Robust Transfer Learning for Active Level Set Estimation with Locally Adaptive Gaussian Process Prior

Giang Ngo, Dang Nguyen, Sunil Gupta

Abstract

The objective of active level set estimation for a black-box function is to precisely identify regions where the function values exceed or fall below a specified threshold by iteratively performing function evaluations to gather more information about the function. This becomes particularly important when function evaluations are costly, drastically limiting our ability to acquire large datasets. A promising way to sample-efficiently model the black-box function is by incorporating prior knowledge from a related function. However, this approach risks slowing down the estimation task if the prior knowledge is irrelevant or misleading. In this paper, we present a novel transfer learning method for active level set estimation that safely integrates a given prior knowledge while constantly adjusting it to guarantee a robust performance of a level set estimation algorithm even when the prior knowledge is irrelevant. We theoretically analyze this algorithm to show that it has a better level set convergence compared to standard transfer learning approaches that do not make any adjustment to the prior. Additionally, extensive experiments across multiple datasets confirm the effectiveness of our method when applied to various different level set estimation algorithms as well as different transfer learning scenarios.

Introduction

Level Set Estimation (LSE) is a fundamental task in scientific computing and engineering, aimed at identifying regions where an expensive-to-evaluate black-box function $f(x)$ crosses a predefined threshold $h$. Applications range from environmental monitoring (identifying "hot spots") to manufacturing quality control (detecting defects). Given the high cost of function evaluations, Active Learning is typically employed to strategically select sample points that maximize information gain.

Transfer learning offers a potential acceleration for LSE by incorporating prior knowledge, such as simulation data or results from previous experiments. However, a significant bottleneck exists: if the prior is inaccurate—due to model mismatch or changing conditions—standard LSE algorithms can be severely biased, leading to performance that is worse than starting from scratch. This paper introduces Adaptive Prior Level Set Estimation (AP-LSE), a framework designed to leverage prior knowledge safely by locally adapting to real-time observations.

Methodology: Adaptive Prior Level Set Estimation (AP-LSE)

Most LSE frameworks utilize Gaussian Processes (GPs) as surrogate models. In a traditional transfer learning setup, a prior function $u_p(x)$ is directly used as the mean function of the GP. If $u_p(x)$ deviates significantly from the true function $f(x)$, the resulting posterior mean remains biased toward the incorrect prior, especially in regions with sparse data.

The Locally Adaptive Adjustment

The core innovation of AP-LSE is the introduction of a locally adaptive adjustment term, $u_{dt}(x)$. Instead of treating the prior as an immutable truth, the algorithm treats it as a baseline that can be corrected. The process follows three main steps:

  1. Discrepancy Observation: At each iteration, the algorithm measures the difference between the actual observed values and the predictions provided by the prior $u_p(x)$.
  2. Error Interpolation: These discrepancies are used to construct a smooth adjustment function that interpolates the errors across the input space.
  3. Prior Correction: The effective prior for the next iteration becomes the sum of the original prior and this adjustment: $u_{new}(x) = u_p(x) + u_{dt}(x)$.
Comparison of posterior distributions: The adaptive posterior (green) successfully corrects the misleading prior (dashed blue) to match observations, whereas the vanilla posterior (orange) remains biased.

As shown in the figure above, the adaptive posterior is able to "snap" to the observations even when the initial prior is significantly offset. In regions where data is available, the adjusted prior aligns with the truth; in unsampled regions, the adjustment naturally decays, allowing the model to revert to the original prior knowledge without introducing global instability.

Experimental Evaluation

The authors validated AP-LSE across a diverse set of synthetic and real-world benchmarks:

  • Synthetic Functions: Testing on Bird, Multi-circle 3D (MC3D), and Mishra03 functions allowed for controlled experiments with varying levels of prior mismatch.
  • Solar Cell Ingot Analysis: Identifying low carrier lifetime regions in silicon ingots by using data from one ingot as a prior for another.
  • Network Latency Mapping: Finding global regions where internet round-trip times (pings) are below 200ms.
  • Algorithmic Assurance (AA): Identifying specific image transformations (e.g., rotation, shear) that cause a pre-trained MNIST classifier's accuracy to drop below a 96% threshold.

Key Results and Insights

The experimental results highlight three major advantages of the AP-LSE approach:

  1. Sample Efficiency: AP-LSE consistently achieves high F1-scores with significantly fewer samples than both "from-scratch" active learning and vanilla transfer learning.
  2. Robustness to Misleading Priors: Unlike standard transfer learning, which can fail catastrophically when the prior is irrelevant, AP-LSE maintains performance at least comparable to non-transfer methods. It effectively "ignores" bad information by localizing the correction.
  3. Algorithmic Versatility: The method is algorithm-agnostic, functioning as a plug-and-play enhancement for various LSE acquisition functions, including STR (Straddle), RMILE, and C2LSE.

Furthermore, the paper provides theoretical guarantees, proving that AP-LSE ensures better data fitting than vanilla transfer learning and offers a superior convergence rate for classification error when using RBF kernels.

Conclusion and Future Work

AP-LSE addresses a critical vulnerability in transfer learning for active LSE. By introducing a locally adaptive adjustment to the prior, it ensures that prior knowledge acts as a helpful guide rather than a misleading bias. While currently optimized for mean-function adaptation with RBF kernels, future work could extend this robustness to the covariance structure of the GP, potentially handling cases where the prior's uncertainty estimates are also unreliable.

Summary updated

Abstract

The objective of active level set estimation for a black-box function is to precisely identify regions where the function values exceed or fall below a specified threshold by iteratively performing function evaluations to gather more information about the function. This becomes particularly important when function evaluations are costly, drastically limiting our ability to acquire large datasets. A promising way to sample-efficiently model the black-box function is by incorporating prior knowledge from a related function. However, this approach risks slowing down the estimation task if the prior knowledge is irrelevant or misleading. In this paper, we present a novel transfer learning method for active level set estimation that safely integrates a given prior knowledge while constantly adjusting it to guarantee a robust performance of a level set estimation algorithm even when the prior knowledge is irrelevant. We theoretically analyze this algorithm to show that it has a better level set convergence compared to standard transfer learning approaches that do not make any adjustment to the prior. Additionally, extensive experiments across multiple datasets confirm the effectiveness of our method when applied to various different level set estimation algorithms as well as different transfer learning scenarios.
BibTeX
@inproceedings{ngo2024robust,
  author = {Giang Ngo and Dang Nguyen and Sunil Gupta},
  title = {{Robust Transfer Learning for Active Level Set Estimation with Locally Adaptive Gaussian Process Prior}},
  booktitle = {ACML Conference Track},
  year = {2024},
  url = {https://openreview.net/forum?id=xt4oAjRL7g},
}
Figure