PaperlayerView exact revision on arXiv

Source-linked AI summary

A Meta-Transfer Objective for Learning to Disentangle Causal Mechanisms

Yoshua Bengio, Tristan Deleu, Nasim Rahaman, Rosemary Ke, Sébastien Lachapelle, Olexa Bilaniuk, Anirudh Goyal, Christopher Pal

arXiv:1901.10912v2cs.LGstat.ML

TL;DR

The paper addresses how learned knowledge can generalize and adapt quickly when distributions change. It meta-learns causal structures and representations by optimizing adaptation speed under sparse mechanism changes, recovering causal direction and supporting better encoder selection. The work demonstrates this approach experimentally while identifying synthetic settings and simple encoders as current boundaries.

  • Problem

    Current machine-learning methods often generalize poorly beyond the training distribution, while causal reasoning requires identifying structure beyond a single observed joint distribution.

  • Method

    The paper uses a regret-like meta-learning objective based on adaptation speed to learn modular causal structures and encoders in which distributional changes are small and sparse.

  • Results

    The approach recovers causal direction and shows that the correct encoder obtains a better expected-regret meta-learning objective, with σ(γ) converging from 0.5 to 1 within 500 episodes.

  • Takeaways & Limitations

    Distributional changes can serve as a training signal for selecting causal structure and disentangling causal variables when those changes are localized and sparse.

  • Takeaways & Limitations

    The evaluation is an initial exploration using synthetic data and a simplest-possible encoder with a single degree of freedom, leaving broader settings and richer representations for future work.

Abstract

from arXiv · show

We propose to meta-learn causal structures based on how fast a learner adapts to new distributions arising from sparse distributional changes, e.g. due to interventions, actions of agents and other sources of non-stationarities. We show that under this assumption, the correct causal structural choices lead to faster adaptation to modified distributions because the changes are concentrated in one or just a few mechanisms when the learned knowledge is modularized appropriately. This leads to sparse expected gradients and a lower effective number of degrees of freedom needing to be relearned while adapting to the change. It motivates using the speed of adaptation to a modified distribution as a meta-learning objective. We demonstrate how this can be used to determine the cause-effect relationship between two observed variables. The distributional changes do not need to correspond to standard interventions (clamping a variable), and the learner has no direct knowledge of these interventions. We show that causal structures can be parameterized via continuous variables and learned end-to-end. We then explore how these ideas could be used to also learn an encoder that would map low-level observed variables to unobserved causal variables leading to faster adaptation out-of-distribution, learning a representation space where one can satisfy the assumptions of independent mechanisms and of small and sparse changes in these mechanisms due to actions and non-stationarities.

1. Introduction

The paper seeks faster transfer across related distributions by representing knowledge as independent, modular causal mechanisms whose changes remain sparse. It proposes using adaptation speed as a meta-learning signal to discover causal structure and representations, with initial experiments on causality and encoding.

  • Motivation: Transfer beyond the training distribution is difficult, motivating representations that support fast adaptation to new but related distributions.The relevant changes may arise from agent actions or other non-stationarities.
  • Core assumption: The approach assumes that distributional changes affect only one or a few mechanisms when knowledge is appropriately modularized.This sparsity assumption is especially relevant to interventions and agent actions that influence few underlying causal variables.
  • Meta-learning objective: Independent causal mechanisms imply that only a small number of parameters need relearning after a distributional change, producing fast adaptation from a well-trained model.The paper links this reduced relearning burden to sparse expected gradients and a lower effective number of degrees of freedom.
  • Meta-learning objective: The paper defines a regret-like meta-learning objective that measures adaptation speed to optimize how knowledge is represented, factorized, and structured.More distributional non-stationarities provide a stronger training signal, analogous to having more meta-examples.
  • Causal discovery: Distribution shifts that would normally be nuisances become training signals for selecting causal structure and disentangling causal variables.The method does not require direct knowledge that a shift came from a standard intervention.
  • Initial evaluation: Initial experiments use synthetic data to learn causal direction and show that the correct encoder achieves a better expected-regret meta-learning objective.The authors characterize this work as an initial exploration requiring evaluation across more settings and parameterizations.

2. Which is Cause and Which is Effect?

The paper identifies causal direction by comparing how quickly competing causal models adapt to sparse changes in a transfer distribution. The correct structure adapts faster because unchanged mechanisms remain reusable, reducing the parameters that must be relearned.

