Mary’s Room is invalid because it sneaks in a false assumption: that if something is knowledge, then it can be learned by study alone. As a counterexample to the story of Mary’s Room, consider the story of Marty’s Bicycle, which I have just made up.
Marty is a brilliant physicist who has suddenly become very interested in bicycles. Unfortunately, his leg is in a cast, so he can’t ride one. While he waits for it to heal, he studies their physics, watches people ride bikes, reads guides on bike riding, and interviews top racers—in fact, he studies everything there is to study about bicycles. When the cast is finally taken off, he buys a bicycle, gets on, and immediately falls over.
The ability to recognize red objects is like the skill of riding a bicycle—it can only be acquired by doing it, not by study, because study can only train the linguistic centers of the brain, not the visual processing centers (Mary’s room) or the balance centers (Marty’s bicycle). This is merely an accident of how our brains work; one could easily imagine a robot that could be told how to recognize red objects or balance a bicycle without having to try it first.
I once read an anecdote about a person who was determined to prove that “purely theoretical” knowledge was just as good as actual experience, and to prove it, he was going to teach himself how to swim by reading about it. After acquiring what he felt was a sufficient “theoretical” knowledge of swimming, he jumped into a pool and was immediately able to swim, much to the amazement of onlookers.
I don’t believe it. If I were one of the onlookers, I wouldn’t be amazed; I just wouldn’t believe his claim never to have swum before.
I was quite surprised to learn one day, after riding bicycles for years, that you steer by leaning to the side. This is information that is quite important for riding a bicycle that had never passed through the verbal part of my mind. And when I wanted to test this, I couldn’t ask my body about it. Instead I had to pay attention while riding to see what I did; also, I was able to lean or turn the handlebars and see what happens. The communication channel between the verbal part of the mind and the mechanical memory is quite narrow. That’s not to say that it’s impossible to send information in either direction, but that is a skill that is quite rare, not a matter of deciding to acquire “theoretical knowledge.” Also, there’s the problem that since people don’t normally learn this way, they don’t record instructions.
Swimming is a lot easier to figure out prior to doing than bike riding is. Falling, for example, is rather more difficult. It is highly unlikely that someone could do this for any sport or activity that involves significant interactive feedback (e.g. balance or aim).
It’s impossible to jump into a pool without learning something about swimming. At the very least, you start learning about how the water responds to your body. And this will remain true even if you don’t consciously try. (There are experiments, for example, where people “learn” to twitch a muscle they can’t consciously control because twitching it becomes correlated with turning off an annoying sound.)
Furthermore, it’s been my experience that people who think they “can’t explain something” really aren’t trying. If you really try (and I’m sure people have), you can come up with pointers that greatly demystify swimming, even if they require you to be in a pool to start making sense.
I don’t know what I would call the hard limits of a first-timer. If they swam at a competitive level the first time, and I had good reason to believe they’d never practiced, I’d have to admit I underestimated what you can learn without swimming. But it’s not unreasonable that someone who was both a quick learner and was told of a known motion that closely approximates good swimming, could swim on the first try.
The first time I went ice skating a couple years ago, I flailed around dangerously for a bit, and was unable to exert much control over my path. I comically ran into a female friend and wound up in an unflattering position, and it was forceful/painful enough for all present to understand that I hadn’t done it on purpose...
Then my roommate took me aside and explained what motion to do, with a few hand gestures. After about 45 seconds of explanation, it clicked in my mind, and I went out and was able to skate around quite gracefully in comparison to a few minutes before. I was surprised at how well it worked.
I suspect that this falls into the “quick learner who is told of a motion that approximates good skating,” though.
What you describe is what Tim Gallwey calls the ‘inner game’. It is, to simplify a bit, training your intuitive subconscious without letting your conscious awareness interfere. Here is a video of him coaching a woman who has never touched a tennis racket to serve using the technique.
The ability to recognize red objects is like the skill of riding a bicycle
I don’t think the question is about ability. Mary already has the ability to recognize red objects by using a light detector, for example. The question is that, when seeing red for the first time, it seems you’ve learned something new—i.e. what it feels like to see red.
The question boils down to whether subjective experiences are a form of knowledge.
This applies to humans, who have fixed systems for different kinds of learning. If marty or mary were AI’s they should be able to gain abilities through knowledge directly.
I basically agree with what you said, but I would go one step further.
That is, we don’t learn about redness as such. When we get the sensory stimuli of redness, we aren’t learning the ability to recognize red objects as such. What we are learning is things and causes in the world and redness is just one attribute of the sensory data. It is useful, because it makes the prediction of things in the world more accurate.
