PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Causal Representation Learning

Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, Yoshua Bengio

arXiv:2102.11107v1cs.LGcs.AI

TL;DR

Machine learning often relies on statistical associations under fixed or i.i.d. conditions, limiting transfer, robustness, and intervention-based reasoning. This paper reviews causal inference, causal discovery, and causal representations, relating them to machine-learning problems and identifying causal mechanisms, interventions, and representation learning as central connections.

  • Problem

    Machine learning is limited in transfer and out-of-distribution generalization, while causal research commonly assumes that causal variables are already given.

  • Method

    The paper reviews causal inference and discovery, develops the Independent Causal Mechanisms perspective, and relates causal representations to machine-learning problems.

  • Results

    The paper identifies causal representations, interventions, and reusable independent mechanisms as central research directions connecting causality with machine learning.

  • Takeaways & Limitations

    Causal modeling offers a framework for reasoning about interventions and distribution changes while connecting causal structure to robustness, generalization, and learning practice.

Abstract

from arXiv · show

The two fields of machine learning and graphical causality arose and developed separately. However, there is now cross-pollination and increasing interest in both fields to benefit from the advances of the other. In the present paper, we review fundamental concepts of causal inference and relate them to crucial open problems of machine learning, including transfer and generalization, thereby assaying how causality can contribute to modern machine learning research. This also applies in the opposite direction: we note that most work in causality starts from the premise that the causal variables are given. A central problem for AI and causality is, thus, causal representation learning, the discovery of high-level causal variables from low-level observations. Finally, we delineate some implications of causality for machine learning and propose key research areas at the intersection of both communities.

I. INTRODUCTION

The paper argues that causality can address machine learning’s limitations in transfer and out-of-distribution generalization by modeling interventions, distribution changes, and reusable structure. It reviews causal concepts, causal discovery, causal representations, and implications for practical machine learning.

  • Motivation: Machine learning remains limited at transfer and problem-to-problem generalization because it commonly treats interventions, domain shifts, and temporal structure as nuisances.Most current successes rely on large-scale pattern recognition using suitably collected i.i.d. data.
  • A Causality Perspective: Causation requires interventions beyond conditional probabilities, distinguishing observed associations from the outcomes of actively changing variables.The umbrella example illustrates why observing a correlation does not determine an intervention’s effect.
  • Contributions: The paper presents causal representations as models of structural knowledge that can support reasoning about interventions and distribution changes.It contrasts these with statistical learning, whose descriptions hold primarily under fixed experimental conditions.
  • Contributions: It reviews classical and deep-neural approaches for learning causal relations from descriptors, emphasizing principles that enable causal discovery.The paper also expands the Independent Causal Mechanisms principle through the Sparse Mechanism Shift hypothesis.
  • Contributions: The paper examines causal perspectives on robustness, generalization, semi-supervised learning, self-supervision, data augmentation, and pre-training.It also discusses scientific applications and combining causal and machine-learning strengths toward more versatile AI.
  • Scope: The present paper expands an earlier work, resulting in partial text overlap.This is stated as a paper-level scope note.

II. LEVELS OF CAUSAL MODELING

The paper places statistical, causal, and physical models at different levels of description. Statistical models are learnable from observations but mainly capture associations, whereas causal models aim to answer intervention questions through data-driven structural modeling.

  • Physical models: Physical models use differential equations to represent mechanisms, predict future behavior, explain systems, and expose causal structure.Under suitable conditions, the system’s causal structure can be read from which variables determine others’ future values.
  • Statistical models: Statistical models learn associations from observational data and can predict under fixed experimental conditions, but they do not generally predict intervention effects.Their strength is learnability from raw perceptual or hand-engineered observations.
  • Scope condition: Mechanistic equations are not automatically causally meaningful when they describe time evolution phenomenologically rather than the underlying mechanisms.Unobserved confounding or causal-distorting course-graining can break that interpretation.
  • Causal models: Causal models lie between physical and statistical descriptions, retaining power for intervention and counterfactual questions while learning structure from data with weak assumptions.They abstract away from full physical realism while preserving selected causal capabilities.
  • Interventions and distribution shift: Accurate statistical prediction can fail after interventions because changing experimental conditions changes the data distribution.The stork and birth-rate example separates predictive association from causal effect.

C. Answering Counterfactual Questions

The paper distinguishes observational, interventional, and counterfactual reasoning, emphasizing that counterfactuals ask about alternative actions in hindsight. It also relates these questions to data modalities and the difficulty of learning causal structure from raw or passively observed data.

