PaperlayerView exact revision on arXiv

Source-linked AI summary

Sequential Graph Convolutional Network for Active Learning

Razvan Caramalau, Binod Bhattarai, Tae-Kyun Kim

arXiv:2006.10219v3cs.CVcs.LG

TL;DR

Active learning must reduce costly annotation while selecting representative data across varied downstream tasks. The paper proposes a sequential, task-agnostic GCN sampler that models labelled–unlabelled relationships and iteratively queries samples. Across six benchmarks, the framework reports higher accuracy than existing sampling methods and state-of-the-art results.

  • Problem

    Active learning seeks representative samples because deep-learning annotation is time-consuming, expensive, and difficult to scale across tasks.

  • Method

    A sequential GCN builds a similarity graph from learner features, distinguishes labelled from unlabelled nodes, and uses confidence scores with uncertainty or CoreSet-style sampling before retraining.

  • Results

    The framework was more accurate than existing methods across six benchmarks and produced state-of-the-art results on the evaluated applications.

  • Takeaways & Limitations

    The sampling mechanism supports task-agnostic integration beyond image classification, including regression and synthetic-data selection.

  • Takeaways & Limitations

    Adapting the method to another task requires modifying the learner while leaving the rest of the pipeline unchanged.

Abstract

from arXiv · show

We propose a novel pool-based Active Learning framework constructed on a sequential Graph Convolution Network (GCN). Each image's feature from a pool of data represents a node in the graph and the edges encode their similarities. With a small number of randomly sampled images as seed labelled examples, we learn the parameters of the graph to distinguish labelled vs unlabelled nodes by minimising the binary cross-entropy loss. GCN performs message-passing operations between the nodes, and hence, induces similar representations of the strongly associated nodes. We exploit these characteristics of GCN to select the unlabelled examples which are sufficiently different from labelled ones. To this end, we utilise the graph node embeddings and their confidence scores and adapt sampling techniques such as CoreSet and uncertainty-based methods to query the nodes. We flip the label of newly queried nodes from unlabelled to labelled, re-train the learner to optimise the downstream task and the graph to minimise its modified objective. We continue this process within a fixed budget. We evaluate our method on 6 different benchmarks:4 real image classification, 1 depth-based hand pose estimation and 1 synthetic RGB image classification datasets. Our method outperforms several competitive baselines such as VAAL, Learning Loss, CoreSet and attains the new state-of-the-art performance on multiple applications The implementations can be found here: https://github.com/razvancaramalau/Sequential-GCN-for-Active-Learning

1. Introduction

The paper presents a task-agnostic sequential GCN framework that selects informative samples by modeling relationships between labelled and unlabelled images. It combines graph-based embeddings and confidence scores with uncertainty and CoreSet-style sampling, then iteratively retrains the learner and sampler.

  • Motivation: Active learning addresses expensive expert annotation and limited data representativeness by selecting meaningful samples within a fixed labelling budget.The framework separates the learner, sampler, and annotator components.
  • Proposed method: The pipeline trains the learner, builds the graph, optimizes the GCN, queries unlabelled examples, annotates them, and repeats the cycle.Figure 1 organizes the method into five sequential phases.
  • Motivation: Previous task-agnostic methods such as VAAL and Learning Loss do not exploit correlations between labelled and unlabelled images through message passing.The paper identifies this relational modeling gap as a limitation of those approaches.
  • Proposed method: The proposed sampler constructs a similarity graph from learner features, applies GCN layers to distinguish labelled from unlabelled nodes, and uses node embeddings and confidence scores for selection.The learner and sampler are separated, making the sampling framework task-agnostic.
  • Evaluation: The framework was evaluated on four real image-classification benchmarks, one depth-based 3D hand-pose dataset, and one synthetic image-classification benchmark.It was compared with competitive methods including CoreSet, VAAL, and Learning Loss.

2. Related Works

Prior active-learning work includes model-based samplers and GCN-based methods, but the paper distinguishes its approach through relational message passing, a separate task-agnostic sampler, and sequential training.

  • Model-based methods: Model-based active-learning methods train a separate model to select representative data, including loss prediction and VAE-based discrimination of labelled from unlabelled samples.The paper positions its method within this model-based category.
  • Model-based methods: The proposed method differs from VAE-based sampling by exploiting relationships between examples through GCN message passing.This provides a relational alternative to latent-space discrimination.
  • GCN-based methods: Existing GCN active-learning methods are commonly applied to graph-structured datasets such as Cora, CiteSeer, and PubMed.The paper contrasts those settings with its image-based sampling framework.
  • GCN-based methods: The paper separates the learner and sampler, whereas cited GCN methods derive selection mechanisms under a graph-learner assumption.It also identifies sequential training as a distinction from those methods.

3. Method

The method uses a sequential GCN to support task-agnostic pool-based active learning by representing labelled and unlabelled examples as connected feature nodes. It repeatedly trains the learner and graph, then queries examples using GCN confidence or graph-feature distance within a fixed budget.

  • Active learning setup: The sampler operates in a pool-based loop that starts from randomly labelled examples and seeks to minimize annotation stages while tracking performance at each query.The learner is retrained after new labels are acquired, while the sampler selects representative unlabelled examples.
  • Graph construction: Learner features initialize graph nodes, while adjacency relations encode similarities among labelled and unlabelled images.The graph propagates similarities between nodes in the learner’s feature space.
  • Graph convolutional network: A two-layer GCN maps graph inputs to node outputs whose sigmoid values distinguish labelled from unlabelled examples.The architecture uses ReLU after the first layer and a sigmoid output ranging from 0 to 1.
  • UncertainGCN: UncertainGCN queries unlabelled samples according to their GCN confidence scores, selecting examples whose scores are closest to a configurable margin.Samples with margin scores closer to 0 are treated as more uncertain and challenging to discriminate.
  • CoreGCN: CoreGCN adapts CoreSet sampling by measuring Euclidean distances between labelled and unlabelled graph features.This transfers geometric selection from the learner feature space to GCN-induced representations.
  • Sequential procedure: The framework is claimed to be task-agnostic when the learner produces feature representations, and experiments cover classification and regression settings.The active-learning loop trains the learner, extracts features, trains the GCN, queries nodes, and adds their labels until the stopping condition is met.