  • The Advantage of the Correct Causal Model: With only training-distribution data, A →B and B →A achieve indistinguishable test performance, so distributional changes provide the causal signal.Both models have O(N^2) parameters and equal maximum-likelihood performance on same-distribution tests.
  • The Advantage of the Correct Causal Model: The method compares the learning curves of competing causal hypotheses during stochastic-gradient adaptation to a related transfer distribution.The transfer distribution may arise from an unknown change, and the learner need not know which intervention occurred.
  • Experiments on Adaptation to the transfer distribution: The correct A →B model adapts faster than B →A, although both eventually reach the same transfer-distribution log-likelihood.The largest difference occurs within the first 10 to 20 transfer examples.
  • Parameter Counting Argument: Modules with correct causal parents and unchanged conditional distributions have zero expected transfer-regret gradients, concentrating adaptation on changed mechanisms.Thus, the effective number of parameters requiring adaptation is reduced under the correct graph.
  • Parameter Counting Argument: For a change only in P(A), the correct model needs to re-estimate N −1 parameters, whereas the incorrect model must update additional conditional mechanisms.This parameter-counting asymmetry explains the faster adaptation of the correct causal structure.
  • Smooth parameterization of the causal structure: A continuous structural parameter γ is optimized by regret, and experiments recover A →B across discrete, nonlinear continuous, and multivariate Gaussian settings.Starting from equal beliefs, σ(γ) converges to 1 within 500 episodes; related experiments report rapid graph recovery and 50–100 sufficient meta-examples.

3. Representation Learning

The paper learns an encoder that maps raw observations into a representation where causal modules support faster adaptation to distributional changes. In a rotation-based experiment, the meta-transfer objective recovers the underlying causal variables up to permutation.

  • Representation Learning: Raw observations may not preserve sparse causal structure, so the learner maps them into a hidden representation space via an encoder.The encoder is optimized jointly with structural meta-parameters using the meta-transfer objective.
  • Representation Learning: The simplified experiment assumes observations (X, Y) arise from causal variables (A, B) through an unknown rotation decoder.The learner must implicitly invert this decoder using an encoder rotation.
  • Representation Learning: Causal modules operate on encoded variables (U, V), which recover (A, B) or (B, A) when the encoder is valid.The two solutions differ by permutation, with possible sign symmetry.
  • Representation Learning: When the encoder is invalid, the hidden variables become superpositions of the underlying variables, removing the adaptation advantage of the correct causal direction.The relevant mismatch is represented by θ = θE + θD.
  • Representation Learning: The proposed objective learns the encoder jointly with causal structure and disentangles the ground-truth variables up to permutation in the simple experiment.Figure 5 illustrates the encoder-training behavior.

4. Related Work

The paper situates its approach among work on disentanglement, causal structure learning, domain adaptation, and transfer learning. It emphasizes learning explanatory variables and causal mechanisms rather than only fitting observed-variable distributions.

  • Representation Learning: Figure 5 is referenced as evidence that the proposed training procedure can learn encoder parameters in the representation-learning setting.The related discussion connects this objective to learning useful causal representations.
  • Disentanglement: The paper relates its objective to disentangling underlying explanatory variables and to representations involving only a few related high-level concepts.It contrasts this with extreme disentanglement assumptions requiring marginal independence.
  • Causal Structure Learning: Earlier Bayesian-network approaches searched discrete structures using methods such as simulated annealing, MDL, or BIC-based scoring.These methods are presented as prior work on structure learning.
  • Domain Adaptation and Transfer: Related domain-adaptation methods seek invariant conditional distributions, counterfactual predictions, treatment effects, or feature subsets across domains.The paper places its method alongside several such approaches.
  • Causal Mechanisms: Other work recovers independent causal mechanisms through competition or uses meta-learning to draw causal inferences from observational data.These approaches provide neighboring links between causal discovery and meta-learning.

5. Conclusion and Future Work

The paper shows that adaptation speed can help select causal structures when distributional changes are sparse and localized in causal mechanisms. It presents this as an initial step, with broader evaluation and richer representation learning left for future work.

  • The demonstrated approach establishes that sparse changes in observed-data distributions can be used to identify causal structure in simple settings.
  • Adaptation rates can select or optimize causal structure and disentangle causal variables when distributional changes are sparse and localized.The paper supports this with theoretical results and experimental validation in simple bivariate settings.
  • Future work includes richer causal graphs, alternative parametrizations and optimization procedures, and more capable encoders for disentangled causal variables.The paper reports experimenting only with a simplest encoder having a single degree of freedom.
  • Scaling these ideas could improve how learning agents handle non-stationarities, with potential benefits for sample complexity and robustness.

