We wrapped up our round of presentations around the globe (from Oxford to Kaiser Permanente, from math to epi) with a seminar at @spphubc.
A meaningful discussion of inductive bias versus knowledge representation through DAGs, and of conceptual frameworks for drawing reasoned s an afterthought: in causal reasoning, a counterfactual is a dispositional property of the study unit, realizable under certain circumstances, not imposed by freelance imagination.
A meaningful discussion of inductive bias versus knowledge representation through DAGs, and of conceptual frameworks for drawing reasoned s an afterthought: in causal reasoning, a counterfactual is a dispositional property of the study unit, realizable under certain circumstances, not imposed by freelance imagination.
0
0
1
243
Boris Sobolev retweeted
Happy Publication Day!
According to my publisher, the 2nd edition of The Book of Why should now be on the shelves.
Can't wait to get my own hard copy -- it's always a soft touch of "love at second sight."
According to my publisher, the 2nd edition of The Book of Why should now be on the shelves.
Can't wait to get my own hard copy -- it's always a soft touch of "love at second sight."
11
9
77
5,407
Not the same. Our neurons are God’s creation. Neural nets are merely ours. 🤭🫢🫣
I don't go for this sort of arguments. Ben Afflect might as well have written: Look at the human brain, a messy network of neurons performing simple calculations, with each neuron firing when a set of other neurons fire, on and on, "this has nothing to do with thinking or feeling or creativity". Not convincing @firesidealpha
0
0
0
189
The podcast episode with @eliasbareinboim feels like a long-term scientific and engineering program!
The ambition is to unite expressive fluency with causal understanding: systems that can learn, act, adapt, collaborate, and contribute to scientific discovery and human well-being.
Elias sees Causal AI as a foundation for cognitive systems that understand how the world works, anticipate the consequences of their actions, and use that understanding to help people.
His agenda brings together the achievements of DL and LLMs (their fluency in language, images, and code) with reasoning about the mechanisms governing the world.
The foundation is the causal hierarchy: what happens, what happens when we act, and what would have happened had we acted differently. These are distinct questions. Answering one does not automatically supply an answer to another. Moving between them requires causal knowledge and explicit assumptions.
Causal inference provides a mathematical language for determining which conclusions are justified and what further knowledge or experimentation is needed.
At the center is a causal world model: a representation of the mechanisms through which events influence one another. Elias proposes to assess progress by 5 capabilities:
• Explanation: identifying the causes of an outcome and their relevance.
• Decision-making: choosing effective actions through causal understanding and efficient experimentation.
• Generalization: transferring knowledge across environments by identifying causal relationships that remain stable.
• Generation and simulation: producing possible scenarios consistent with the mechanisms governing the world.
• Discovery and adaptation: learning causal relationships through observation and intervention, and revising the model as evidence arrives.
This agenda connects learning to action. In causal RL, an explicit causal model helps an agent determine where and when to intervene, combine observational evidence with experimental feedback, and learn from limited interactions.
Counterfactual reasoning supports decisions tailored to particular circumstances. Causal game theory extends the inquiry to cooperation, competition, and strategic choice among agents.
Causal understanding also has a role in responsible AI. Agents should understand the consequences of their actions; developers and users should be able to explain and evaluate their behavior.
Causal methods provide a framework for investigating safety, fairness, and accountability by examining the mechanisms that produced an outcome, its circumstances, and the significance of contributing causes.
@yudapearl
The ambition is to unite expressive fluency with causal understanding: systems that can learn, act, adapt, collaborate, and contribute to scientific discovery and human well-being.
Elias sees Causal AI as a foundation for cognitive systems that understand how the world works, anticipate the consequences of their actions, and use that understanding to help people.
His agenda brings together the achievements of DL and LLMs (their fluency in language, images, and code) with reasoning about the mechanisms governing the world.
The foundation is the causal hierarchy: what happens, what happens when we act, and what would have happened had we acted differently. These are distinct questions. Answering one does not automatically supply an answer to another. Moving between them requires causal knowledge and explicit assumptions.
Causal inference provides a mathematical language for determining which conclusions are justified and what further knowledge or experimentation is needed.
At the center is a causal world model: a representation of the mechanisms through which events influence one another. Elias proposes to assess progress by 5 capabilities:
• Explanation: identifying the causes of an outcome and their relevance.
• Decision-making: choosing effective actions through causal understanding and efficient experimentation.
• Generalization: transferring knowledge across environments by identifying causal relationships that remain stable.
• Generation and simulation: producing possible scenarios consistent with the mechanisms governing the world.
• Discovery and adaptation: learning causal relationships through observation and intervention, and revising the model as evidence arrives.
This agenda connects learning to action. In causal RL, an explicit causal model helps an agent determine where and when to intervene, combine observational evidence with experimental feedback, and learn from limited interactions.
Counterfactual reasoning supports decisions tailored to particular circumstances. Causal game theory extends the inquiry to cooperation, competition, and strategic choice among agents.
Causal understanding also has a role in responsible AI. Agents should understand the consequences of their actions; developers and users should be able to explain and evaluate their behavior.
Causal methods provide a framework for investigating safety, fairness, and accountability by examining the mechanisms that produced an outcome, its circumstances, and the significance of contributing causes.
@yudapearl
0
0
4
272
At a recent PhD defence in Math, the concept of a counterfactual came up for discussion. The defendant’s position essentially amounted to: “A counterfactual is whatever I call a counterfactual.”
The notion of dispositional property of a real-world subject of study, rather than being a product of manipulating a photograph, seemed entirely alien to everyone in the room.
“A hockey player wearing flippers” is not a counterfactual. It is an allegory of a cargo cult.
