Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models

TL;DR

The paper tackles the bottleneck of identifying detrimental training samples in deep models by reframing influence-function based detection as gradient-space outlier analysis, eliminating the costly Hessian inversion. It introduces Outlier Gradient Analysis, using detectors like Isolation Forest and thresholding on per-sample gradients to label outliers and trim them without retraining from scratch. Across synthetic data, noisy-label vision tasks, NLP data selection, and LLM influential data benchmarks, the approach achieves high detection accuracy and practical gains while remaining computationally efficient. This Hessian-free, data-centric method broadens the applicability of data valuation techniques to large-scale non-convex models and real-world noisy-data scenarios, enabling scalable improvements in model utility.

Abstract

A core data-centric learning challenge is the identification of training samples that are detrimental to model performance. Influence functions serve as a prominent tool for this task and offer a robust framework for assessing training data influence on model predictions. Despite their widespread use, their high computational cost associated with calculating the inverse of the Hessian matrix pose constraints, particularly when analyzing large-sized deep models. In this paper, we establish a bridge between identifying detrimental training samples via influence functions and outlier gradient detection. This transformation not only presents a straightforward and Hessian-free formulation but also provides insights into the role of the gradient in sample impact. Through systematic empirical evaluations, we first validate the hypothesis of our proposed outlier gradient analysis approach on synthetic datasets. We then demonstrate its effectiveness in detecting mislabeled samples in vision models and selecting data samples for improving performance of natural language processing transformer models. We also extend its use to influential sample identification for fine-tuning Large Language Models.

Figures (5)

• Figure 1: Illustrating our outlier gradient analysis approach on two synthetic datasets and convex/non-convex models. A-D showcase our outlier gradient analysis approach on a 2D linearly separable synthetic dataset. This dataset includes a small subset of detrimental samples with incorrect labels used to train a Logistic Regression binary classification model. Meanwhile, E-H depict our outlier gradient analysis on a non-linear synthetic dataset with mislabeled samples employed in training a Multilayer Perceptron (MLP) neural network. In subfigures A and E, the training sets are represented with class labels 0 (red) and 1 (blue) in the convex and non-convex cases, respectively. Detrimental samples with incorrect class labels are marked with $\times$, while regular samples are marked with $\circ$. B and F denote the test sets used to evaluate model performance. C and G display the influence scores calculated by Eq. (\ref{['eq:influence']}). Note that G demonstrates that influence scores are not reliable indicators for detecting detrimental samples in the non-convex case. After applying outlier analysis on the gradient space of the non-convex MLP model, most detrimental samples are detected. D and H showcase the gradient space obtained for each sample from the Logistic Regression and MLP models, respectively. It is evident that the outlier samples correspond to detrimental samples with mislabeled classes, which are linearly or non-linearly separated from inliers. Note that the benefits of outlier gradient trimming can be clearly observed—removing predicted outlier samples via iForest and retraining the MLP enhances classification performance from $90\% \rightarrow 96\%$ on the test set (refer to Table \ref{['tab:synthetic']}).
• Figure 2: Detrimental samples detected using our outlier gradient analysis. Top row: CIFAR-10N; bottom row: CIFAR-100N. Top label (red): noisy label; bottom label (green): correct class.
• Figure 3: Performance of the data selection task using outlier gradient trimming and other influence baselines for fine-tuning RoBERTa.
• Figure 4: Results for outlier gradient analysis on LLM influential data identification benchmarks.
• Figure : : Outlier Gradient Analysis and Trimming