Appendix A. Results on Non-Identifiability of Causal Structure

The appendix shows that the two bivariate causal models have identical maximum-likelihood distributions and joint likelihood on the training data. Training and same-distribution testing therefore cannot distinguish their causal direction at convergence.

  • Maximum likelihood estimation gives both A →B and B →A models the same estimated joint distribution over A and B.
  • The two causal models use separate marginal and conditional probability modules to represent their respective factorizations.The parameters are defined as θ_i, θ_j|i for A →B and η_j, η_i|j for B →A.
  • The maximum-likelihood estimators use empirical marginal counts and joint counts: ˆθ_i = n_i/n, ˆθ_j|i = n_ij/n_i, ˆη_j = n_j/n, and ˆη_i|j = n_ij/n_j.
  • Both fitted models assign the same joint probability, ˆP_A→B(A,B) = ˆP_B→A(A,B) = n_ij/n.
  • SGD training may fit A →B faster, but both models reach the same training- and test-distribution log-likelihoods at convergence.This demonstrates indistinguishability from data sampled from the same distribution, even on test data.

Appendix B. Proof of the Zero-Gradient Proposition

The zero-gradient proposition analyzes a correctly structured, perfectly pretrained modular model whose transfer distribution changes only a few mechanisms. The appendix also gives the meta-transfer procedure that pretrains, adapts, and updates structural meta-parameters.

  • The proposition assumes the correct causal structure, perfect training on P1, and changes between P1 and P2 limited to selected conditional mechanisms.The changed mechanisms are indexed by i ∈ C.
  • The joint log-likelihood decomposes into a sum of module-specific log probabilities, making each module’s parameter contribution explicit.
  • For unchanged mechanisms, the expected transfer gradient contribution vanishes under the proposition’s assumptions.The proof uses parameter modularity and equality of the corresponding transfer and training conditionals.
  • Algorithm 1 pretrains a learner on training data, samples transfer distributions, adapts on transfer minibatches, and updates meta-parameters by SGD.The learner parameters may optionally be reset to their pretrained values.

Appendix D. Proof of the Structural Parameter Gradient Proposition

The appendix derives how transfer-data likelihood differences determine the structural posterior and the gradient-descent limit for a continuous causal-choice parameter. The optimization converges to the model with the larger expected transfer log-likelihood.

  • The transfer-data posterior for A →B is obtained by Bayes’ rule from the two hypotheses’ transfer likelihoods and their prior structural probabilities.
  • The log-likelihood difference Δ = log L_A→B − log L_B→A summarizes the transfer-data evidence between the two causal hypotheses.
  • The mixture likelihood is M = σ(γ)L_A→B + (1−σ(γ))L_B→A, and transfer regret is R = −log M.
  • SGD on the structural parameter converges to σ(γ) = 1 when E_D2[log L_A→B] exceeds E_D2[log L_B→A], and to 0 otherwise.
  • The proof rules out interior solutions and leaves only endpoint choices, with the minimizing endpoint determined by the larger expected transfer log-likelihood.

Appendix F. More Than Two Causal Hypotheses

The appendix extends causal-structure meta-learning beyond two hypotheses by independently parameterizing incoming edges and decomposing the regret across nodes. Although enumerating all graphs is intractable, node-wise stochastic gradient estimates make optimization efficient.

  • The method considers O(2^m^2) possible causal graphs over m variables, with each potential directed edge represented as a binary decision.The ground-truth experiments use directed graphs, but the learning procedure does not directly assume acyclicity.
  • Each edge probability p_ij is independently parameterized, with B_ij indicating whether V_j is a direct parent of V_i.The parent set of V_i contains variables V_j for which B_ij = 1, excluding self-edges.
  • Each node-specific mechanism f_i uses only parent variables selected by the binary mask B_i and combines them with an independent noise source N_i.The masked neural generators resemble input-masked or dropout-style networks.
  • The graph-level regret is formed from node-wise likelihood terms whose product explains the full adaptation sequence across candidate structures.Each node contribution evaluates how well its incoming edges model the observed examples during adaptation.
  • The expectation over all graph structures is intractable, but stochastic gradient estimates can be computed separately for each node using samples of incoming-edge vectors.The estimator uses multiple graph samples and weights individual binomial gradients by each sampled graph’s relative regret.
  • The resulting loss and gradient decomposition enables the general multi-variable causal-structure meta-transfer procedure in Algorithm 1.Each independent term depends only on a node’s incoming-edge vector rather than the full graph.

