Causal Wizard retweeted
I just posted a new paper on arXiv:
“Active perception and disentangled representations allow continual, episodic, zero-shot and few-shot learning”
The core idea: stop forcing fast learners to generalize. Delegate generalization to a perceptual system.
arxiv.org/abs/2602.19355
Causal Wizard retweeted
Replying to @CausalWizard
Cool! This reminds me of the fixed-point phenomenon discussed in arxiv.org/abs/2210.12867#:~:… and Fixed point reusing arxiv.org/abs/2204.08442 ! In the latter one, it also actually mentioned the deep supervision step. It’s intriguing that TRM and HRM aren’t explicitly trained with the DEQ method, yet the models still learn to converge toward fixed points, which represents the final answer. Still much to explore about why.
Causal Wizard retweeted
HRM-Agent: Using the Hierarchical Reasoning Model in Reinforcement Learning
Paper: arxiv.org/abs/2510.22832
The Hierarchical Reasoning Model (HRM) has impressive reasoning abilities given its small size, but has only been applied to supervised, static, fully-observable problems.
Causal Wizard retweeted
There is a result in causal inference called the 'causal hierarchy theorem' (Thm. 27.1, dl.acm.org/doi/10.1145/35017…), which serves as our 'no free lunch' theorem for inferences across the layers of the Pearl Causal Hierarchy (PCH). As discussed in Sec. 27.3, if we have observational data (from layer 1), no inferences about interventions (layer 2) or counterfactuals (layer 3) can be made without further assumptions. We can jump, scream, or cry (...), but these inferences are not generally possible (with Lebesgue measure 0). Of course, learning the laws of physics from observational data is only possible under special conditions. This observation motivates us to study causal inference and the kinds of inductive biases
and assumptions needed to extrapolate across distributions induced by the underlying system.
@analisereal @ccaballeroh10 @aleixrvr @yudapearl @GaryMarcus @VC31415
Let me put it another way. Would you say that in physical settings, PCH layers 3 and 2 'collapse' to layer 1 (Sec. 1.3, around p. 20, causalai.net/r60.pdf)? If so, this would imply that from purely observational data, one should be able to recover the entire SCM, and the laws of physics. This is not true in general, of course; but is it something specific about this particular setting discussed that would imply collapsibility?
@yudapearl @GaryMarcus
Causal Wizard retweeted
The paper introduces causal graphical models as a framework for addressing nonresponse and selection bias in surveys.
The authors show how to 1) to improve survey methods using causal graphs, 2) encode assumptions about nonprobability samples and survey nonresponse, and 3) determine whether it's possible to identify population quantities.
• Common practice of selecting adjustment variables based on correlations with sample selection and outcome variables is not well-justified.
• Design weights without nonresponse adjustment often rely on strong, but unrealistic, implicit assumptions.
• Using DAGs, the authors walk us through the occurrence of selection bias at sampling, response, and analysis.
• The paper illustrates how conditional independence assumptions for nonresponse adjustment can be intuitively represented and analyzed using causal graphs.
• The authors argue that imperfect nonresponse weighting may be worse than no weighting at all in causal inference from survey data.
• The paper shows how graphs can help understand collider bias and other subtle selection effects in survey data.
• The paper discusses applications of causal graphs to other survey methodology challenges like measurement error and transportability of causal effects.
Causal Wizard retweeted
Do you want to use causal graphs to improve your research but never know where to start?
In a series of papers, we used the same data and research question—one where we are pretty darn sure we know the real causal effect is that there isn’t one—to compare graph drawing methods.
Causal Wizard retweeted
The current state of AI is that a third of Americans used AI in any given week in August 2024, mostly ChatGPT but a decent amount of Gemini.
This is a pretty ubiquitous technology at this point. static1.squarespace.com/stat…
Causal Wizard retweeted
We are now in receipt of edited video recordings of all sessions from the last Causal AI conference in San Francisco (C-AI)
youtube.com/playlist?list=PL…
My Fire Side Chat is recorded here: youtube.com/watch?v=kgfwav5X…
Causal Wizard retweeted
Our weekly harvest of new papers of Causal Inference is knocking at our door: ucla.in/3Y8OFlN,
It contains of 8 papers presented at the 9th Causal Inference Workshop at UAI 2024, July 19, Barcelona, which I had to miss.
