[Paper Review] Learning Causal Models Online

Khurram Javed, Martha White+1 more|arXiv (Cornell University)|Jun 12, 2020

Domain Adaptation and Few-Shot Learning25 references14 citations

TL;DR

This paper proposes an online algorithm, PwB*, that detects and removes spurious features in neural networks by exploiting their temporal instability—features whose correlation with the target shifts over time. By iteratively perturbing network parameters and minimizing both loss and feature variance, the method converges to models with strong generalization, outperforming baseline methods like Oracle IRM in online settings while discovering non-spurious features from raw data.

ABSTRACT

Predictive models -- learned from observational data not covering the complete data distribution -- can rely on spurious correlations in the data for making predictions. These correlations make the models brittle and hinder generalization. One solution for achieving strong generalization is to incorporate causal structures in the models; such structures constrain learning by ignoring correlations that contradict them. However, learning these structures is a hard problem in itself. Moreover, it's not clear how to incorporate the machinery of causality with online continual learning. In this work, we take an indirect approach to discovering causal models. Instead of searching for the true causal model directly, we propose an online algorithm that continually detects and removes spurious features. Our algorithm works on the idea that the correlation of a spurious feature with a target is not constant over-time. As a result, the weight associated with that feature is constantly changing. We show that by continually removing such features, our method converges to solutions that have strong generalization. Moreover, our method combined with random search can also discover non-spurious features from raw sensory data. Finally, our work highlights that the information present in the temporal structure of the problem -- destroyed by shuffling the data -- is essential for detecting spurious features online.

Motivation & Objective

• To address the brittleness of predictive models trained on incomplete data distributions by incorporating causal invariance.
• To enable online continual learning with causal generalization, avoiding reliance on full data coverage.
• To develop a scalable, non-gradient-based method for identifying and removing spurious features in real-time.
• To demonstrate that temporal structure in data—preserved in online learning—is essential for detecting spurious correlations.
• To show that random search over parameter perturbations can effectively discover non-spurious features from raw sensory inputs.

Proposed method

• The method uses online perturbation-based search to detect spurious features by measuring changes in feature variance over time.
• It maintains two objectives: minimizing prediction loss and minimizing the variance (v) of feature weights across time, with v computed via offline estimation for efficiency.
• Perturbations are applied selectively to convolutional and fully connected layers to avoid conflict between loss and variance minimization.
• The algorithm uses a binarized feature representation (PwB*) to estimate v efficiently, enabling online operation.
• It leverages the temporal instability of spurious features—whose correlation with the target fluctuates over time—as a signal for removal.
• The approach avoids gradient-based optimization for feature selection, enabling credit assignment across long time horizons.

Experimental results

Research questions

• RQ1 Can spurious features be detected and removed in an online, continual learning setting without prior knowledge of interventions or full data coverage?
• RQ2 Does the temporal instability of spurious features provide a reliable signal for causal feature discovery in real-time?
• RQ3 Can a non-gradient-based, perturbation-driven method effectively identify and eliminate spurious correlations in neural networks?
• RQ4 To what extent can this method discover non-spurious features directly from raw sensory data?
• RQ5 How does the temporal structure of data—preserved in online learning—enable better detection of spurious correlations than shuffled data?

Key findings

• PwB* achieved 68.83 ± 0.49 test accuracy on the Online Colored MNIST benchmark, outperforming baseline methods in online settings.
• The variant PwB* (0.85), which fixes the spurious correlation to be stable, used color information as expected, confirming the method’s sensitivity to correlation instability.
• A strong Pearson correlation (r = +0.91) was found between online and offline estimates of v, validating the use of offline v estimation for speed without sacrificing accuracy.
• The method successfully discovered non-spurious features from raw images using random search, demonstrating applicability to sensory data.
• PwB* was shown to be scalable and compatible with online continual learning, unlike gradient-based methods such as Oracle IRM.
• The results highlight that temporal structure in data is essential for spurious feature detection—shuffling destroys this signal and undermines detection.