generation instead of full state-space updates. So we can speed up the evolution of our target even quicker.
There's nothing I haven't said here but here's my blog post about this idea if you're interested in reading more: https://scrollto.com/life-a-universe-simulation/
I have always thought that machine learning is the less likely path to artificial human-level intelligence. It makes more sense to create a universe. I know your laughing now - but please hear me out. A digital universe can be merely a bunch of cells on a grid with specified rules for state-space updates. Cellular automata are the most well known example ie. Game of Life.
If we look at the requirements for evolution and the requirements for information stability and movement, we can then proceed to create a 2d universe, tinker enough until we get life. Information stability and movement is mostly just the entities of matter and energy, and their conversion and total conservation. The life may not be all that remarkable at first but we run the universe more cycles, throw more compute at it, and we can push the clock forward quite a large amount. The trick of course, is not to try to make OUR universe. Too much information. Too much compute power required. Too much space...likely without life -- so wasted resources for our goal. And lastly, too many layers - even going from bosons and fermions we have: Protons/Neutrons/Electrons/Atoms/Molecules/Amino Acids/Proteins/RNA/DNA/Single Cells/Multicellular Life Forms/Mammals/Primates/Humans. The last 3 could be considered as evolution - so not all have to be present at the same time. However, all the other abstractions are a real dimensionality crusher. Are they totally necessary for an intelligent being? I doubt it.
Once we find the local cluster of life we are interested in, in our digital universe, we could even do procedural generation instead of full state-space updates. So we can speed up the evolution of our target even quicker.
There's nothing I haven't said here but here's my blog post about this idea if you're interested in reading more: https://scrollto.com/life-a-universe-simulation/
Maybe AGI is just a problem of computation power and complexity.
Can infinite monkeys write a brain?
https://en.wikipedia.org/wiki/Infinite_monkey_theorem
Can infinite monkeys write a brain?
https://en.wikipedia.org/wiki/Infinite_monkey_theorem
It almost certainly is. But we can at least say that our universe is a solution. Still isn't guaranteed to help remove the infinite though since we don't know if our universe is just one flavor of an infinite class of parameters - some which may be unobservable in full precision.
You might find this paper interesting, if you haven't heard of it already: https://www.simulation-argument.com/simulation.html
It is possible we are prone to sudden termination or tinkering because we are instantiated by equipment outside our own possible perceptions -- but does that make it a simulation? Our very definition of information manipulation is computation. And our science of physics is founded on the principles of unitarity. Saying there is a God is the same as saying our universe is a simulation. I think the only important factor is whether we are prone to manipulation by external unobservable actions -- It is funny then to just see strongly religious folk as believers in that nice surly sysadmin Big G. And then throw blackhat redboxing Satan in there too, he hacks the box occasionally and overheats the case. Poor misunderstood hacker..
"Simulation" is such a catch 22. There is no way to know if anything runs itself in compartmentalized fashion so it needs no higher equipment. Its the snake eating its own tail. If its possible then maybe we are isolated...Black holes as isolated universes somewhat accomplishes this only due to continuous values..time dilation...infinite time inside..finite time outside. Its a pretty thought.
But I'd still raise the flag to say that technically, we could eventually create black holes. And that slippery slope leads to the question: If THEY chose the seed but cant intervene - are we technically a simulation? Or are we free?
All the usual questions about God and divine intervention..and the asymmetrical cosmic background radiation possibly due to the big bang (seed?) come to bear.
Thanks for making me think about these things. Really phanatasmagoric ideas here.
"Simulation" is such a catch 22. There is no way to know if anything runs itself in compartmentalized fashion so it needs no higher equipment. Its the snake eating its own tail. If its possible then maybe we are isolated...Black holes as isolated universes somewhat accomplishes this only due to continuous values..time dilation...infinite time inside..finite time outside. Its a pretty thought.