Appendix G. Results on Learning which is Cause and which is Effect

The appendix evaluates cause-effect discovery on discrete variables using modular causal models, transfer distributions, and meta-learning of a structural parameter. Across the setup, the learned belief converges to the correct direction A →B.

  • The experiments cover discrete, multimodal continuous, and multivariate Gaussian variables while fixing the ground-truth direction to A →B.The appendix reports setups and supplementary results for all three domains.
  • For discrete bivariate variables, the competing A →B and B →A models each use separately parameterized marginal and conditional modules.The two models have the same parameter count, N^2 + N.
  • The structural parameter γ is updated across multiple sampled interventions using the adaptation likelihoods of the two causal hypotheses.The update can use gradient descent or a first-order optimizer such as RMSprop.
  • The learned belief P(A →B) = σ(γ) converges to 1, recovering the correct causal direction.The result is reported as the evolution of σ(γ) over meta-training steps or sampled interventions.
  • The causal mechanisms are implemented with one MLP per variable, whose inputs are masked according to the sampled graph structure.Each edge indicator B_ij determines whether variable j can enter the MLP for variable i.

G.3 Continuous Multimodal Variables

The continuous multimodal experiment generates training and transfer distributions from a spline-based SCM, then compares how quickly the two causal directions adapt. Meta-learning increases belief in A →B as transfer regret favors that structure.

  • The SCM generates B = f(A) + N_B, where f is a random spline and N_B is unit-normal noise.The spline uses K = 8 knots with R_A = R_B = 8 in the reported setup.
  • Training samples come from P_μ=0(A, B), while transfer samples use shifted distributions P_μ=-4(A, B) and P_μ=+4(A, B).Figure G.1 distinguishes training samples from the two transfer distributions by color.
  • The A →B and B →A models use Mixture Density Networks for conditionals and Gaussian Mixture Models for marginals.The conditional networks have two layers, 32 hidden units, and 10 components; each marginal GMM also has 10 components.
  • The structural objective mixes exponentiated regrets for A →B and B →A using σ(γ), favoring A →B when its expected regret is lower.The reported optimization minimizes structural regret for 200 updates and expects R_A→B < R_B→A.
  • Figure G.2 tracks σ(γ) over training iterations and indicates that it should increase when A →B is the true causal graph.The figure provides the optimization trajectory for the structural meta-parameter.

G.4 Linear Gaussian Model

The linear Gaussian experiment tests causal-direction learning with 100-dimensional vector variables and analytically specified, symmetrically parameterized models. After transfer adaptation episodes, the causal belief converges to the correct direction.

  • The experiment uses vector variables A and B in R^d with d = 100 and a linear-Gaussian ground-truth causal model.The model includes Gaussian covariances parameterized through Cholesky decompositions.
  • The competing A →B and B →A models are parameterized symmetrically and initialized to exact values implied by the ground-truth distribution.The reverse-direction parameters are computed analytically, making both models equivalent before transfer adaptation.
  • Each meta-learning episode samples a transfer distribution with μ_A ∼ N(0, I) and trains both models for 10 iterations.Accumulated log-likelihoods during adaptation define L_A→B and L_B→A for the meta-objective.
  • After 200 episodes, σ(γ) converges to 1, indicating successful causal-direction learning on this task.The result is shown in Figure G.3 for the linear Gaussian experiments.

Appendix H. Results on Learning the Correct Encoder

The appendix learns an encoder for hidden causal variables by meta-training directional models and their adaptation regret. The system converges to the encoder and structural parameter settings corresponding to the correct causal direction U → V.

  • Encoder setup: The experiment maps causal variables (A, B) to observations (X, Y) through an unknown rotation decoder and then to hidden states (U, V) through a learned rotation encoder.
  • Encoder setup: Ground-truth parameters μA, β1, and β0 use Gaussian sampling, while ΣA and ΣB use inverse-Wishart sampling.
  • Meta-training procedure: Each meta-training iteration trains U → V and V → U models for 20 iterations on Pμ=0(U, V), then computes their regrets after 5 adaptation iterations.
  • Meta-training procedure: The gradients of the regret with respect to γ and θE update the meta-parameters, and this process repeats for 1000 meta-iterations.
  • Results: The encoder converges to θE = 0, implying the correct causal direction is U → V, while γ increases during meta-training.
Loading 1901.10912v2…