Preface

Min-Ling Zhang, Sheng-Jun Huang, Mingsheng LongJCST 2021 · DOI

Abstract

Modern machine learning, particularly deep learning, faces a significant bottleneck: the "data hunger" required for high performance. In many real-world scenarios, obtaining large-scale, human-annotated datasets is impractical due to high costs, time constraints, or the need for specialized expertise. This special section of the Journal of Computer Science and Technology (JCST) presents three innovative research papers that address the "small sample problem" through advanced techniques in partial label learning, source-free domain adaptation, and robust biometric recognition.

Introduction: The Small Sample Challenge

While deep learning has achieved remarkable success in computer vision and natural language processing, its efficacy is often tied to the availability of massive datasets. However, in fields like medicine or specialized industrial applications, data is inherently "small." For instance, labeling electrocardiogram (ECG) signals requires the time and expertise of a cardiologist, making large-scale annotation prohibitively expensive.

The core objective of this special section is to explore how models can be trained effectively when labeled data is scarce, noisy, or structurally complex. The editors curated a selection of research that moves beyond standard supervised learning paradigms to tackle these real-world constraints.

Methodological Highlights

From a pool of 26 submissions, three papers were selected for their significant contributions to the field:

1. Conditional Label Aware Disambiguation for Partial Label Learning

In Partial Label Learning (PLL), each training instance is associated with a set of candidate labels, only one of which is the ground truth. To address this ambiguity, the authors proposed a unified framework that simultaneously trains the model using label constraints and employs "pseudo-labeling" to identify the correct label within the noisy candidate set.

2. Sample Transport Learning for Source-Free Domain Adaptation

Source-Free Unsupervised Domain Adaptation (SF-UDA) involves transferring knowledge from a source domain to a target domain without access to the original source data—a critical requirement for privacy-sensitive applications. This method utilizes an intelligent instance selection strategy to identify and label the most "valuable" samples in the target domain, facilitating effective knowledge transport while maintaining data privacy.

3. Multi-Scale Deep Cascade Bi-Forest for ECG Recognition

Biometric recognition using ECG signals often suffers from noisy data and limited training samples. This research introduces a deep cascade framework that combines multi-scale signal coding with adaptive weighted pooling. This architecture extracts highly discriminative features, ensuring model robustness even when training examples are minimal.

Key Insights and Results

The research presented in this section offers several critical insights for the machine learning community:

  • Robust Disambiguation: Effective learning from noisy or partial labels requires a robust process to filter out incorrect candidates during training.
  • Privacy-Preserving Transfer: Knowledge can be successfully transferred between domains without exposing raw source data, overcoming significant legal and ethical hurdles in AI deployment.
  • Architectural Inductive Bias: Standard deep networks are prone to overfitting on small datasets. Specialized structures, such as deep cascades and multi-scale coding, provide the necessary inductive bias to achieve high performance with limited data.

The selected papers demonstrate state-of-the-art performance in their respective tasks, significantly reducing the reliance on large-scale human annotation while maintaining high accuracy and robustness.

Conclusion

This special section highlights the ongoing shift in machine learning research from "big data" to "smart data." By developing algorithms that can learn from small, noisy, or transferred samples, the field moves closer to deploying AI in domains where data scarcity was previously a prohibitive barrier.

Summary updated

BibTeX
@article{zhang2021preface,
  author = {Min-Ling Zhang and Sheng-Jun Huang and Mingsheng Long},
  title = {{Preface}},
  journal = {J. Comput. Sci. Technol.},
  year = {2021},
  doi = {10.1007/s11390-021-0004-1},
}
Figure