But, an important question still needs to be asked. Is sensory perception and how its input gets organised in our minds the sole basis of our internal representations of the world or is there something else that might placate any creeping errors from perception?
Yes, there is: action. Organisms learn about the world not merely by sitting there looking, but by acting and observing the results.
Action still depends on perception. I would agree that there is more, e.g. proprioception, but at the end of the day you are still just learning about the environment by interacting with it and analysing the feedback. This interaction can be action or it can be simply looking. I agree with your overall point, though, if you want to know something about the territory, you actually have to interact with it. You can’t rely on thought experiments or your maps as they can be faulty and the territory could have changed since you last interacted with it and built up your models of it.
Scott Aaronson once posed the question in his blog, of whether an entity with only the power to observe, not to act, could discover causal relationships. Despite the eminence of some of the people commenting there, it seems to me that they came up with little of substance.
I think we have to first establish what it means to learn a causal relationship (e.g. we want both “true” and “justified” and agree on the “justified” part), and decide on how we are going to formalize causality. If we are Popperian that would mean we can test it. If we don’t believe in an adversarial universe, and like minimum description length or something, then we don’t have to test, but we then have to justify our optimism.
We could formalize causation via interventions (as I would), or via something else.
If we are interventionist Popperians, then from the point of view of the entity itself, the answer is “no” (because causal relationships are defined by acting, if we are interventionists, and the entity can never act to verify what it learned as genuinely causal, nor can it verify the assumptions it needs to equate observational constraints with causation). If we are interventionist optimists, the entity could perhaps construct an MDL story to make itself believe it learned a causal relationship. But if it can’t act, it just feels like a mental game it’s playing with itself, not genuine knowledge.
From the point of view of some hypothetical omniscient being that programmed the entity’s universe, the answer could very well be “yes”, regardless (if, for example the entity is running some big data version of the FCI algorithm, or something like that). The entity will run into certain limits on what is possible, but there exist universes we could program where the entity could extract information about causal structure via conditional independence tests (but only given some assumptions the entity itself cannot test, but that the omniscient universe programmers know happen to hold).
If we are not interventionists, then I don’t know. The answer may depend on how you formalize non-interventionist causality.
So: given the way I think about what it means to learn a causal relationship (I am a Popperian interventionist), the answer is: “no from the point of view of the entity itself, but possibly yes from the God’s eye view of the creators of the entity’s universe.”
The question Scott originally posed is whether we must intervene to discover causes—that is, whether interventionism (we must) or non-interventionism (we need not) is correct.
The answer may depend on how you formalize non-interventionist causality.
Is there such a formalism? Even with formalisations of both, that would just define the two positions more precisely. There is still the question of which of the two describes the world.
The question Scott originally posed is whether we must intervene to discover causes—that is, whether
interventionism (we must) or non-interventionism (we need not) is correct.
I disagree with your interpretation of Scott’s question (that is I agree, with the first, but not the second part of your sentence). But Scott is aware of this thread, he may pipe up himself!
There are two ways interventions come up: definitionally and operationally. We may choose to define causation via interventions or not, and we may need to resort to interventions (or not..) to discover causation. A ton of modern causal inference is precisely about finding ways of avoiding having to resort to interventions while still trying to get at causation defined via interventions.
I also think a big part of Scott’s question is about justifying what you believe.
Is there such a formalism?
Sure. Hume himself had like 3-4 definitions on his own (including his “proto-counterfactual” definition, which is fairly close to how I think about it).
There is still the question of which of the two describes the world.
I am not sure if causality is in the territory or not. I used to think “no,” but now I am not sure.
The question Scott originally posed is whether we must intervene to discover causes—that is, whether interventionism (we must) or non-interventionism (we need not) is correct.
I disagree with your interpretation of Scott’s question (that is I agree, with the first, but not the second part of your sentence). But Scott is aware of this thread, he may pipe up himself!
I hadn’t noticed that the OP is also called Scott. The Scott I intended to refer to there was Aaronson. ScottL’s question doesn’t mention causation, but is more general: is perception alone enough to learn everything that we learn? To which I am suggesting that no, action is also required, although I don’t have a rigorous argument to that effect.
Human babies may learn causality from observations, but:
(a) In this scenario the baby is the armchairian, and the parents are the omniscient Universe programmers. The parents know the baby is right in some cases, but how does the baby itself know it is right, without actually pushing things around. Armchairians aren’t allowed to push things around, either to learn or to verify what they learned.
