Identifying Wrongly Predicted Samples: A Method for Active Learning

Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino

Abstract

State-of-the-art machine learning models require access to significant amount of annotated data in order to achieve the desired level of performance. While unlabelled data can be largely available and even abundant, annotation process can be quite expensive and limiting. Under the assumption that some samples are more important for a given task than others, active learning targets the problem of identifying the most informative samples that one should acquire annotations for. Instead of the conventional reliance on model uncertainty as a proxy to leverage new unknown labels, in this work we propose a simple sample selection criterion that moves beyond uncertainty. By first accepting the model prediction and then judging its effect on the generalization error, we can better identify wrongly predicted samples. We further present an approximation to our criterion that is very efficient and provides a similarity based interpretation. In addition to evaluating our method on the standard benchmarks of active learning, we consider the challenging yet realistic scenario of imbalanced data where categories are not equally represented. We show state-of-the-art results and better rates at identifying wrongly predicted samples. Our method is simple, model agnostic and relies on the current model status without the need for re-training from scratch.

Introduction

Deep learning's success is heavily dependent on large-scale annotated datasets. However, the process of labeling data is often the primary bottleneck in model development. Active Learning (AL) seeks to optimize this process by selecting only the most informative samples for human annotation. Traditional AL strategies typically rely on model uncertainty—selecting samples where the model's confidence is lowest. However, high uncertainty does not always correlate with incorrect predictions, and conversely, a model can be highly confident yet wrong.

This paper addresses these shortcomings by proposing a selection criterion that identifies samples the model is likely to mispredict. This approach is particularly effective in real-world scenarios involving imbalanced data, where standard methods often struggle to identify rare but critical samples.

Methodology

The core innovation is a "hypothesize and test" approach to sample selection. Instead of measuring uncertainty, the method evaluates how a sample's predicted label (pseudo-label) affects the model's generalization error on a small, labeled holdout set.

Visual representation of the selection process: (a) initial state, (b) testing a wrong prediction, (c) testing a correct prediction, (d) final selection.

The Selection Criterion

  1. Pseudo-labeling: For an unlabeled sample, the model's current prediction is treated as the ground truth.
  2. Local Update: The model is briefly updated (via a single gradient step) using this pseudo-label.
  3. Error Measurement: The change in loss is measured on a small holdout set. If the loss increases significantly, it suggests the pseudo-label was likely incorrect, identifying the sample as a high-priority candidate for manual annotation.

Efficient Approximation (Ours-app)

To avoid the computational cost of updating the model for every unlabeled sample, the authors derive a first-order approximation. This reduces the criterion to a dot product between the gradient of the unlabeled sample and the gradient of the holdout set, providing a similarity-based interpretation linked to the Neural Tangent Kernel (NTK).

Experimental Setup

The method was evaluated on several image classification benchmarks (MNIST, KMNIST, SVHN, CIFAR-10) and a semantic segmentation task (Cityscapes). Crucially, the researchers tested these methods on imbalanced variants of the datasets to simulate realistic data distributions where certain classes are underrepresented.

Results and Discussion

The proposed method consistently outperformed standard AL baselines, particularly in the presence of data imbalance.

Performance comparison across annotation steps, showing 'Ours' and 'Ours-app' outperforming standard methods.

Key Findings

  • Higher Mistake Hit Rate: The method is significantly more effective at identifying mispredicted samples. On MNIST, it selected wrongly predicted samples approximately 90% of the time in the initial step, compared to 70% for uncertainty-based methods.
  • Efficiency in Imbalanced Settings: In imbalanced MNIST tests, the method reached target accuracy in half the annotation steps required by random sampling. It naturally focuses on underrepresented classes where the model's error rate is higher.
  • Robustness of the Holdout Set: The method is surprisingly robust to the size of the holdout set; as few as two samples per class are sufficient to serve as an effective anchor for identifying errors.

Conclusion

By moving beyond simple uncertainty and focusing on the impact of predictions on generalization error, this method provides a more robust framework for active learning. It is model-agnostic, computationally efficient via its first-order approximation, and excels in the challenging but common scenario of imbalanced data. Future work may explore dynamic updates to the holdout set to further enhance long-term training stability.

Summary updated

Abstract

State-of-the-art machine learning models require access to significant amount of annotated data in order to achieve the desired level of performance. While unlabelled data can be largely available and even abundant, annotation process can be quite expensive and limiting. Under the assumption that some samples are more important for a given task than others, active learning targets the problem of identifying the most informative samples that one should acquire annotations for. Instead of the conventional reliance on model uncertainty as a proxy to leverage new unknown labels, in this work we propose a simple sample selection criterion that moves beyond uncertainty. By first accepting the model prediction and then judging its effect on the generalization error, we can better identify wrongly predicted samples. We further present an approximation to our criterion that is very efficient and provides a similarity based interpretation. In addition to evaluating our method on the standard benchmarks of active learning, we consider the challenging yet realistic scenario of imbalanced data where categories are not equally represented. We show state-of-the-art results and better rates at identifying wrongly predicted samples. Our method is simple, model agnostic and relies on the current model status without the need for re-training from scratch.
BibTeX
@inproceedings{aljundi2022identifying,
  author = {Rahaf Aljundi and Nikolay Chumerin and Daniel Olmeda Reino},
  title = {{Identifying Wrongly Predicted Samples: A Method for Active Learning}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
  url = {https://openreview.net/forum?id=TjBHwYl5Pig},
}
Figure