Let’s take one type of cause in the world that we learn: objects. When we are exposed to visual stimuli of a certain type of object, we learn to expect it to have some type of common pattern. One aspect of this is the color of the object. If the color is nearly always red, we learn to predict that this type of object is typically red. So our prediction of this type of cause in the world is enhanced by recognizing this attribute of the sensory data.
If the sensory data lacks color completely, or if the color is not what we expect it to be, the prediction of our brain is not as accurate. But we might be able to still make the correct inference of the world, it’s just a bit harder. If the color is completely different of what we’ve learned it to be, we are typically still able to recognize the object, but we are surprised about the color, because it doesn’t match our representation. Our attention is drawn to the peculiar color.
The high level representation of the object of certain type, then, is composed of these different aspects of what we have seen before. These different aspects are things like the pattern of the visual stimuli, the intensity of the visual signal, the color of the signal, etc. A person who has been exposed to color stimuli all her life, has these extra bits of information about objects of the world she has learned about. So “sensation of red” isn’t a thing or place in the brain. It’s not even a process in the brain. Redness (and color stimuli in general) is part of our knowledge of the world. It’s part of the objects and categories we have learned. It’s part of the representation of these objects in our brain.
So when Mary, in her black and white room, tries to learn about colors, she cannot directly influence her representation of the world in her brain. She would have to twiddle every piece of knowledge in her head with regards of its color.
As an example, let’s say she has seen a lot of roses from her black and white TV. (Let’s assume for simplicity that roses were pretty much always red.) Now, from her sources, she learns that roses are typically red. She would then have to locate the representation of roses in her head and add a probabilistic adjustment if the color of the visual stimuli is red. It’s something like: “if the visual pattern is like this and if the color is red, then it’s likely that’s it’s a rose.” Or more precisely, it’s some sort of conditional probabilistic table.
So for Mary to learn redness before being exposed to color stimuli, she would have to modify all the bits of knowledge of the world that she has acquired through visual data. This would mean all types of objects and other visual information she has. She would have to add this conditional clause to every type of object she has learned where the color red has any significance to the inference of the object. She would have to add this attribute also to every specific knowledge of the world she has too.
So in summary, this thought experiment seems to me like a huge misunderstanding of how we learn about the world.
Mary’s Room is invalid because it sneaks in a false assumption: that if something is knowledge, then it can be learned by study alone. As a counterexample to the story of Mary’s Room, consider the story of Marty’s Bicycle, which I have just made up.
Marty is a brilliant physicist who has suddenly become very interested in bicycles. Unfortunately, his leg is in a cast, so he can’t ride one. While he waits for it to heal, he studies their physics, watches people ride bikes, reads guides on bike riding, and interviews top racers—in fact, he studies everything there is to study about bicycles. When the cast is finally taken off, he buys a bicycle, gets on, and immediately falls over.
The ability to recognize red objects is like the skill of riding a bicycle—it can only be acquired by doing it, not by study, because study can only train the linguistic centers of the brain, not the visual processing centers (Mary’s room) or the balance centers (Marty’s bicycle). This is merely an accident of how our brains work; one could easily imagine a robot that could be told how to recognize red objects or balance a bicycle without having to try it first.
This is a standard reply, known as the Lewis-Nemirow “Ability Hypothesis”. See here for a critique.
I once read an anecdote about a person who was determined to prove that “purely theoretical” knowledge was just as good as actual experience, and to prove it, he was going to teach himself how to swim by reading about it. After acquiring what he felt was a sufficient “theoretical” knowledge of swimming, he jumped into a pool and was immediately able to swim, much to the amazement of onlookers.
It’s probably false, though.
I don’t believe it. If I were one of the onlookers, I wouldn’t be amazed; I just wouldn’t believe his claim never to have swum before.
I was quite surprised to learn one day, after riding bicycles for years, that you steer by leaning to the side. This is information that is quite important for riding a bicycle that had never passed through the verbal part of my mind. And when I wanted to test this, I couldn’t ask my body about it. Instead I had to pay attention while riding to see what I did; also, I was able to lean or turn the handlebars and see what happens. The communication channel between the verbal part of the mind and the mechanical memory is quite narrow. That’s not to say that it’s impossible to send information in either direction, but that is a skill that is quite rare, not a matter of deciding to acquire “theoretical knowledge.” Also, there’s the problem that since people don’t normally learn this way, they don’t record instructions.
Swimming is a lot easier to figure out prior to doing than bike riding is. Falling, for example, is rather more difficult. It is highly unlikely that someone could do this for any sport or activity that involves significant interactive feedback (e.g. balance or aim).
