Lazy Estimation of Variable Importance for Large Neural Networks

Yue Gao, Abby Stevens, Garvesh Raskutti, Rebecca Willett

As neural networks (NNs) become central to high-stakes decision-making in fields like healthcare and climate science, their opaque "black box" nature becomes a significant liability. In many critical applications, we don't just need a prediction; we need to know why the model made it. Specifically, we need to quantify Variable Importance (VI)—the contribution of each input feature to the model's overall performance.

In the paper "Lazy Estimation of Variable Importance for Large Neural Networks," researchers introduce LazyVI. This novel method bridges the gap between computationally expensive "retraining" techniques and fast but often inaccurate "dropout" methods, providing a mathematically rigorous way to peek inside large-scale models.

The Variable Importance Bottleneck

There are two traditional ways to measure model-agnostic variable importance, both of which have severe drawbacks:

  1. Retraining (The Gold Standard): You train a "full" model on all variables, then train a "reduced" model for every single variable you want to test (excluding that specific variable). The difference in predictive power is your VI. While highly accurate, this is computationally ruinous for large NNs or high-dimensional data. If you have 1,000 variables, you must train 1,001 separate models.
  2. Dropout/Plug-in (The Fast Approximation): You train one model and simply "zero out" or replace a variable with its mean during testing. This is fast, but often wrong. If two variables are highly correlated, dropping one doesn't hurt the model much because the other acts as a proxy, leading to a severe underestimation of the dropped variable's true importance.

The core challenge is finding a method that is as accurate as retraining but as fast as dropout, while also providing statistical confidence in the results.

The LazyVI Approach: Linearization and NTK

To solve this, the authors propose a "lazy" retraining framework. Instead of training a new neural network from scratch for every excluded variable, they linearize the neural network around the parameters of the already-trained full model.

The core insight is that for sufficiently wide networks, the change in parameters required to adapt to a missing variable is relatively small. Therefore, a linear approximation isn't just a shortcut—it is mathematically justified by Neural Tangent Kernel (NTK) theory.

By using a first-order Taylor expansion of the NN, the problem of retraining becomes a convex Ridge Regression problem in the gradient space of the model. By solving a linear system rather than a non-convex optimization, LazyVI finds the exact "parameter correction" needed to simulate removing a variable. Furthermore, unlike dropout, LazyVI correctly identifies the importance of correlated features because it approximates what the model would have learned if that feature were never there.

Key Results and Performance

The authors validated LazyVI across several regimes, including synthetic simulations, MNIST image data, and real-world climate data (predicting winter precipitation using Sea Surface Temperature anomalies).

Speed vs. Accuracy

In empirical tests, LazyVI proved to be a massive leap forward in efficiency. In binary classification tests, LazyVI achieved nearly the same accuracy as full retraining but was over 10x faster.

Error vs. Time comparison showing LazyVI achieving low error at a fraction of the computational cost of full retraining.

As shown above, LazyVI clusters near zero seconds with minimal error, vastly outperforming standard dropout methods in accuracy while avoiding the heavy computational burden of full retraining. The authors also demonstrated that LazyVI can speed up the calculation of Shapley values—a popular but notoriously slow VI metric—by roughly 5 times without sacrificing performance.

Statistical Rigor

A major breakthrough of LazyVI is its statistical foundation. The authors prove that their estimator is asymptotically normal. This allows researchers to generate Wald-type confidence intervals, meaning they can state with 95% confidence that a variable's importance falls within a specific range.

Coverage and Bias comparison between LazyVI and full retraining across multiple variables, demonstrating LazyVI's statistical reliability.

When comparing coverage and bias across multiple variables, LazyVI closely tracks the gold-standard retraining method, proving that its speed does not come at the cost of statistical reliability.

Limitations and Future Work

While powerful, LazyVI has boundaries. The theoretical guarantees rely on the "NTK regime," meaning the networks must be sufficiently wide. While it works well in practice on medium-sized networks, the theory is strongest for very large ones. Additionally, the current work focuses primarily on feedforward (MLP) and convolutional (CNN) architectures; extending these guarantees to Transformers or Recurrent Neural Networks remains an open area for future research. Finally, the method requires tuning a ridge penalty hyperparameter, which adds a minor layer of computational overhead.

Despite these limitations, LazyVI represents a significant step forward in making large, complex neural networks interpretable, accountable, and scientifically useful.

Summary updated

Abstract

As opaque predictive models increasingly impact many areas of modern life, interest in quantifying the importance of a given input variable for making a specific prediction has grown. Recently, there has been a proliferation of model-agnostic methods to measure variable importance (VI) that analyze the difference in predictive power between a full model trained on all variables and a reduced model that excludes the variable(s) of interest. A bottleneck common to these methods is the estimation of the reduced model for each variable (or subset of variables), which is an expensive process that often does not come with theoretical guarantees. In this work, we propose a fast and flexible method for approximating the reduced model with important inferential guarantees. We replace the need for fully retraining a wide neural network by a linearization initialized at the full model parameters. By adding a ridge-like penalty to make the problem convex, we prove that when the ridge penalty parameter is sufficiently large, our method estimates the variable importance measure with an error rate of O(1/n) where n is the number of training samples. We also show that our estimator is asymptotically normal, enabling us to provide confidence bounds for the VI estimates. We demonstrate through simulations that our method is fast and accurate under several data-generating regimes, and we demonstrate its real-world applicability on a seasonal climate forecasting example.
BibTeX
@inproceedings{gao2022lazy,
  author = {Yue Gao and Abby Stevens and Garvesh Raskutti and Rebecca Willett},
  title = {{Lazy Estimation of Variable Importance for Large Neural Networks}},
  booktitle = {International Conference on Machine Learning},
  year = {2022},
  url = {https://openreview.net/forum?id=x6eUFxLjuy},
}
Figure