(b) I think ScottL’s question is also about justifying your beliefs (I don’t think it is such an easy problem). But I am (perhaps naturally) more interested in Scott Aaronson’s question.
The parents know the baby is right in some cases, but how does the baby itself know it is right, without actually pushing things around.
The baby is pushing things around, not being an Armchairian. That’s what I intended as the point of the video. Causality is one of the earliest things we learn, even before walking and talking, and we learn it by acting and perceiving the effects.
The question Scott originally posed is whether we must intervene to discover causes—that is, whether interventionism (we must) or non-interventionism (we need not) is correct.
I don’t know about all that.. My point was that we don’t know anything apriori. That is everything that what we do know has a basis in interaction with the world or is from inference from premises. So, if we know something then at its root this knowledge must have been learnt from us or our ancestors interacting with the world and percieving the feedback. Even observation is interaction, see observer effect. The reason that we can only learn from intervention is that we cannot observe without intervention. We are not observers of the universe, but are part of the universe. It is impossible for us not to intervene when observing or learning about how the world is. We can, however, lessen the impact from our interventions to some degree.
I disagree with your interpretation of Scott’s question (that is I agree, with the first, but not the second part of your sentence). But Scott is aware of this thread, he may pipe up himself!
You guys seem to be asking a different question. Namely, whether an entity with only the power to observe, not to act, could discover causal relationships. I don’t really know the exact meaning of what it means to observe only without acting or what it takes to discover a causal relationship.
Maybe, it would help if you try thinking about at what level of intervention you can no longer discover causal relationships. At what point in the below points, assuming you trust the results and instruments etc, can you no longer discover causal relationships?
Someone performs an action and notices the effect.
Someone sets up a scientific experiment with scientific instruments and notices the results and what those instruments observed
Someone reads about the results of the scientific experiment
An entity that can only observe sees the results of the scientific experiment
At what point in the below points, assuming you trust the results and instruments etc, can you no longer discover causal relationships?
For the first three, clearly you can. The fourth is the tricky case. The difference between the third and the fourth is that in the third, the “someone” has already learned about causation, so when they read about what was done, it is as good as having done it themselves. At least, they will understand the causal relationships claimed, even if the paper does not contain enough detail for them to replicate it.
In the fourth case, the Armchairian (Scott Aa’s name for them) has never interacted with the world, only watched it as if on a television screen or through a read-only internet connection. One can consider two different versions of the Armchairians, depending on whether they have the power to direct their gaze wherever they choose or not (or whether they have the power to type in URLs or not), but in either case it is not clear to me what the answer is.
To give a more concrete example than the parable of the Armchairians, which could be run as a practical causal analysis challenge, could a program whose only input was the Facebook firehose discover causal relationships in the data?
Yes, there is: action. Organisms learn about the world not merely by sitting there looking, but by acting and observing the results.
Action still depends on perception. I would agree that there is more, e.g. proprioception, but at the end of the day you are still just learning about the environment by interacting with it and analysing the feedback. This interaction can be action or it can be simply looking. I agree with your overall point, though, if you want to know something about the territory, you actually have to interact with it. You can’t rely on thought experiments or your maps as they can be faulty and the territory could have changed since you last interacted with it and built up your models of it.
Scott Aaronson once posed the question in his blog, of whether an entity with only the power to observe, not to act, could discover causal relationships. Despite the eminence of some of the people commenting there, it seems to me that they came up with little of substance.
I think we have to first establish what it means to learn a causal relationship (e.g. we want both “true” and “justified” and agree on the “justified” part), and decide on how we are going to formalize causality. If we are Popperian that would mean we can test it. If we don’t believe in an adversarial universe, and like minimum description length or something, then we don’t have to test, but we then have to justify our optimism.
We could formalize causation via interventions (as I would), or via something else.
If we are interventionist Popperians, then from the point of view of the entity itself, the answer is “no” (because causal relationships are defined by acting, if we are interventionists, and the entity can never act to verify what it learned as genuinely causal, nor can it verify the assumptions it needs to equate observational constraints with causation). If we are interventionist optimists, the entity could perhaps construct an MDL story to make itself believe it learned a causal relationship. But if it can’t act, it just feels like a mental game it’s playing with itself, not genuine knowledge.