It’s impossible to jump into a pool without learning something about swimming. At the very least, you start learning about how the water responds to your body. And this will remain true even if you don’t consciously try. (There are experiments, for example, where people “learn” to twitch a muscle they can’t consciously control because twitching it becomes correlated with turning off an annoying sound.)
Furthermore, it’s been my experience that people who think they “can’t explain something” really aren’t trying. If you really try (and I’m sure people have), you can come up with pointers that greatly demystify swimming, even if they require you to be in a pool to start making sense.
I don’t know what I would call the hard limits of a first-timer. If they swam at a competitive level the first time, and I had good reason to believe they’d never practiced, I’d have to admit I underestimated what you can learn without swimming. But it’s not unreasonable that someone who was both a quick learner and was told of a known motion that closely approximates good swimming, could swim on the first try.
The first time I went ice skating a couple years ago, I flailed around dangerously for a bit, and was unable to exert much control over my path. I comically ran into a female friend and wound up in an unflattering position, and it was forceful/painful enough for all present to understand that I hadn’t done it on purpose...
Then my roommate took me aside and explained what motion to do, with a few hand gestures. After about 45 seconds of explanation, it clicked in my mind, and I went out and was able to skate around quite gracefully in comparison to a few minutes before. I was surprised at how well it worked.
I suspect that this falls into the “quick learner who is told of a motion that approximates good skating,” though.
Yep, sounds exactly right.
What you describe is what Tim Gallwey calls the ‘inner game’. It is, to simplify a bit, training your intuitive subconscious without letting your conscious awareness interfere. Here is a video of him coaching a woman who has never touched a tennis racket to serve using the technique.
Another similar technique is drawing on the right side of the brain.
Good analogy!
I don’t think the question is about ability. Mary already has the ability to recognize red objects by using a light detector, for example. The question is that, when seeing red for the first time, it seems you’ve learned something new—i.e. what it feels like to see red.
The question boils down to whether subjective experiences are a form of knowledge.
This applies to humans, who have fixed systems for different kinds of learning. If marty or mary were AI’s they should be able to gain abilities through knowledge directly.
I basically agree with what you said, but I would go one step further.
That is, we don’t learn about redness as such. When we get the sensory stimuli of redness, we aren’t learning the ability to recognize red objects as such. What we are learning is things and causes in the world and redness is just one attribute of the sensory data. It is useful, because it makes the prediction of things in the world more accurate.
Let’s take one type of cause in the world that we learn: objects. When we are exposed to visual stimuli of a certain type of object, we learn to expect it to have some type of common pattern. One aspect of this is the color of the object. If the color is nearly always red, we learn to predict that this type of object is typically red. So our prediction of this type of cause in the world is enhanced by recognizing this attribute of the sensory data.
If the sensory data lacks color completely, or if the color is not what we expect it to be, the prediction of our brain is not as accurate. But we might be able to still make the correct inference of the world, it’s just a bit harder. If the color is completely different of what we’ve learned it to be, we are typically still able to recognize the object, but we are surprised about the color, because it doesn’t match our representation. Our attention is drawn to the peculiar color.
The high level representation of the object of certain type, then, is composed of these different aspects of what we have seen before. These different aspects are things like the pattern of the visual stimuli, the intensity of the visual signal, the color of the signal, etc. A person who has been exposed to color stimuli all her life, has these extra bits of information about objects of the world she has learned about. So “sensation of red” isn’t a thing or place in the brain. It’s not even a process in the brain. Redness (and color stimuli in general) is part of our knowledge of the world. It’s part of the objects and categories we have learned. It’s part of the representation of these objects in our brain.
So when Mary, in her black and white room, tries to learn about colors, she cannot directly influence her representation of the world in her brain. She would have to twiddle every piece of knowledge in her head with regards of its color.
As an example, let’s say she has seen a lot of roses from her black and white TV. (Let’s assume for simplicity that roses were pretty much always red.) Now, from her sources, she learns that roses are typically red. She would then have to locate the representation of roses in her head and add a probabilistic adjustment if the color of the visual stimuli is red. It’s something like: “if the visual pattern is like this and if the color is red, then it’s likely that’s it’s a rose.” Or more precisely, it’s some sort of conditional probabilistic table.
So for Mary to learn redness before being exposed to color stimuli, she would have to modify all the bits of knowledge of the world that she has acquired through visual data. This would mean all types of objects and other visual information she has. She would have to add this conditional clause to every type of object she has learned where the color red has any significance to the inference of the object. She would have to add this attribute also to every specific knowledge of the world she has too.
So in summary, this thought experiment seems to me like a huge misunderstanding of how we learn about the world.