But I'd still raise the flag to say that technically, we could eventually create black holes. And that slippery slope leads to the question: If THEY chose the seed but cant intervene - are we technically a simulation? Or are we free?
All the usual questions about God and divine intervention..and the asymmetrical cosmic background radiation possibly due to the big bang (seed?) come to bear.
Thanks for making me think about these things. Really phanatasmagoric ideas here.
"Other areas of cortex can support 'declarative' memory formation. For example you can learn things like the definition of a new word (lang), features of a novel object (vis), where you parked your car outside your new apartment (spatial/episodic), within just a few seconds."
Can you provide other examples of this phenomenon? It seems like the ones you put forth here are all examples of relying on previous experience, which seems more on the "habitual" end of the spectrum, which is the articles claim in the general case.
I can learn new word's definitions because I've spent a significant amount of time learning definitions to all sorts of words throughout my life. Thus, the process is very familiar to me, and is a habit.
I have learned over the course of my life that novelties are by definition outliers. Just like all animals with a survival instinct, I've learned that strange things can be dangerous, so it is an effective heuristic to expect danger from strange things. Another heuristic I favor is that the best way to defang danger is to understand it enough to avoid its mechanisms.
I have learned over the course of my driving experience that keeping track of where I parked my car is crucial to locating the car again in the future. Thus, through experience, I know that it is an important piece of information to remember.
These three examples seem to fit perfectly well within the hypothesis of the article. They aren't new tricks. They are by definition the application of old tricks.
Can you provide other examples of this phenomenon? It seems like the ones you put forth here are all examples of relying on previous experience, which seems more on the "habitual" end of the spectrum, which is the articles claim in the general case.
I can learn new word's definitions because I've spent a significant amount of time learning definitions to all sorts of words throughout my life. Thus, the process is very familiar to me, and is a habit.
I have learned over the course of my life that novelties are by definition outliers. Just like all animals with a survival instinct, I've learned that strange things can be dangerous, so it is an effective heuristic to expect danger from strange things. Another heuristic I favor is that the best way to defang danger is to understand it enough to avoid its mechanisms.
I have learned over the course of my driving experience that keeping track of where I parked my car is crucial to locating the car again in the future. Thus, through experience, I know that it is an important piece of information to remember.
These three examples seem to fit perfectly well within the hypothesis of the article. They aren't new tricks. They are by definition the application of old tricks.
This might also explain why learning a 2nd language is usually harder than the 1st, and why getting rid of, or changing, an accent is so hard. All that time invested in those old neural circuits, better make use of those first.
I'm still not convinced learning a 2 language is harder, even as an adult. I now know 3 and if you put some effort and time (or even better get immersed) you will learn it. Children have YEARS of amazing immersion, adult patience and care. Who wouldn't learn a language in such fertiile conditions.
As for the accent which i have, i don't know yet. I will try to correct it and see how it goes.
As for the accent which i have, i don't know yet. I will try to correct it and see how it goes.
The step from one language to two is much, much larger than the step from two to three or beyond. I did this seriously for the first time when I was 21/22, and it was initially difficult to get used to the concept of thinking not in terms of the temporal order in which words are introduced (i.e. the syntax of the language in which you are speaking) or at least shifting your process to generate sentences on the fly in a new order. In my experience (native English, learned German) I had a relatively smooth transition but even so it was difficult not to construct sentences as I spoke in the typical English construction. Once you get comfortable with shutting off that reflex to formulate thoughts according to the syntax and vocabulary of your own language, learning language becomes a lot smoother. This reflex is of course more strongly burned-in the older one gets due to the decrease in neuroplasticity as we age.
Kids have incredible motivation and time far beyond any adult.
If you spent 1 hour a day learning a new language most people would consider you a dedicated language learner, but a child spends almost every waking our in contact with language.
Children also have to keep up with their peers and siblings (and adults after a little while).
If I had unlimited money I'd love to do a research project with adults in full immersion environments similar to what children experience for years at a time with no chance to 'run away' to their native language.
