Co-Training with Active Contrastive Learning and Meta-Pseudo-Labeling on 2D Projections for Deep Semi-Supervised Learning

TL;DR

This paper tackles the challenge of training deep CNNs with limited labeled data by introducing active-DeepFA, a two-network co-training framework that blends supervised contrastive learning, meta-pseudo-labeling, and active learning within a DeepFA-based paradigm. It leverages label propagation on 2D projections of deep features via OPFSemi and periodically expands the labeled set by querying an oracle for the most informative samples, thereby reducing annotation effort. The approach achieves state-of-the-art results on three challenging biological image datasets with only 5% labeled data and remains competitive with as little as 3% labeled, demonstrating strong data efficiency and generalization. The combination of dual-network cross-training, proactive sample labeling, and 2D feature-space propagation offers a practical route to high-performance SSL for non-pretrained architectures in domains with scarce annotations and abundant unlabeled data.

Abstract

A major challenge that prevents the training of DL models is the limited availability of accurately labeled data. This shortcoming is highlighted in areas where data annotation becomes a time-consuming and error-prone task. In this regard, SSL tackles this challenge by capitalizing on scarce labeled and abundant unlabeled data; however, SoTA methods typically depend on pre-trained features and large validation sets to learn effective representations for classification tasks. In addition, the reduced set of labeled data is often randomly sampled, neglecting the selection of more informative samples. Here, we present active-DeepFA, a method that effectively combines CL, teacher-student-based meta-pseudo-labeling and AL to train non-pretrained CNN architectures for image classification in scenarios of scarcity of labeled and abundance of unlabeled data. It integrates DeepFA into a co-training setup that implements two cooperative networks to mitigate confirmation bias from pseudo-labels. The method starts with a reduced set of labeled samples by warming up the networks with supervised CL. Afterward and at regular epoch intervals, label propagation is performed on the 2D projections of the networks' deep features. Next, the most reliable pseudo-labels are exchanged between networks in a cross-training fashion, while the most meaningful samples are annotated and added into the labeled set. The networks independently minimize an objective loss function comprising supervised contrastive, supervised and semi-supervised loss components, enhancing the representations towards image classification. Our approach is evaluated on three challenging biological image datasets using only 5% of labeled samples, improving baselines and outperforming six other SoTA methods. In addition, it reduces annotation effort by achieving comparable results to those of its counterparts with only 3% of labeled data.

Figures (6)

• Figure 1: The networks' architecture consisting of an encoder $E$ and two projection heads $P_{x}$ and $P_{c}$.
• Figure 2: The weight-sharing Siamese architecture adopted by each network during contrastive learning.
• Figure 3: Overview of the active-DeepFA method. For each network, the output deep features of $E({\cal Z})$ are projected into 2D using t-SNE and labeling propagation is executed by OPFSemi on the 2D space. Next, the sets ${\cal Z}_{U}^{pl(i)},{{\cal Z}_{A}^{(i)}}$ are constructed for $N_{i},\,i\in\{1,2\}$. The set ${\cal Z}'_{A}$ is created by joining the sets ${\cal Z}_{A}^{(1)},{\cal Z}_{A}^{(2)}$ and picking the $k_{\text{active}}$ samples whose pseudo-labels have the lowest $v$ value. Then, the elements of ${\cal Z}'_{A}$ are sent to the oracle to be labeled and added to ${\cal Z}_{L}$. Afterward, the losses ${\cal L}_{S}^{cl},{\cal L}_{S}$ are computed on ${\cal Z}_{L}$, while ${\cal L}_{U}^{ssl}$ is calculated on ${\cal Z}_{U}^{pl(2)}$ for $N_{1}$ and vice versa for $N_{2}$ in a cross-training manner.
• Figure 4: Images from the Helminth larvae, Helminth eggs and Protozoan cysts datasets. The first row shows images of parasites, while the second row exhibits images of impurities.
• Figure 5: Heatmap of the labeling confidence values after labeling propagation on the 2D projection of the Helminth eggs's deep features via OPFSemi. Left: the prototypes from $\phi({\cal Z}_{L})$ are non-gray-colored with a different color per class, while samples from $\phi({\cal Z}_{U})$ are gray-colored. Middle: result of label propagation from $\phi({\cal Z}_{L})$ to $\phi({\cal Z}_{U})$. Right: heatmap of the labeling confidence values, where blue-colored samples with the lowest labeling confidence values lie close to decision boundary regions.