Few-Shot Partial Multi-View Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Real-world data often presents multiple views, crucial for representative information.
• Data scarcity and missing views are common issues due to collection and preprocessing limitations.
• These challenges hinder effective pattern classification, especially when co-occurring.

Purpose of the Study:

• Introduce a new task: few-shot partial multi-view learning.
• Address the dual challenges of data scarcity and missing views in low-data regimes.
• Propose a method to overcome the negative impact of view missing under data scarcity.

Main Methods:

• Propose a unified Gaussian dense-anchoring method.
• Learn unified dense anchors for limited partial multi-view data.
• Map data into a unified dense representation space to mitigate data scarcity and view missing.

Main Results:

• Extensive experiments validate the proposed method's effectiveness.
• Demonstrated performance on diverse datasets including Cub-googlenet-doc2vec, Handwritten, Caltech102, Scene15, Animal, ORL, tieredImagenet, and Birds-200-2011.
• The method successfully alleviates the influence of data scarcity and view missing.

Conclusions:

• The proposed few-shot partial multi-view learning task and method offer a solution for challenging data conditions.
• The unified Gaussian dense-anchoring approach effectively handles both data scarcity and missing views.
• This work aims to draw community attention to the simultaneous challenge of view missing and data scarcity.