4. Experiments

The experiments evaluate the proposed active-learning sampler across real and synthetic image classification and depth-based hand-pose estimation. CoreGCN and UncertainGCN generally match or outperform established sampling methods, with qualitative and ablation analyses examining their selection behavior and components.

  • Experimental Settings: The evaluation covers RGB and grayscale real-image classification, depth-image regression for 3D hand pose estimation, and synthetic-image classification.The classification benchmarks include CIFAR-10, CIFAR-100, SVHN, and FashionMNIST.
  • Classification: CoreGCN achieves 90.7% accuracy on CIFAR-10 after selecting 10,000 labelled examples, the highest reported performance among the compared literature.The comparison includes UncertainGCN and five existing methods, with CoreSet identified as the closest competitor.
  • Classification: With 40% of CIFAR-100 training data, CoreGCN reaches 69% accuracy, only 4% below training with the entire dataset.The CIFAR-100 comparison follows trends similar to CIFAR-10, while VAAL performs better than on CIFAR-10.
  • Classification: On FashionMNIST and SVHN, the proposed methods achieve at least similar performance to existing methods or outperform them.FeatProp shows consistently modest performance on these datasets, possibly because it does not generalise to the unstructured data.
  • Qualitative Comparisons: CoreGCN preserves sparsity across selection stages, maintaining message passing through uncertain areas while CoreSet samples closer to cluster centres.The resulting selection provides a stronger balance between in-distribution and out-of-distribution samples as more data becomes available.
  • Regression: On ICVL 3D hand-pose estimation, CoreGCN and UncertainGCN consistently outperform CoreSet and random sampling from the second selection stage.Performance is reported using mean squared error averaged over five trials and its standard deviation; lower values are better.
  • Sub-sampling of Synthetic Data: For synthetic expression data, UncertainGCN achieves higher accuracy with less variance than random sampling, although both methods decline after the fourth selection stage.The decline is attributed to selecting more noisy data after the useful synthetic examples have largely been selected.

5. Conclusions

The paper presents a GCN-based active-learning methodology for image classification and regression, with UncertainGCN and CoreGCN achieving state-of-the-art results across six benchmarks.

  • The proposed methodology applies Graph Convolutional Networks to active learning for image classification and regression.
  • UncertainGCN and CoreGCN are adapted sampling techniques designed for the proposed GCN-based framework.
  • The experiments report state-of-the-art results on six benchmarks.
  • Qualitative distributions indicate that the selection functions maximise informativeness within the data space.
  • The sampling mechanism permits integration into other learning tasks and supports combining optimised selection criteria with a GCN sampler.

A. Datasets

The evaluation uses diverse image-classification, hand-pose-estimation, and synthetic face-expression datasets with varied modalities, resolutions, class structures, and dataset sizes.

  • The image-classification benchmarks are CIFAR-10, CIFAR-100, FashionMNIST, and SVHN.They vary in modality, class granularity, image resolution, and dataset size.
  • CIFAR-10 and CIFAR-100 contain 50,000 training and 10,000 testing images, while CIFAR-10 has 10 classes and CIFAR-100 has 100 classes.
  • FashionMNIST is grayscale with 60,000 training and 10,000 testing images, whereas the other classification datasets are RGB.
  • The ICVL benchmark evaluates depth-based 3D hand pose estimation using 16,004 training and 1,600 testing images.The images contain varied articulation and hand positions and are pre-processed to 128x128 resolution.
  • RaFD contains eight facial expressions and uses StarGAN-generated images as the unlabelled set.The experiment uses 7,200 training images, 840 test images, and 57,600 generated unlabelled images.

B. Experiments

Experiments evaluate the proposed samplers across imbalanced classification, learner-architecture changes, hyper-parameter settings, and qualitative sample selection, with reported strengths and sensitivity patterns.

  • CIFAR-10 imbalanced dataset: 80.05% mean average accuracy at 10,000 labelled samples gives UncertainGCN a 2% advantage over VAAL and Learning Loss on imbalanced CIFAR-10.
  • CIFAR-10 imbalanced dataset: 84.5% top performance is achieved by CoreGCN together with CoreSet on imbalanced CIFAR-10.
  • Ablation study - GCN parameter search: Increasing GCN dropout from 0.3 to 0.5 or 0.8 produced poorer selection, while hidden sizes of 256 and 512 did not affect UncertainGCN on CIFAR-10.
  • VGG-11 learner for CIFAR-10 image classification for 3 selection stages: Changing the learner to VGG-11 left the proposed methods robust and enabled them to surpass the compared state-of-the-art methods at an early sampling stage.The GCN settings were unchanged in this experiment.
  • Hyper-parameters Study: Performance improved as the uncertainty margin decreased from 0.4 to 0.1 and then remained stable, while changes in λ produced a smooth performance drift.
  • Extended qualitative analysis on the AL method: Qualitative analysis found difficult CIFAR-10 cat examples among selected samples and noisy RaFD images among those left unlabelled.ICVL selections showed closer and easier hand articulations than the initial labelled set.
Loading 2006.10219v3…