Can a model learn causes?

Connect structural causal models, identification, invariant prediction and counterfactual explanations.

AI & computationTopic · Selected readingSome background helpful

Background and goal

Basic probability, directed graphs and simple modelling are required. Start with Probability and Causality if these are unfamiliar; Python helps with optional simulations.

A causal analysis that separates predictive accuracy from intervention validity.

Useful preparation

Start here if the background is new. Equivalent experience is enough.

Causality
Use causal diagrams, interventions and assumption checks to assess a model’s causal claims.
PhilosophySome background
Probability & statistics
Use conditional probability to separate an observed association from an intervention question.
MathematicsBeginner

Your pathway

5 steps · 16 hours

01
Specify the intervention question
2h · week 1
Causal Inference in Statistics: A Primer
Pearl, Glymour · Jewell · book · 2016
Interventions, graphs and confounding
Check reading access
Open the reading
Causal Models
Stanford Encyclopedia of Philosophy editors · reference
Interventions, graphs and confounding
Check reading access
Open the reading
Reading notes and task
Causal Inference in Statistics: A Primer. Use one small graph to distinguish observational association, confounding and intervention.
Causal Models. Draw a small graph and distinguish a correlation from an intervention question.
Try thisUse the primer and causal-model overview to turn a prediction task into an intervention question. Draw variables and a possible confounder.
Your outputA question pair and an initial causal graph.
02
Compare identification strategies
4h · week 1
Causality
Judea Pearl · book · 2000
Graph-based identification and structural assumptions
Check reading access
Open the reading
Elements of Causal Inference
Peters, Janzing · Schölkopf · book · 2017
Graph-based identification and structural assumptions
Check reading access
Open the reading
Reading notes and task
Causality. Use a selected graph-based argument to identify the assumptions behind an intervention claim.
Elements of Causal Inference. Inspect one identification strategy and a case where its assumptions fail.
Try thisRead selected Pearl and Peters, Janzing and Schölkopf arguments. Explain what assumptions let an observational model answer the intervention question.
Your outputAn assumption table with two proposed strategies.
03
Simulate confounding
4h · week 2
Causal Inference: What If
Hernán · Robins · book · 2020
A confounded effect-estimation example
Check reading access
Open the reading
Causal Inference in Statistics: A Primer
Pearl, Glymour · Jewell · book · 2016
Interventions, graphs and confounding
Check reading access
Open the reading
Reading notes and task
Causal Inference: What If. Choose a worked adjustment example and list the identification assumptions.
Causal Inference in Statistics: A Primer. Use one small graph to distinguish observational association, confounding and intervention.
Try thisUse a small spreadsheet or program to generate a confounded observational example. Compare the naive association with the effect under a simulated intervention and an appropriate adjustment.
Your outputA reproducible simulation with three estimates.
04
Test invariance carefully
4h · week 3
Causal inference by using invariant prediction: identification and confidence intervals
Peters, Bühlmann · Meinshausen · paper · 2016
Cross-environment assumptions and invariance claims
Check reading access
Open the reading
Invariant Risk Minimization
Arjovsky et al. · paper · 2019
Cross-environment assumptions and invariance claims
Check reading access
Open the reading
Reading notes and task
Causal inference by using invariant prediction: identification and confidence intervals. Record the environmental assumptions needed for the invariance claim to identify a cause.
Invariant Risk Minimization. Distinguish the optimisation proposal from a guarantee of causal identification.
Try thisCompare invariant prediction with invariant risk minimisation. Describe what changes across environments, what remains stable and why the available environments may not identify causality.
Your outputAn environment comparison with a failed-identification case.
05
Write a causal modelling report
2h · week 4
Reading notes and task
Try thisPresent the intervention question, graph, simulated data, estimates and assumptions. Explain which claims are identified in the example and which would need real evidence.
Your outputA reproducible causal report with an identification limit.

Compare perspectives

Editorial perspectives based on selected works, rather than author-endorsed reading lists.

Reference synthesis

Editorial synthesis

An editorial comparison of the route’s selected accounts and evidence.

A curated synthesis cannot be neutral or exhaustive. Follow the primary sources when an interpretation matters.

Judea Pearl

Reconstructed perspective

An editorial reconstruction using the selected work of Judea Pearl.

This is not an author-endorsed syllabus. The reconstruction highlights selected works and may omit other commitments.

