PaperlayerView exact revision on arXiv

Source-linked AI summary

Influence Selection for Active Learning

Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, Conghui He

arXiv:2108.09331v1cs.LGcs.CV

TL;DR

Active-learning methods often rank samples by uncertainty or dataset diversity, and computer-vision research has limited coverage beyond classification. ISAL estimates influence for unlabeled samples using expected gradients, and reports state-of-the-art performance across classification and detection with lower annotation costs.

  • Problem

    Existing approaches select samples using uncertainty or diversity criteria, while computer-vision active learning has been limited particularly in object detection.

    Source: §1 · pp. 1–2

  • Method

    ISAL uses UUIC to estimate the expected gradients of unlabeled, untrained samples and select those with the most positive influence on model performance.

    Source: §1 · p. 2

  • Results

    ISAL achieves state-of-the-art performance across image classification and object detection, saving 12%, 13%, and 16% annotation on CIFAR10, VOC2012, and COCO, respectively.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors state that ISAL may extend to active-learning problems beyond computer vision.

    Source: §5 · p. 8

    The influence analysis assumes a twice-differentiable loss that is strictly convex in θ, with a positive-definite Hessian by assumption.

    Source: §3.2 · p. 4

Abstract

from arXiv · show

The existing active learning methods select the samples by evaluating the sample's uncertainty or its effect on the diversity of labeled datasets based on different task-specific or model-specific criteria. In this paper, we propose the Influence Selection for Active Learning(ISAL) which selects the unlabeled samples that can provide the most positive Influence on model performance. To obtain the Influence of the unlabeled sample in the active learning scenario, we design the Untrained Unlabeled sample Influence Calculation(UUIC) to estimate the unlabeled sample's expected gradient with which we calculate its Influence. To prove the effectiveness of UUIC, we provide both theoretical and experimental analyses. Since the UUIC just depends on the model gradients, which can be obtained easily from any neural network, our active learning algorithm is task-agnostic and model-agnostic. ISAL achieves state-of-the-art performance in different active learning settings for different tasks with different datasets. Compared with previous methods, our method decreases the annotation cost at least by 12%, 13% and 16% on CIFAR10, VOC2012 and COCO, respectively.

1. Introduction

Active learning seeks better model performance with fewer labeled samples, but computer-vision methods are limited in task coverage and use uncertainty or diversity as selection proxies. ISAL instead estimates unlabeled-sample influence and reports state-of-the-art results across classification and detection.

1. Introduction

  • Uncertainty-based selection may misjudge importance because neural networks can be overconfident on unfamiliar samples, while diversity methods ignore model state and may require quadratic pairwise-distance computation.The latter cost becomes more apparent on large-scale datasets.Source: §1 · pp. 1–2
  • Active-learning research for object detection is rare, and most existing methods are designed for specific architectures such as Faster R-CNN or SSD.Object-detection annotation is extremely time-consuming.Source: §1 · p. 2
  • The authors frame ISAL as applicable across tasks and architectures, estimating sample influence directly rather than relying on feature distance or expected loss as proxies.The cited alternatives are diversity-based coreset and uncertainty-based learning loss.Source: §1 · p. 2
  • ISAL selects samples with the most positive influence on model performance, estimated for unlabeled, untrained samples by UUIC using their expected gradients.UUIC uses model gradients, making ISAL task-agnostic and model-agnostic; more negative influence values indicate more positive influence on performance.Source: §1 · pp. 1–2
  • ISAL saves 12%, 13%, and 16% annotation versus the best comparing methods on CIFAR10, VOC2012, and COCO, respectively, and achieves state-of-the-art classification and object-detection performance.Source: §1 · p. 2
  • The paper introduces ISAL as a task-agnostic, model-agnostic algorithm that selects samples by calculated influence, and UUIC estimates that influence for unlabeled, untrained samples.The authors provide theoretical and experimental analyses to validate UUIC.Source: §1 · p. 2

2. Related Work

Prior active-learning work is organized around uncertainty and diversity criteria, with many methods developed for classification. Object-detection work is less common and often architecture-specific; influence functions have not been established for untrained samples.

