Deep learning’s impressive performance on complex classification applications has made deep neural networks the standard tool for many applications, such as image classification, document summarization, and speaker identification. This impressive performance is commonly achieved with supervised learning in which deep neural networks train on very large labeled datasets. Moreover, benchmark datasets in these topic areas typically contain thousands to millions of labeled images. Consequently, research in deep learning had exploded in the last decade following the creation of large labeled benchmark datasets such as ImageNet.

However, collecting, cleaning, and labeling training, validation, and test data for potential new applications of deep learning is hard, even for those applications where the raw data might be plentiful. Manually labeling thousands of data samples is often impractical for new applications because labeling is labor-intensive and often can only be correctly classified by a limited number of experts, such as in medical, defense, and scientific applications. Furthermore, care must be taken in labeling training samples or else label noise can interfere with the model's accuracy and generalization capabilities.

On the other hand, humans learn to recognize new object types and words (both written and auditory) with one or a few examples. Designing deep neural networks that can learn with limited labeled examples is an open and active research area. Novel deep learning systems that learn with few labeled examples to recognize new classes and to adapt continuously to changing scenarios would greatly reduce the effort required to develop deep learning systems for new applications and expand the lifetime of production systems.

Furthermore, significant progress has been made in semi-supervised and unsupervised learning. Recent approaches are achieving performances that are comparable to fully supervised training on large labeled datasets. These new approaches lower the barrier for applying deep learning to new applications.

This Research Topic focuses on learning with fewer labels for deep neural networks. Application areas can include vision, language processing, multimedia, and speech (i.e., machine language translation). Multi-modal tasks come with their own set of challenges and are of particular interest. The topics of interest include (but are not limited to) the following areas:

• Self-supervised and unsupervised learning methods
• Semi-supervised learning methods
• Weakly-supervised learning methods
• New methods for few-/zero-shot learning
• Meta-learning methods
• New applications in vision, text, and speech
• Multi-modal learning with limited labels (i.e., VQA, fusion)
• Life-long/continual/incremental learning methods
• Novel domain adaptation methods
• Theoretical understanding of learning with limited labels
• Biologically inspired learning with limited labels
• Novel evaluation metrics

Submit your idea

Impact

Frontiers in Artificial Intelligence

Participating Journals

Impact