Causal Wizard retweeted
🚨New paper!🚨Real-world data can be used to contextualize findings of single-arm trials, but what if endpoints are measured differently in trial vs. real-world settings? We demystify bias in #RWD outcomes due to measurement error and quantify its impact: frontiersin.org/journals/dru…
Causal Wizard retweeted
😀Our 5th #Fribourg #WinterSchool in #DataAnalytics & #MachineLearning is happening from Feb 3-14 2025 at @ses_unifr @unifr!Registration is already open for our hybrid courses in machine/deep learning & #CausalAnalysis using #Python,#RStats,#Julia,#Knime:
unifr.ch/appecon/en/wintersc…
Causal Wizard retweeted
I am in receipt of a new book titled "Causal Artificial Intelligence", by Judith Hurwitz and John Thompson, which I highly recommend: amazon.com/Causal-Artificial…
This book bridges the gap between conventional
methods of business analytics and modern techniques of causal analysis and machine learning. Business leaders and data scientists trained in prescriptive statistics, simulation, and optimization techniques, will be happy to discover how critical problems in their field, commonly left to intuition or contentious judgement can now be systematically conceptualized, articulated, and solved using simple techniques of causal artificial intelligence.
The lucid exposition of this book and the variety of practical examples provided promise to make business analytic the new beneficiary of the causal revolution.
Causal Wizard retweeted
This was a shocking book. I read it sometime last year.
Every Machine Learning practitioner should learn about causal inference. This is not optional.
It's a different way of thinking. It will help you look at the world with different eyes.
amzn.to/3wPqcHb
Causal Wizard retweeted
We aim to build opportunities for accessible #causalinference research.
@CAUSALab awarded 42 tuition waivers for students to join us for #2024CAUSALabcourses!
Recipients represented:
🎓 25 institutions
🌎 13 countries
📍 7 US states
Congrats to this year's cohort!
Causal Wizard retweeted
Excited to present recent work from the Causal ML lab @SpotifyResearch, about estimating long-term effects when only short-run experiments are available (work joint w/ @notoriousgvg @lucasmaystre), at the first @nabe_econ Tech Economics Seminar this week.
The first #Tech #Economics Seminar in this summer's NABE members-only series is this Thursday, June 20 at 12 PM EDT - "Using Short-Term Metrics to Study Long-Term Outcomes" with @quantumciaran, @nathankallus & @Wayfair's Colin Gray as moderator. Sign up for the full series now!⤵️
Causal Wizard retweeted
Meehl’s pleas for reproducibility may sound depressingly familiar. But, even though it’s slow and painful, we’re making progress. argmin.net/p/towards-interpr…
Causal Wizard retweeted
#statstab #120 Why do we do matching for causal inference vs regressing on confounders?
Thoughts: A funny and insightful #stackexchange discussion on the pros and cons of causal inference and matching.
#causalinference #stats #matching #confounds
stats.stackexchange.com/ques…
Causal Wizard retweeted
If you want to see how ChatGpt fails on understanding causation, ask it: Can you give me a real-life example of three events, A, B, C, where A and B are correlated, B and C are correlated, but A and C are uncorrelated.
This question could be part of the Turing Test for deciding if a system has world model or not.
@GaryMarcus @ylecun @YoavShoham67461 #Bookofwhy
This Post is from an account that no longer exists. Learn more
Causal Wizard retweeted
Our paper (joint with Michela Bia & @lukaslaffers) on double machine learning for treatment evaluation with nonrandomly missing outcome variables has been published in the Journal of Business & Economic Statistics doi.org/10.1080/07350015.202…
#EconTwitter #CausalTwitter #StatsTwitter



