2. Related Work

  • Existing active-learning methods are grouped into uncertainty-based and diversity-based approaches, and many target image classification or adapt to it with little change.The paper's related-work discussion treats these as the two broad categories.Source: §2 · p. 2
  • Uncertainty methods use posterior probabilities, margins, entropy, committee agreement, Monte Carlo Dropout, predicted loss, or expected output changes to rank samples.The passage attributes these approaches to,,, Seung, Gal, Yoo, and Freytag, respectively.Source: §2 · p. 2
  • Diversity methods address batch-query sampling bias through clustering, discrete optimization, core-set construction, context-aware distances, or discrimination between labeled and unlabeled samples.The passage cites,, [31],, and Sinha for these approaches.Source: §2 · pp. 2–3
  • Object-detection active learning is described as rare and often complex, with methods using SSD layer offsets, Faster R-CNN uncertainty, bbox-level uncertainty, or large-scale evaluations.Haussmann found that the best-performing method selected images with more boxes, increasing annotation cost; detection cost is closely related to box count rather than image count.Source: §2 · p. 3
  • Influence functions began in robust statistics and were later used to study neural networks and trained unlabeled samples in semi-supervised learning [28], but not untrained samples.The passage notes that Cook's derivation is based on trained samples, leaving no solid theoretical support for applying the function to untrained samples.Source: §2 · p. 3

3. Method

ISAL estimates how unlabeled samples may affect model performance and selects samples for annotation using those influence scores. UUIC estimates the gradients needed to calculate influence without ground-truth labels.

3.1. Problem Definition

  • In each active-learning round, the method evaluates unlabeled samples, annotates a selected batch, updates the labeled and unlabeled sets, and retrains until validation performance is satisfactory or the annotation budget is exhausted.Source: §3.1 · p. 3

3.2. The Influence of an Untrained Sample

  • An influence function approximates the parameter shift from adding a labeled candidate, and the method measures its effect on reference-set loss.The first-step validation set serves as the reference set without additional annotation; using labeled data is also possible, but the validation set performs better.Source: §3.2 · p. 4

3.3. Untrained Unlabeled sample Influence Calculation

  • UUIC estimates an unlabeled sample’s expected gradient by averaging gradients for top-K predicted classes, weighted by their posterior probabilities.For object detection, the algorithm uses predicted bounding boxes that pass a threshold as pseudo-labels.Source: §3.3 · p. 4
  • Experiments report that K=1 gives the best performance at each step, and applying this gradient-estimation approach to object detection also achieves state-of-the-art performance.Source: §3.3 · p. 4

3.4. Influence Selection for Active Learning

  • ISAL computes influence for each unlabeled sample, ranks samples by influence, and selects a batch according to the active-learning setting; more negative values indicate more positive influence.Source: §3.4 · p. 5

4. Experiment

Experiments test ISAL across image classification and object detection datasets, including large-scale detection, then examine its components and selected samples. The results cover annotation savings, ablations, and visualizations.

4.1. Image Classification

  • Across CIFAR10, CIFAR100, and SVHN, ISAL saves 12%, 2.9%, and 12% versus coreset sampling, and 23%, 9.3%, and 14% versus random sampling, respectively.These savings are measured when ISAL reaches each baseline’s final performance after 10 iterations.Source: §4.1 · p. 5, Figure 3 · p. 6

4.2. Object Detection

  • On VOC2012, ISAL saves 13% of annotations versus coreset sampling at 42% mAP; on COCO, it saves 16% versus random and 59% versus learning loss at the target AP.The VOC2012 comparison also reports 26% fewer annotations than localization stability sampling.Source: §4.2 · p. 6, Figure 4 · p. 7

4.3. Large Scale Experiment in Object Detection

  • At 34% AP on large-scale COCO, ISAL uses 8% fewer annotations than coreset sampling and needs 60.7% of dataset annotations to reach 94.4% of full-data performance.The comparison is shown in Figure 4(c).Source: §4.3 · p. 7, Figure 4 · p. 7

4.4.1 The Effect of K in Expected gradient

  • With K=1, ISAL performs best at each step; averaging gradients across predicted classes can diminish components when their directions oppose one another.The ablation varies K in the expected gradient Gz′.Source: §4.4.1 · p. 7, Table 1 · p. 7

4.4.2 The Effect of H−1

  • Gradient Similarity performs similarly to ISAL, though ISAL performs better; H^-1 helps ISAL identify positively influential samples with different expected gradients.Gradient Similarity compares reference-set gradients with expected gradients of unlabeled samples.Source: §4.4.2 · p. 7, Table 2 · p. 7

4.4.3 The Selection of Reference Set

  • Using L1 or the labeled set as reference produces slightly worse results than validation-set ISAL, though both variants outperform random sampling.L1 is closer in distribution to the unlabeled pool, but its trained-model gradients are smaller and may make influence estimates less precise.Source: §4.4.3 · p. 8, Table 3 · p. 8

4.5. Visualization Analysis

  • On CIFAR10, ISAL selects more cat, bird, and deer images, classes where the model has lower accuracy; uniform class sampling would add redundancy.On COCO, it selects images with fewer but larger, less-overlapping boxes, then later selects images with more objects and complex scenes.Source: §4.5 · p. 8, Figure 5 · p. 8, Figure 6 · p. 8

5. Conclusion

ISAL uses UUIC to select unlabeled samples for positive influence on model performance. The authors report state-of-the-art performance across tasks and settings, with a task- and model-agnostic design.

