Contrastive Learning and Iterative Meta-Pseudo-Labeling on 2D Projections for Deep Semi-Supervised Learning

Autor(es)

Aparco-Cardenas, David ; Gomes, Jancarlo E. ; Falcao, Alexandre X. ; de Rezende, Pedro J.

• Tipo de documento: Artigo Científico
• Fonte: 2024 37TH SIBGRAPI CONFERENCE ON GRAPHICS, PATTERNS AND IMAGES, SIBGRAPI 2024; v. N/A, p. 6-pg., 2024-01-01.

Resumo

The scarcity of accurately labeled data critically hampers the usage of deep learning models. This issue is highlighted in areas (e.g., biological sciences) where data annotation results in an expert-demanding, labor-intensive and error-prone task. While state-of-the-art semi-supervised approaches have proven effective in circumventing this limitation, their reliance on pre-trained architectures and large validation sets to deliver effective solutions still poses a challenge. In this work we introduce an iterative contrastive-based meta-pseudo-labeling method for training non-pre-trained custom CNN architectures for image classification in conditions of limited labeled and abundant unlabeled data, with no dependency on a validation set. It generates multiple models across a few iterations, which are in turn exploited in an ensemble manner to label the unlabeled data and train a final classifier. Our approach starts by capitalizing on contrastive learning to enhance the representation ability of two collaborative networks while eliminating the need of pre-trained architectures. Then, during each iteration, the networks are trained within a teacher-student based cross-training setup, where OPFSemi (teacher) propagates labels from labeled to unlabeled on the non-linear 2D latent space projections of each network's (student) deep features; afterward, the pseudo-labels with the highest top 10% confidence, per class, are picked to fine-tune the other network in a cross-training manner, jointly mitigating confirmation bias and overfitting while improving the generalization ability of the networks as iterations evolve. Our method is evaluated on three challenging biological image datasets with only 5% of labeled samples, demonstrating its effectiveness and robustness when compared to two direct baselines and six state-of-the-art methods from three different semi-supervised learning paradigms. (AU)

• Processo FAPESP: 14/12236-1 - AnImaLS: Anotação de Imagem em Larga Escala: o que máquinas e especialistas podem aprender interagindo?
• Beneficiário: Alexandre Xavier Falcão
• Modalidade de apoio: Auxílio à Pesquisa - Temático
• Processo FAPESP: 13/07375-0 - CeMEAI - Centro de Ciências Matemáticas Aplicadas à Indústria
• Beneficiário: Francisco Louzada Neto
• Modalidade de apoio: Auxílio à Pesquisa - Centros de Pesquisa, Inovação e Difusão - CEPIDs
• Processo FAPESP: 23/14427-8 - Ciência de Dados para a Indústria Inteligente (CDII)
• Beneficiário: José Alberto Cuminato
• Modalidade de apoio: Auxílio à Pesquisa - Programa Centros de Pesquisa em Engenharia