Training Set Debugging Using Trusted Items

Xuezhou Zhang, Xiaojin Zhu, Stephen J. Wright

Abstract: Training set bugs are flaws in the data that adversely affect machine learning. The training set is usually too large for manual inspection, but one may have the resources to verify a few trusted items. The set of trusted items may not by itself be adequate for learning, so we propose an algorithm that uses these items to identify bugs in the training set and thus improves learning. Specifically, our approach seeks the smallest set of changes to the training set labels such that the model learned from this corrected training set predicts labels of the trusted items correctly. We flag the items whose labels are changed as potential bugs, whose labels can be checked for veracity by human experts. To find the bugs in this way is a challenging combinatorial bilevel optimization problem, but it can be relaxed into a continuous optimization problem. Experiments on toy and real data demonstrate that our approach can identify training set bugs effectively and suggest appropriate changes to the labels. Our algorithm is a step toward trustworthy machine learning.

Introduction

Machine learning models are fundamentally limited by the quality of their training data. Large-scale datasets often contain "bugs"—mislabeled items resulting from human error, sensor noise, or deep-seated historical biases. While domain experts can verify a small subset of "trusted items," manually inspecting entire datasets is unfeasible.

The challenge lies in identifying systematic bugs, such as historical biases against specific demographics, which may appear self-consistent to a learning algorithm. Standard heuristics often fail to capture the non-linear relationship between training labels and model predictions. This paper introduces DUTI (Debugging Using Trusted Items), a framework that leverages a small set of verified data to identify and suggest fixes for mislabeled training points.

The DUTI Framework

DUTI frames training set debugging as a bilevel optimization problem. The core intuition is to find the minimal set of changes to the training labels that would result in a model capable of correctly predicting the trusted items.

  1. Outer Loop: This objective minimizes the difference between the original and "corrected" training labels, while also minimizing loss on the trusted items.
  2. Inner Loop: This represents the standard Empirical Risk Minimization (ERM) process that learns model parameters from the corrected labels.
Figure 1: (a) Training data with systematic bias and a small set of trusted items (red). (b) Bugs flagged by the DUTI algorithm in the Harry Potter toy example.

To solve this NP-hard problem, the authors relax it into a continuous space. By applying the Karush–Kuhn–Tucker (KKT) conditions and the Implicit Function Theorem, they calculate how infinitesimal changes in training labels affect final model parameters. This allows for gradient-based optimization to identify bugs and suggest corrected labels.

Experimental Evaluation

The authors evaluated DUTI across several contexts:

  • Systematic Bias: A "Harry Potter" toy example and UCI datasets (German Loan, Adult Income) were used to simulate historical biases (e.g., age or gender discrimination).
  • Label Noise: The MNIST dataset was used to simulate human error by blurring images and using a weak model for initial labeling.
  • Regression: A sine wave experiment tested the algorithm's ability to correct systematic functional errors.

Key Results

DUTI consistently outperformed baseline methods like Influence Functions (INF), Nearest Neighbor (NN), and Label Noise Detection (LND). Unlike local heuristics, DUTI considers the global behavior of the learning algorithm, making it uniquely capable of identifying systematic errors that span large regions of the feature space.

Figure 2: Comparison of bug detection methods (INF, NN, LND) and the resulting Precision-Recall curves, where DUTI demonstrates superior performance.

In the German Loan and Adult Income datasets, DUTI successfully isolated simulated historical biases where baselines only flagged random points near the trusted set. On MNIST, the algorithm recovered over half of the 133 buggy labels within its first 200 flagged items, demonstrating high precision in a "budgeted" human-in-the-loop scenario.

Limitations and Future Work

The primary bottleneck for DUTI is scalability. Because the optimization involves an implicit function of the learner, it is computationally expensive for datasets beyond a few thousand points. Future work will focus on faster approximation methods. Additionally, the effectiveness of DUTI depends on the informativeness of the trusted set; if trusted items do not cover biased regions, bugs may remain hidden. Finally, if flagged items are found to be correct by experts, it serves as a diagnostic signal that the model architecture itself may be misspecified.

Summary updated

Abstract

Training set bugs are flaws in the data that adversely affect machine learning. The training set is usually too large for man- ual inspection, but one may have the resources to verify a few trusted items. The set of trusted items may not by itself be adequate for learning, so we propose an algorithm that uses these items to identify bugs in the training set and thus im- proves learning. Specifically, our approach seeks the smallest set of changes to the training set labels such that the model learned from this corrected training set predicts labels of the trusted items correctly. We flag the items whose labels are changed as potential bugs, whose labels can be checked for veracity by human experts. To find the bugs in this way is a challenging combinatorial bilevel optimization problem, but it can be relaxed into a continuous optimization problem. Ex- periments on toy and real data demonstrate that our approach can identify training set bugs effectively and suggest appro- priate changes to the labels. Our algorithm is a step toward trustworthy machine learning.
BibTeX
@inproceedings{zhang2018training,
  author = {Xuezhou Zhang and Xiaojin Zhu and Stephen J. Wright},
  title = {{Training Set Debugging Using Trusted Items}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11610},
}

Loading related content...

Figure