  • Counterfactual questions: Counterfactual questions ask why an outcome occurred and what would have happened under a different action, making them harder than interventional questions.The paper identifies counterfactual reasoning as especially relevant to reinforcement learning.
  • A continuum of questions: Conditional-interventional quantities form a continuum between interventions and counterfactuals, depending on how much information about background noise is conditioned upon.Reinforcement-learning Q functions are described as closer to interventions because they condition on state and action rather than all noise terms.
  • Data modalities: Observational i.i.d. data, known interventional data, and data with unknown shifts represent different positions between passive sampling and explicit interventions.The intermediate case is passively observed despite containing shifts or interventions unknown to the learner.
  • Structured versus raw data: Raw perceptual data does not directly expose causal variables, while causal learning often requires multiple environments or interventions.Hand-engineered variables may partially correspond to high-level causal variables.
  • Limits of i.i.d. learning: Machine-learning methods perform well under i.i.d. assumptions but often degrade under context changes, corruptions, and other distribution shifts.The paper notes that universal consistency guarantees apply to i.i.d. data, not generally to these violations.

B. The Reichenbach Principle: From Statistics to Causality

The Reichenbach principle connects statistical dependence with possible common causes, but observational data alone may not identify the causal direction. Structural causal models add graph-based assignments, independent noises, and intervention semantics to represent causal mechanisms.

  • B. The Reichenbach Principle: Reichenbach’s common-cause principle states that dependent observables share a variable that explains their dependence when conditioned upon.The common cause may coincide with either observable.
  • B. The Reichenbach Principle: Without additional assumptions, observational data cannot distinguish direct causation in either direction from a common-cause explanation.These alternatives generate the same class of observational distributions over two variables.
  • Structural causal models: An SCM assigns each observable to a deterministic function of its graph parents and an unexplained random variable, with directed edges representing direct causation.The noise term allows the assignment to represent a general conditional distribution given the parents.
  • Structural causal models: Recursive application of structural assignments produces an observational joint distribution satisfying the causal Markov condition.Conditioned on its parents, each variable is independent of its non-descendants.
  • Causal discovery: Independent noises leave conditional-dependence footprints that can reveal aspects of graph structure, although they may not uniquely identify it.With only two observables, the footprint is insufficient to pin down a unique causal structure.
  • Causal factorization: The causal factorization decomposes the joint distribution into conditionals corresponding to structural assignments, treating them as distinct causal mechanisms.This factorization also offers computational advantages during inference.
  • Latent variables and confounders: Unobserved variables can confound observed variables, making causal inference particularly challenging.Confounding can create or alter statistical relations without being directly observed.
  • Interventions: SCM interventions modify assignments, such as setting a variable constant or changing its function or noise term.The intervention set is relevant when learning a causal model because the SCM depends on the interventions considered.

D. Difference Between Statistical Models, Causal Graphical Models, and SCMs

Statistical models describe a single observational distribution, whereas causal models represent interventions and their resulting distributions. Structural causal models additionally specify causal variables and structural equations, enabling counterfactual reasoning.

  • A statistical model specifies a probability distribution, but a causal model represents distributions for possible interventions.Interventions disconnect a variable from its parents, fix its value, and enable ancestral sampling on its children.
  • A graphical model becomes causal when its graph edges represent causal relations rather than merely statistical dependencies.Noncausal graphs can belong to the same Markov equivalence class while implying identical conditional independencies.
  • A structural causal model contains causal variables and structural equations with noise distributions or causal conditionals.SCMs support both interventional distributions and counterfactuals, the latter requiring fixed noise-variable values.
  • Causal learning exploits causal conditionals and the causal factorization of the joint distribution rather than only a response-prediction function.This provides a richer set of assumptions than statistical learning based on the joint distribution and function classes.

IV. INDEPENDENT CAUSAL MECHANISMS

The Independent Causal Mechanisms principle models a causal process as autonomous modules whose mechanisms do not inform or influence one another. Its Sparse Mechanism Shift consequence predicts that distribution changes usually affect only a small subset of causal factors, unlike entangled noncausal factorizations.

  • Independent Causal Mechanisms: Changing viewpoint is expected to leave lighting, object position, and structure unaffected, supporting inference of 3D information from motion.The example treats vantage point and the remaining variables as independent components of the generative process.
  • Independent Causal Mechanisms: Causal models are expected to generalize from observational to certain interventional distributions because they explicitly model interventions.The SMS hypothesis further assumes that much causal structure remains invariant across related environments or contexts.
  • Independent Causal Mechanisms: The ICM principle states that causal generative processes comprise autonomous modules whose mechanisms do not inform or influence one another.In probabilistic models, each conditional distribution given its causes is treated as a mechanism independent of the others.
  • Independent Causal Mechanisms: Mechanism independence entails separate intervenability, modularity, autonomy, and invariance across changes to other mechanisms.For two variables, it reduces to independence between the cause distribution and the mechanism producing the effect.
  • Independent Causal Mechanisms: The Sparse Mechanism Shift hypothesis says small distribution changes tend to affect a sparse or local subset of causal-factorization terms.This follows from the idea that changing one mechanism leaves the others unchanged.
  • Independent Causal Mechanisms: A noncausal factorization can change many or all terms simultaneously when one physical mechanism changes, making it entangled.Causal structure is presented as more fundamental because it generates the statistical dependencies observed after unexplained variables are randomized.