From the point of view of some hypothetical omniscient being that programmed the entity’s universe, the answer could very well be “yes”, regardless (if, for example the entity is running some big data version of the FCI algorithm, or something like that). The entity will run into certain limits on what is possible, but there exist universes we could program where the entity could extract information about causal structure via conditional independence tests (but only given some assumptions the entity itself cannot test, but that the omniscient universe programmers know happen to hold).
If we are not interventionists, then I don’t know. The answer may depend on how you formalize non-interventionist causality.
So: given the way I think about what it means to learn a causal relationship (I am a Popperian interventionist), the answer is: “no from the point of view of the entity itself, but possibly yes from the God’s eye view of the creators of the entity’s universe.”
...
The question Scott originally posed is whether we must intervene to discover causes—that is, whether interventionism (we must) or non-interventionism (we need not) is correct.
Is there such a formalism? Even with formalisations of both, that would just define the two positions more precisely. There is still the question of which of the two describes the world.
I disagree with your interpretation of Scott’s question (that is I agree, with the first, but not the second part of your sentence). But Scott is aware of this thread, he may pipe up himself!
There are two ways interventions come up: definitionally and operationally. We may choose to define causation via interventions or not, and we may need to resort to interventions (or not..) to discover causation. A ton of modern causal inference is precisely about finding ways of avoiding having to resort to interventions while still trying to get at causation defined via interventions.
I also think a big part of Scott’s question is about justifying what you believe.
Sure. Hume himself had like 3-4 definitions on his own (including his “proto-counterfactual” definition, which is fairly close to how I think about it).
I am not sure if causality is in the territory or not. I used to think “no,” but now I am not sure.
I hadn’t noticed that the OP is also called Scott. The Scott I intended to refer to there was Aaronson. ScottL’s question doesn’t mention causation, but is more general: is perception alone enough to learn everything that we learn? To which I am suggesting that no, action is also required, although I don’t have a rigorous argument to that effect.
This is how humans learn causality.
I meant Aaronson, also.
Human babies may learn causality from observations, but:
(a) In this scenario the baby is the armchairian, and the parents are the omniscient Universe programmers. The parents know the baby is right in some cases, but how does the baby itself know it is right, without actually pushing things around. Armchairians aren’t allowed to push things around, either to learn or to verify what they learned.
(b) I think ScottL’s question is also about justifying your beliefs (I don’t think it is such an easy problem). But I am (perhaps naturally) more interested in Scott Aaronson’s question.
The baby is pushing things around, not being an Armchairian. That’s what I intended as the point of the video. Causality is one of the earliest things we learn, even before walking and talking, and we learn it by acting and perceiving the effects.
I agree!
I don’t know about all that.. My point was that we don’t know anything apriori. That is everything that what we do know has a basis in interaction with the world or is from inference from premises. So, if we know something then at its root this knowledge must have been learnt from us or our ancestors interacting with the world and percieving the feedback. Even observation is interaction, see observer effect. The reason that we can only learn from intervention is that we cannot observe without intervention. We are not observers of the universe, but are part of the universe. It is impossible for us not to intervene when observing or learning about how the world is. We can, however, lessen the impact from our interventions to some degree.
You guys seem to be asking a different question. Namely, whether an entity with only the power to observe, not to act, could discover causal relationships. I don’t really know the exact meaning of what it means to observe only without acting or what it takes to discover a causal relationship.
Maybe, it would help if you try thinking about at what level of intervention you can no longer discover causal relationships. At what point in the below points, assuming you trust the results and instruments etc, can you no longer discover causal relationships?
Someone performs an action and notices the effect.
Someone sets up a scientific experiment with scientific instruments and notices the results and what those instruments observed
Someone reads about the results of the scientific experiment
An entity that can only observe sees the results of the scientific experiment
For the first three, clearly you can. The fourth is the tricky case. The difference between the third and the fourth is that in the third, the “someone” has already learned about causation, so when they read about what was done, it is as good as having done it themselves. At least, they will understand the causal relationships claimed, even if the paper does not contain enough detail for them to replicate it.
In the fourth case, the Armchairian (Scott Aa’s name for them) has never interacted with the world, only watched it as if on a television screen or through a read-only internet connection. One can consider two different versions of the Armchairians, depending on whether they have the power to direct their gaze wherever they choose or not (or whether they have the power to type in URLs or not), but in either case it is not clear to me what the answer is.
To give a more concrete example than the parable of the Armchairians, which could be run as a practical causal analysis challenge, could a program whose only input was the Facebook firehose discover causal relationships in the data?