Bernhard Schölkopf

Reconstructed perspective

An editorial reconstruction using the selected work of Bernhard Schölkopf.

This is not an author-endorsed syllabus. The reconstruction highlights selected works and may omit other commitments.

Jonas Peters

Reconstructed perspective

An editorial reconstruction using the selected work of Jonas Peters.

This is not an author-endorsed syllabus. The reconstruction highlights selected works and may omit other commitments.

Key ideas

View connections

Structural causal models Identification Invariance Causal discovery

All 20 resources

Background, different viewpoints, and further reading.

Causality
Judea Pearlbook · 2000core
Use a selected graph-based argument to identify the assumptions behind an intervention claim.
Graph-based identification and structural assumptions
The Book of Why
Judea Pearl · Dana Mackenziebook · 2018core
Compare this account’s mechanism with the preceding reading; note where their predictions differ.
Selected chapters and argument
Elements of Causal Inference
Peters, Janzing · Schölkopfbook · 2017core
Inspect one identification strategy and a case where its assumptions fail.
Graph-based identification and structural assumptions
Causal Inference in Statistics: A Primer
Pearl, Glymour · Jewellbook · 2016core
Use one small graph to distinguish observational association, confounding and intervention.
Interventions, graphs and confounding
Causal Inference: What If
Hernán · Robinsbook · 2020commentary
Choose a worked adjustment example and list the identification assumptions.
A confounded effect-estimation example
Toward Causal Representation Learning
Schölkopf et al.paper · 2021commentary
Identify the proposed model class and what evidence could distinguish its causal account.
Causal representation and additive-noise model assumptions
Invariant Risk Minimization
Arjovsky et al.paper · 2019commentary
Distinguish the optimisation proposal from a guarantee of causal identification.
Cross-environment assumptions and invariance claims
Causal inference by using invariant prediction: identification and confidence intervals
Peters, Bühlmann · Meinshausenpaper · 2016commentary
Record the environmental assumptions needed for the invariance claim to identify a cause.
Cross-environment assumptions and invariance claims
The Seven Tools of Causal Inference, with Reflections on Machine Learning
Judea Pearlpaper · 2019criticism
Reconstruct the main argument and identify the assumptions linking it to the route’s question.
Full paper; methods and limitations
Nonlinear causal discovery with additive noise models
Hoyer et al.paper · 2008criticism
Inspect the additive-noise and independence assumptions and try an alternative model.
Causal representation and additive-noise model assumptions
Counterfactual Explanations without Opening the Black Box
Wachter, Mittelstadt · Russellpaper · 2017background
Extract one claim and distinguish the evidence supporting it from the author’s interpretation.
Full paper; methods and limitations
Causal Models
Stanford Encyclopedia of Philosophy editorsreferencebackground
Draw a small graph and distinguish a correlation from an intervention question.
Interventions, graphs and confounding
Causation and Manipulability
Stanford Encyclopedia of Philosophy editorsreferencebackground
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples
Probabilistic Causation
Stanford Encyclopedia of Philosophy editorsreferencebackground
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples
Models in Science
Stanford Encyclopedia of Philosophy editorsreferencebackground
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples
Bayes’ Theorem
Stanford Encyclopedia of Philosophy editorsreferencebackground
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples
The Metaphysics of Causation
Stanford Encyclopedia of Philosophy editorsreferenceextension
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples
Philosophy of Statistics
Stanford Encyclopedia of Philosophy editorsreferenceextension
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples
Understanding Deep Learning
Simon J. D. Princebook · 2023extension
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples
Artificial Intelligence
Stanford Encyclopedia of Philosophy editorsreferenceextension
Use the entry’s distinctions and bibliography to test the route’s central argument, rather than treating a summary as a substitute for the primary source.
Named section and worked examples

Continue learning

Deep learning
Contrast held-out predictive performance with the validity of an intervention claim.
AI & computationSome background

Finding the next step…

Can a model learn causes?

Useful preparation

Your pathway

Specify the intervention question

Compare identification strategies

Simulate confounding

Test invariance carefully

Write a causal modelling report

Reference synthesis

Key ideas

Continue learning

Can a model learn causes?

Useful preparation

Your pathway

Specify the intervention question

Compare identification strategies

Simulate confounding

Test invariance carefully

Write a causal modelling report

Reference synthesis

Key ideas

Continue learning