V. CAUSAL DISCOVERY AND MACHINE LEARNING

Causal discovery from observational data can recover graph properties under assumptions, but finite-sample conditional-independence testing is difficult and two-variable cases are uninformative. Function-class restrictions and distribution shifts provide additional routes for identifying causal direction and learning models that generalize across environments.

  • Causal Discovery and Machine Learning: Observational causal discovery can recover aspects of a graph under assumptions such as faithfulness, but generally only up to a Markov equivalence class.Graphs in the same class share an undirected skeleton and immoralities.
  • Causal Discovery and Machine Learning: Finite-sample conditional-independence testing is difficult, especially with continuous, high-dimensional conditioning sets, and two variables yield no nontrivial Markov implications.These limitations motivate additional assumptions beyond the causal Markov condition.
  • Causal Discovery and Machine Learning: Restricting function classes can make causal models learnable from finite data and break the symmetry between cause and effect.For additive-noise models, the reverse direction generally cannot fit the same model, subject to genericity assumptions and exceptions such as linear Gaussian systems.
  • Causal Discovery and Machine Learning: Smooth dependence on an unobserved selector variable can reduce the effective function-class complexity relative to arbitrary dependence on its support.A local Taylor expansion motivates this smoothness restriction when the selector is relatively concentrated.
  • Causal Discovery and Machine Learning: Distribution shifts across environments can help identify causal structure because causal mechanisms are expected to remain invariant while only some mechanisms change.Neural approaches have connected this invariance to faster adaptation to interventions and continuous optimization of causal graphs.

VI. LEARNING CAUSAL VARIABLES

Causal representation learning seeks to recover high-level causal variables and their relations from low-level, often entangled observations. The paper connects this goal to disentangled representations, modular neural architectures, and robustness or transfer under distribution shifts, while noting fundamental identifiability and scope challenges.

  • Learning Causal Variables: Causal representation learning discovers causal variables from high-dimensional observations rather than assuming those variables are given.The observations may be images or other sensor measurements that provide entangled views of an unknown causal system.
  • Learning Causal Variables: The learned representation may expose only the causal structure needed for downstream tasks when full recovery is unreasonable.Separating object properties from pose and appearance can support robustness to shifts in the latter variables.
  • Learning Causal Variables: Under the SMS hypothesis, an intervention can produce a localized causal change even when its pixel-space effect is distributed across the observation.The figure illustrates a finger-position intervention followed by the object falling.
  • Learning Causal Variables: Combining an SCM with representation learning uses an encoder, structural assignments, and a decoder to map observations through high-level causal variables.The proposed architecture can embed causal structure inside neural models and support modular fine-tuning and repurposing.
  • Learning Causal Variables: Disentangled representations of independent factors correspond to the special case of a trivial causal graph, while ICM also permits dependent causal variables.Object-centric representations can serve as building blocks in a broader causal factorization.
  • Learning Causal Variables: Recovering exogenous noise variables is ill-defined in the i.i.d. setting because infinitely many equivalent solutions can yield the same observational distribution.Additional assumptions or biases may favor particular solutions.

VII. IMPLICATIONS FOR MACHINE LEARNING

The paper calls for learning beyond the usual i.i.d. assumption, targeting new distributions that involve mostly the same causal mechanisms while inferring abstract causal variables from low-level inputs.

  • Causal learning should handle application data from a possibly different distribution involving mostly the same causal mechanisms.
  • A central challenge is inferring abstract causal variables from available low-level input features.

A. Semi-Supervised Learning (SSL)

The paper uses the Independent Causal Mechanisms principle to distinguish semi-supervised learning by causal direction: unlabeled inputs are uninformative in the causal direction but may help in the anticausal direction.

  • In the causal direction, P(X) should contain no information about P(Y|X), making self-supervised learning futile when it relies on additional unlabeled P(X).
  • In the anticausal direction, semi-supervised learning may be possible because the effect distribution can inform the conditional distribution of causes given effects.
  • The study empirically corroborated these predictions and motivated follow-up work connecting learning-problem structure with physical properties of data-generating processes.
  • This view aligns with semi-supervised assumptions that constrain how the labeling function behaves within input clusters or across low-density regions.