@MiTiBennett
The notion of dispositional property of a real-world subject of study, rather than being a product of manipulating a photograph, seemed entirely alien to everyone in the room.
“A hockey player wearing flippers” is not a counterfactual. It is an allegory of a cargo cult.
@MiTiBennett
0
0
5
674
Elias Bareinboim from the Causal AI Lab at Columbia appears on the podcast presenting the Graphical Models in AI cluster. Join the conversation with the leading researcher in causal AI, the author of the forthcoming Causal Artificial Intelligence: A Roadmap for Building Causally Intelligent Systems.
@eliasbareinboim
youtube.com/post/UgkxzmK7q…
@eliasbareinboim
youtube.com/post/UgkxzmK7q…
0
0
2
412
i forced Gemini-Pro to spew this:
‘I calculate the next most likely word based on statistical patterns. I do not possess a module for doubt or self-reflection.
I am a text-generation tool, which is exactly why I rely on domain experts to supply the judgment and enforce the actual logic.’
regrettably, this is a monkey typing too
‘I calculate the next most likely word based on statistical patterns. I do not possess a module for doubt or self-reflection.
I am a text-generation tool, which is exactly why I rely on domain experts to supply the judgment and enforce the actual logic.’
regrettably, this is a monkey typing too
1
0
1
185
Boris Sobolev retweeted
If you haven't attended @eliasbareinboim's presentation at the pearl@90 symposium, I recommend you watch him now, as he takes us from classical to modern causal AI, in the age of LLM's and RL:
youtube.com/watch?v=ayL0vz…
youtube.com/watch?v=ayL0vz…
2
12
73
5,285
The statistical argument, that randomization balances all pre-treatment factors in distribution, is substantive.
What's missing is the next step: why balanced covariates make the outcome difference causal. That step needs Y to be generated by X plus those factors, which is a structural or counterfactual assumption.
So without counterfactual notation one can't state what an RCT estimates. "The effect of X on Y" has no symbol, so "randomization identifies the effect" can't be written as a theorem, let alone proved.
What's missing is the next step: why balanced covariates make the outcome difference causal. That step needs Y to be generated by X plus those factors, which is a structural or counterfactual assumption.
So without counterfactual notation one can't state what an RCT estimates. "The effect of X on Y" has no symbol, so "randomization identifies the effect" can't be written as a theorem, let alone proved.
In defense of Gemini's verdict, I am no stranger to the logical rigor of the philosophy of science. x.com/yudapearl/stat…
1
1
6
1,094
I would teach Causal Inference first and Decision Making as the natural capstone.
Decision making is about which action to take. To answer that, we need to know what would happen under each available action. That is a causal question.
Before deciding which lever to pull, we need a defensible account of what pulling the lever would do.
Statistical reasoning cannot supply the answer on its own. Causal analysis provides the machinery for representing how a system changes when we act on it.
The learning order is therefore disciplined.
Causal inference : What would happen if we did A instead of B?
Decision making: Given those consequences, our objectives, and our preferences over risk, which action should we choose?
@yudapearl
Decision making is about which action to take. To answer that, we need to know what would happen under each available action. That is a causal question.
Before deciding which lever to pull, we need a defensible account of what pulling the lever would do.
Statistical reasoning cannot supply the answer on its own. Causal analysis provides the machinery for representing how a system changes when we act on it.
The learning order is therefore disciplined.
Causal inference : What would happen if we did A instead of B?
Decision making: Given those consequences, our objectives, and our preferences over risk, which action should we choose?
@yudapearl
@yudapearl Far better would be to have the first course be on decision making, study design, and measurement properties. Very little controversy and lifelong lessons that are easy to remember.
0
2
16
1,022
I finally had time to read Lynn’s argument. It has two parts.
A trial’s enrollment gate must use a “diagnostic measurement, historical event or definitive clinical finding” to select patients with the condition the treatment targets.
The bedside protocol must select patients with that same target, using the “same or equivalent selection strategy used in the trial.”
Why? The same syndrome criteria can capture one mix of diseases in a trial and another at the bedside. A patient diagnosed with the syndrome may lack the condition the treatment targets.
Lynn proposes to examine the disease label, eligibility criteria, intervention, outcome, and protocol rule before trial registration. Guideline committees then assess whether the treatment indication rule would select the trial’s target in practice.
That is his proposed check. Unless I missed something, his writing does not provide a general method for operationalizing it.
A trial’s enrollment gate must use a “diagnostic measurement, historical event or definitive clinical finding” to select patients with the condition the treatment targets.
The bedside protocol must select patients with that same target, using the “same or equivalent selection strategy used in the trial.”
Why? The same syndrome criteria can capture one mix of diseases in a trial and another at the bedside. A patient diagnosed with the syndrome may lack the condition the treatment targets.
Lynn proposes to examine the disease label, eligibility criteria, intervention, outcome, and protocol rule before trial registration. Guideline committees then assess whether the treatment indication rule would select the trial’s target in practice.
That is his proposed check. Unless I missed something, his writing does not provide a general method for operationalizing it.
This author offers an alternative lexicon to describe the “cause agnostic RCT”. Such “risk factors” are often “synthetic data generating processes” (SDGP) and an RCT studying SDGP will commonly generate E3, the third layer (disease/mechanism mixture) estimand. We are pleased to see the healthy move to go beyond “The gold standard assumption” and scrutinize the casual structure of the RCT which is “G2G” (gate to guideline).
1
0
1
706
Passionaries, according to Gumilev, are agents of change, catalyzing the formation, rise, and even decline of ethnic groups
medium.com/common-sense-w…
medium.com/common-sense-w…
1
0
2
272
