This looks very interesting. I always felt most of ML is about one aspect of intelligence - prediction from existing data. While this may work for specific tasks like driving a car or analysing a scanned image, the human intelligence has always been about the Why? on anything. Answering the Why? on a pile of data by connecting the dots using a causal model is also prediction in a sense, but a more generic one than a prediction on a bunch of specific classes or outcomes. For example, a self-driving car algorithm trained on a pre-classified data to detect obstacles versus a algorithm that can answer Why something is a obstacle is vastly different and is much more effective I guess.
> The equations of physics are algebraic and symmetrical, whereas causal relationships are directional.
I don't agree with this. Quantum measurements are projective and they are very much one way. People seem to want to dismiss this as being just "epistemic" like the way entropy (thermodynamics) is one-way, but not fundamentally one way. Entropy increases only because we can't see all the details. Quantum measurements are not like that.
As far as I know the theory does not completely exclude time-reversal formulations of QM. Non-symmetry of quantum measurements has not been experimentally confirmed.
The only confirmed T-symmetry violation is that of K and B mesons, I think. But that does not follow from quantum theory I think. It's experimental result.
I remember reading a post that entropy increases because we classify higher entropy states as essentially the same and there are more of them. Is there a good explanation of this?
Actually, quantum 'measurements' are like that; decoherence happens when a entangled photon (or whatever paticle) escapes far enough away that you can't retrieve it and force it's horizonal and vertical states to feed amplitude into the same outcome. This generally happens because you lost track of that photon (and countless others), the same way you lose track of the blackbody radiation photons your heat engine emitted to cool down.
Edit: And actually, causal relationships aren't directional, they just look that way because we usually care about "What can we infer about the more complicated [thermodynamically later] system state given knowledge about the less complicated [thermodynamically earlier] one?" rather than vice versa. (Well, also because most physical processes exchange information about about particle velocities (heat differentials) for information about the state and layout of macroscopic-ish objects (knowledge), and the laws of thermodynamics result in those exchanges all going in the same direction.)
It’s important to note this is causal inference from Pearl’s darling theory using do-calculus and graphical models.
I work in causal inference professionally (large scale causal impact measurement for observational studies that customers require causal interpretations of, for clinical trial data and advertising data mostly).
In practice, the stuff from Pearl is just unhelpful, and it’s very valuable to also read sources from Rubin, Imbens, Gelman and many others, because there are whole other approaches to the problem, approaches which more or less completely sidestep and have no need for do-calculus concepts, but which are extremely successful in practice, with real observational data sets.
Methods like propensity matching, propensity weighting, and hierarchical models using different treatment levels as strata, are absolutely critical for applying this stuff in the real world.
If you only ever read Pearl, you’ll mistakenly think those methods are somehow “inferior to” or “subsumed by” do-calculus when this is not at all true, certainly not in practice.
My understanding is that the do is needed for counterfactual reasoning, so if I were to do x assuming condition y. And that is lacking in other approaches? Or am I mistaken?
You definitely don’t need do for counterfactual reasoning, that is a big misconception. The misconception is usually rooted in a misunderstabding of what “doing causal inference” means, and what specific assumptions are being made, e.g. exclusion restriction, SUTVA, exogeniety in a propensity model, balance between treated and non-treated groups.
In some cases, you would need the framework of do-calculus, but very often when other assumptions are reasonable, you don’t.
When using matching, regression adjustment, propensity techniques, or hierarchical models that separate strata based on the treatment to measure causal effects of treatments, and you believe the underlying assumptions are generally satisfied, you have utterly no need for anything do-calculus related, not even theoretically, and certainly not in terms of practically actually fitting models to the data.
The cases where you do need it are in creating a generalized AI, perhaps your use case is only for epidemiology, and of course that’s Perl’s background too. But his arguing for general purpose AI and using his epidemiological expirence for demonstrating since it’s what he knows well. That’s my understanding from The Book of Why at least.
> Methods like propensity matching, propensity weighting, and hierarchical models using different treatment levels as strata, are absolutely critical for applying this stuff in the real world.
Have you got some readings? I've been reading the Pearl books on this subject.
A good starting point is chapters 9 & 10 from [0]. Then many of the same topics are re-discussed in the second half of the book through the lens of Bayesian hierarchical models.
Another good reference is [1]. Rubin invented a lot of observational data methods for correcting to measure causal effect. Imbens is also a prolific author in this area, and even just googling for propensity model papers from Imbens will leads to many methods and many other papers.
there are whole other approaches to the problem... which are extremely successful in practice, with real observational data sets.
I'm going to echo my sibling commenter, would love to hear it if you have any pointers on where to start reading.
I'm a software engineer but data work comes up all the damn time. Seeing coworkers approach causality in a hand-wavy way without knowing a more rigorous approach to suggest is frustrating! I had hoped Pearl's book was exactly that.
>> The astonishing success of big-data and machine learning reflects our
under-estimating how much can be achieved by the low hanging fruits of
model-free curve-fitting. But when we look at the limitations unveiled by the
calculus of causation we understand that human-level AI requires two more
layers: intervention and counterfactuals.
Yes, well, the problem with that is that the vast majority of researchers in
AI know very well that human-level AI is many, many years away still. Whereas
those "low-hanging fruit"? They're just hanging there, ripe for the picking
and large companies are very eager to throw a shitload of money at people who
can pick them, right. now.
And- let's be fair. Anyone who knows how to do this "model-free curve-fitting"
that the good professor so despises has a brilliant career for upwards of 30
years laid out for them- and those are 30 years in which they won't have to
think about causality and Judea Pearl even once.
