Learning to Explore Sample Relationships.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models often struggle with data scarcity in real-world applications.
• Existing methods for handling data scarcity typically explore sample relationships simplistically.
• These methods focus on either the input data or the loss function.

Purpose of the Study:

• To introduce a novel module, BatchFormer, that enables deep neural networks to learn sample relationships effectively.
• To generalize BatchFormer for dense prediction tasks (BatchFormerV2) and address train-test inconsistencies.
• To improve deep learning performance in data-scarce scenarios across diverse visual recognition tasks.

Main Methods:

• Proposed BatchFormerV1 module to equip neural networks with learnable sample relationship exploration capabilities.
• Introduced BatchFormerV2, generalizing the module for pixel-/patch-level dense representations.
• Devised a two-stream training pipeline to resolve train-test inconsistencies, removing BatchFormerV2 during inference.

Main Results:

• BatchFormer enables data collaboration, allowing head-class samples to aid tail-class learning.
• The plug-and-play module incurs no extra inference cost.
• Evaluated on over ten datasets across various data scarcity settings and visual recognition tasks.

Conclusions:

• BatchFormer significantly enhances deep learning model performance, particularly in data-scarce environments.
• The module's adaptability to different tasks and its efficiency make it a valuable contribution.
• Addresses fundamental limitations in current deep learning approaches for real-world data challenges.