5. Conclusion

  • ISAL achieves state-of-the-art performance across different tasks in commonly used and newly designed large-scale settings.The authors propose extending it to active-learning problems beyond computer vision.Source: §5 · p. 8
  • UUIC calculates each unlabeled sample’s influence value, and ISAL selects samples expected to provide the most positive influence on model performance.Source: §5 · p. 8
  • The authors characterize ISAL as task-agnostic and model-agnostic and say it helps neural networks learn more effectively while decreasing annotation cost.Source: §5 · p. 8

Supplementary Material

The supplementary material provides additional derivations, implementation details, complexity analysis, and experiments beyond the main paper.

Supplementary Material

A. The Derivation of the Influence of Untrained Samples

The derivation estimates how adding an untrained sample changes model parameters, then uses that change to evaluate the sample’s influence.

A. The Derivation of the Influence of Untrained Samples

B. The Implementation Details of stest Calculation

The stest calculation approximates an inverse-Hessian product using a Taylor expansion and stochastic estimates from randomly sampled labeled examples.

B. The Implementation Details of stest Calculation

C. The Time Complexity Analysis

The analysis reduces influence computation to a fixed-cost estimate from labeled images followed by a linear pass over unlabeled samples. It also records implementation choices for the classification and detection baselines.

C. The Time Complexity Analysis

  • With a fixed number of labeled images, estimating stest costs a constant C, and calculating influence for n unlabeled samples then costs O(n).The method samples labeled images to estimate stest rather than directly calculating H^-1.Source: C. The Time Complexity Analysis · p. 11

D.1. Image Classification

  • For classification, coreset sampling uses K-Ceter-Greedy, while learning-loss sampling predicts loss on a random subset of 10,000 unlabeled images.The loss-prediction module's gradient is stopped from back-propagating to ResNet-18 after 120 epochs, with λ set to 1.Source: §D.1 · p. 11

D.2. Object Detection

  • For detection, coreset features concatenate pooled FCOS regression and classification features; learning-loss sampling uses five FPN maps and blocks loss gradients to the backbone.Using FPN features for coreset sampling did not perform better; localization stability sampling was implemented for VOC2012 and COCO.Source: §D.2 · p. 11

D.3. Large Scale Experiment in Object Detection

The supplied passages identify the compared-method implementation and the scopes of two result figures. They do not report the figures' outcomes.

D.3. Large Scale Experiment in Object Detection

  • The compared methods use the same implementation details as those described in D.2.Source: §D.3 · p. 11
  • Figure 7 presents results for CIFAR10 in a large-scale active-learning setting.Source: Figure 7 · p. 12
  • Figure 8 presents results for COCO with Faster R-CNN.Source: Figure 8 · p. 12

E. Additional Experiments

The additional experiments examine large-scale CIFAR10 classification and COCO object detection. They report improved CIFAR10 performance through step 5 and fewer bounding-box annotations than random sampling for Faster R-CNN at a specified AP.

E.1. Image Classification

  • On CIFAR10, the experiment starts with 5,000 labeled images and adds 5,000 per step for 10 active-learning steps.Source: §E.1 · p. 12
  • Before step 6, ISAL outperforms all compared methods on large-scale CIFAR10, and ISAL and coreset sampling both perform best at step 5.Figure 7 shows the large-scale CIFAR10 results.Source: §E.1 · p. 12
  • After step 5, CIFAR10 model performance deteriorates as the labeled set grows, so the authors say continuing the iteration is unnecessary in this setting.They interpret this as indicating that some CIFAR10 training images negatively influence ResNet-18 performance on the test set.Source: §E.1 · p. 12

E.2. Object Detection

  • For COCO, images are added in batches of 1,000 until the trained model reaches 22 ± 0.3% AP, to account for differences in selected box counts.Source: §E.2 · p. 12
  • The COCO experiment uses Faster R-CNN with a ResNet-50 backbone and estimates stest from at most 500 labeled images, averaging four estimates.Influence gradients are calculated for specified RPN convolutional and RoI fully connected layers, using the validation set as reference.Source: §E.2 · p. 12
  • At 21.8 AP, ISAL uses 7.1k fewer bounding boxes than random sampling, a 10.4% annotation saving with Faster R-CNN.The authors say this supports effectiveness with both one-stage and two-stage detectors.Source: §E.2 · p. 12

F. The Experiment Results

The passages map the paper's experiment tables to classification and object-detection datasets and settings. They provide table scopes but no numeric cell results.

F. The Experiment Results

Loading 2108.09331v1…