If you spent 1 hour a day learning a new language most people would consider you a dedicated language learner, but a child spends almost every waking our in contact with language.
Children also have to keep up with their peers and siblings (and adults after a little while).
If I had unlimited money I'd love to do a research project with adults in full immersion environments similar to what children experience for years at a time with no chance to 'run away' to their native language.
I thought that the more languages one learns the easier it becomes?
Isn't that more to do with age. You can pick up any new language or accent before age of 15.
I know very little about learning, but it sounds sorta like multi armed bandit. Put a lot of emphasis on past success, rather than exploring. I would guess a really good tennis player would have a hard time getting good at squash, as opposed to a rank beginner. There is nothing to unlearn for the beginner.
Indeed this is a well-known phenomenon: https://en.wikipedia.org/wiki/Negative_transfer_(memory)
Skills should be leaned on when needed but we should always approach new things, and revisit things you know, with a beginners mind zen.
I believe a large part of being an engineer or problem solver is to take complex things and make them simple as can be but no simpler as Einstein stated, that way others can also come into that with a beginners mind.
Basically be Feynman and learn how things really work and when you find out, explain them as basic/simple as possible for understanding [1]. If you can’t explain something in simple terms, you don’t understand it.
[1] https://kottke.org/17/06/if-you-cant-explain-something-in-si...
I believe a large part of being an engineer or problem solver is to take complex things and make them simple as can be but no simpler as Einstein stated, that way others can also come into that with a beginners mind.
Basically be Feynman and learn how things really work and when you find out, explain them as basic/simple as possible for understanding [1]. If you can’t explain something in simple terms, you don’t understand it.
[1] https://kottke.org/17/06/if-you-cant-explain-something-in-si...
The most interesting thing I read recently about brain plasticity is that, the brain recycles connections formed in earlier memories when you learn something new. So you will lose the ability to easily recall old memories if tasked to learn new tricks.
So neurons don’t disassemble their current patterns to find the optimal pattern for the latest training goal? That sounds like a good neural algorithm because how do you know that the old should be fully replaced with the new? If I were God, my creatures’ brains would adapt older patterns so they could do double duty, especially in the short term. Because there is still a high Bayesian likelihood that the older behaviors might need to be recalled. This neural model creates creatures that can adapt to an expanding set of training goals.
If you are learning something that requires high level control and conscious decision making and correction, what is typical rate of full feedback loop? Maybe 1-2 times per second?
If you train nonstop for an hour, that's just 3600 - 7200 times. That's relatively small batch size to learn something. Even 10 or 100 hours of training is not that much. It's not surprise if brains reuses old stuff as much as possible.
If you train nonstop for an hour, that's just 3600 - 7200 times. That's relatively small batch size to learn something. Even 10 or 100 hours of training is not that much. It's not surprise if brains reuses old stuff as much as possible.
"""Yu [...] also noted that their experiments only tracked the brain’s activity for one or two hours. The researchers can’t yet rule out the possibility that reassociation is a fast interim way for the brain to learn new tasks; over a longer time period, realignment or rescaling might still show up."""
I know that for musical instrument practice the common advice I've read is that it is better to practice 30 minutes every day rather than a 3½-hour session at the weekend in part because what you learn actually sinks in when you are resting. Maybe the realignment or rescaling happened when the monkeys slept?
I know that for musical instrument practice the common advice I've read is that it is better to practice 30 minutes every day rather than a 3½-hour session at the weekend in part because what you learn actually sinks in when you are resting. Maybe the realignment or rescaling happened when the monkeys slept?
Probably the cause of the "Einstellung Effect" [0]
[0] https://en.wikipedia.org/wiki/Einstellung_effect
[0] https://en.wikipedia.org/wiki/Einstellung_effect
A defense for ageism?
[0] https://en.wikipedia.org/wiki/Catastrophic_interference