B. Adversarial Vulnerability

The paper connects causal structure with adversarial vulnerability and robust out-of-distribution prediction, while emphasizing that robustness optimization depends on specifying the causal model and intervention set.

  • Adversarial examples violate the i.i.d. assumption because modified test inputs are not drawn from the training distribution.
  • The observed robustness of current classifiers differs from human robustness, which motivates distinguishing pixel-based adversarial perturbations from broader robustness measures.
  • If a predictor approximates a transferable causal mechanism, adversarial examples may be harder to find.
  • Adversarial attacks may still exploit errors in the parameterized approximation of a structural equation.
  • Causal features can make credit scoring more robust to strategic behavior than features people can readily manipulate.
  • Out-of-distribution robustness can be formalized by minimizing empirical risk over distributions induced by a causal model and its interventions.
  • The robustness objective is challenging because the observed environments may not cover all possible environments, causing arbitrarily large worst-case estimation error.

D. Pre-training, Data Augmentation, and Self-Supervision

The paper interprets pre-training, data augmentation, and self-supervision as ways to approximate robustness objectives by enriching data, simulating interventions, and learning from unlabeled inputs.

  • Enriching training distributions helps only when the added data is sufficiently diverse to contain information about other distributions.
  • Data augmentation increases data diversity through artificially generated interventions such as image rotations, translations, and flips.
  • Self-supervised pre-training learns about P(X) from unlabeled data and can support good results with few class labels or efficient downstream transfer.
  • The intervention set may represent natural interventions rather than adversarial attacks.
  • Robust predictive learning can combine large-scale training, data augmentation, self-supervision, and robust fine-tuning across multiple potentially simulated environments.

E. Reinforcement Learning

The paper connects reinforcement learning to causality through world models, interventions, invariances, counterfactuals, and offline learning, while identifying transfer, robustness, and data-efficiency challenges.

  • World Models: Model-based reinforcement learning relates to causality by modeling how actions affect the current world state, especially through generative world models.These models can provide imagined spaces for training agents and may capture causal relations underlying environments.
  • Generalization, Robustness, and Fast Transfer: Causal graph invariances and independent causal mechanisms are proposed as routes toward extrapolating experience across environments and tasks.Interventions may help agents discover causal factors, while changing distributions need not require relearning every mechanism.
  • Counterfactuals: Counterfactual reasoning can improve reinforcement-learning data efficiency and performance, support communication about past experiences, and guide testable intervention hypotheses.The paper argues that counterfactuals should be treated as a critical component of acting in imagined spaces.
  • Offline RL: Offline reinforcement learning addresses the expense of collecting diverse, high-fidelity trajectories by learning policies from fixed datasets without new experiments.The supplied passage introduces offline RL as a response to data-collection bottlenecks.
  • Applications: Causal models are presented as relevant to scientific and medical applications where distribution changes, confounding, or uncontrolled experimental conditions limit purely statistical approaches.Examples include personalized medicine, causal mediation analysis during Covid-19, and exoplanet identification under instrumental confounding.
  • Motivation: The broader motivation is that current machine learning remains narrow and struggles with transfer across substantially different problems, whereas causal concepts may support more versatile systems.The paper frames combining deep learning with causality as a possible step toward versatile AI.

VIII. CONCLUSION

The conclusion positions causal modeling and representation learning as a bridge between machine learning and causality, outlining research priorities for scalable causal relations, causal variables, inductive biases, and world models.

  • Conclusion: The paper synthesizes causal and statistical models, independent mechanisms, invariance, causal discovery, causal representation learning, and machine-learning problems through a unified causal perspective.It covers causal relations from observational and interventional data when causal variables are observed, then turns to causal representation learning and machine-learning applications.
  • Learning Non-Linear Causal Relations at Scale: Learning nonlinear causal relations at scale requires conditions for learning, scalable training frameworks, and evidence for improved generalization, repurposing, and transfer.The paper notes that interventions may be observed across multiple environments and that meta- and multi-task learning align with causal-modeling goals.
  • Learning Causal Variables: Learning causal variables remains open because current disentangled representations are distributed vectors with arbitrary dimension ordering and fixed representation size.The paper links recovering causal variables to understanding robustness under interventions in predictive tasks.
  • Understanding the Biases of Existing Deep Learning Approaches: Strong generalization over specified interventions could probe the inductive biases of existing deep-learning methods, training schemes, and datasets.The paper specifically calls for understanding how pre-training choices affect downstream transfer and robustness in a causal sense.
  • Learning Causally Correct Models of the World and the Agent: Robust and versatile model-based reinforcement learning requires causal descriptions of both the agent and its environment, including abstract causal variables derived from pixels.The conclusion highlights recovering causal graphs when abstract state representations are unavailable.
Loading 2102.11107v1…