Very interesting. As someone with a background in philosophy, I am wondering if causality has been excluded because of mistaken metaphysical and epistemological assumptions when the sciences were originally developed.
>> JP: Correct. Formally, Bayesian networks are just efficient
evidence-to-hypothesis inference machines. However, in retrospect, their
success emanated from their ability to “secretly” represent causal knowledge.
In other words, they were almost always constructed with their arrows pointing
from causes to effect, thus achieving modularity. It is only due to our
current understanding of causality that we can reflect back and speculate on
why they were successful; we did not know it then.
Hang on a sec. If Bayesian networks are perfectly capable of representing
causality relations, and in fact they've been doing just that all along
(albeit "secretely") then why the hell do we need a different formalism to
represent causality?
To give an analogy - if we can represent context-free langugaes with regular
automata, then what's the point of context-free languages? Instead, we
classify languages that can be represented by both regular automata and
pushdown automata as regular, and reserve the context-free designation for
languages that cannot be represented by regular (or finite) automata.
In the same way, if causality relations can be represented by Bayesian
networks, then higher-order representations are not really needed, or must be
reserved for some object that Bayesian networks can't represent.
In any case, this is just a huge piece of ret-con. Bayesian networks always
represented causality relations, only they did so "secretly"! That's up there
with the original Klingon's flat heads being the result of a virus infection;
or how Jean Grey didn't really die and it was the Phoenix Force that had taken
her form.
22 comments
[ 2.4 ms ] story [ 62.4 ms ] threadI don't agree with this. Quantum measurements are projective and they are very much one way. People seem to want to dismiss this as being just "epistemic" like the way entropy (thermodynamics) is one-way, but not fundamentally one way. Entropy increases only because we can't see all the details. Quantum measurements are not like that.
The only confirmed T-symmetry violation is that of K and B mesons, I think. But that does not follow from quantum theory I think. It's experimental result.
Edit: And actually, causal relationships aren't directional, they just look that way because we usually care about "What can we infer about the more complicated [thermodynamically later] system state given knowledge about the less complicated [thermodynamically earlier] one?" rather than vice versa. (Well, also because most physical processes exchange information about about particle velocities (heat differentials) for information about the state and layout of macroscopic-ish objects (knowledge), and the laws of thermodynamics result in those exchanges all going in the same direction.)
I work in causal inference professionally (large scale causal impact measurement for observational studies that customers require causal interpretations of, for clinical trial data and advertising data mostly).
In practice, the stuff from Pearl is just unhelpful, and it’s very valuable to also read sources from Rubin, Imbens, Gelman and many others, because there are whole other approaches to the problem, approaches which more or less completely sidestep and have no need for do-calculus concepts, but which are extremely successful in practice, with real observational data sets.
Methods like propensity matching, propensity weighting, and hierarchical models using different treatment levels as strata, are absolutely critical for applying this stuff in the real world.
If you only ever read Pearl, you’ll mistakenly think those methods are somehow “inferior to” or “subsumed by” do-calculus when this is not at all true, certainly not in practice.
In some cases, you would need the framework of do-calculus, but very often when other assumptions are reasonable, you don’t.
When using matching, regression adjustment, propensity techniques, or hierarchical models that separate strata based on the treatment to measure causal effects of treatments, and you believe the underlying assumptions are generally satisfied, you have utterly no need for anything do-calculus related, not even theoretically, and certainly not in terms of practically actually fitting models to the data.
Have you got some readings? I've been reading the Pearl books on this subject.
Another good reference is [1]. Rubin invented a lot of observational data methods for correcting to measure causal effect. Imbens is also a prolific author in this area, and even just googling for propensity model papers from Imbens will leads to many methods and many other papers.
[0]: < http://www.stat.columbia.edu/~gelman/arm/ >
[1]: < https://www.amazon.com/Causal-Inference-Statistics-Biomedica... >
I'm going to echo my sibling commenter, would love to hear it if you have any pointers on where to start reading.
I'm a software engineer but data work comes up all the damn time. Seeing coworkers approach causality in a hand-wavy way without knowing a more rigorous approach to suggest is frustrating! I had hoped Pearl's book was exactly that.
Yes, well, the problem with that is that the vast majority of researchers in AI know very well that human-level AI is many, many years away still. Whereas those "low-hanging fruit"? They're just hanging there, ripe for the picking and large companies are very eager to throw a shitload of money at people who can pick them, right. now.
And- let's be fair. Anyone who knows how to do this "model-free curve-fitting" that the good professor so despises has a brilliant career for upwards of 30 years laid out for them- and those are 30 years in which they won't have to think about causality and Judea Pearl even once.
Hang on a sec. If Bayesian networks are perfectly capable of representing causality relations, and in fact they've been doing just that all along (albeit "secretely") then why the hell do we need a different formalism to represent causality?
To give an analogy - if we can represent context-free langugaes with regular automata, then what's the point of context-free languages? Instead, we classify languages that can be represented by both regular automata and pushdown automata as regular, and reserve the context-free designation for languages that cannot be represented by regular (or finite) automata.
In the same way, if causality relations can be represented by Bayesian networks, then higher-order representations are not really needed, or must be reserved for some object that Bayesian networks can't represent.
In any case, this is just a huge piece of ret-con. Bayesian networks always represented causality relations, only they did so "secretly"! That's up there with the original Klingon's flat heads being the result of a virus infection; or how Jean Grey didn't really die and it was the Phoenix Force that had taken her form.