Deep Metric Learning with BIER: Boosting Independent Embeddings Robustly.

Area of Science:

• Computer Science
• Machine Learning
• Deep Learning

Background:

• Deep neural networks learn similarity functions for image pairs, but embeddings often have highly correlated activations.
• This correlation can limit the robustness and retrieval accuracy of learned embeddings.

Purpose of the Study:

• To improve the robustness and retrieval accuracy of deep network embeddings by reducing activation correlation.
• To leverage larger embedding sizes more effectively through ensemble techniques.

Main Methods:

• Dividing the final embedding layer of a deep network into an embedding ensemble.
• Formulating the ensemble training as an online gradient boosting problem with reweighted samples.
• Proposing two novel loss functions to increase diversity within the embedding ensemble.

Main Results:

• Significantly reduced correlation within the embedding layer.
• Increased retrieval accuracy across multiple benchmark image retrieval datasets (CUB-200-2011, Cars-196, etc.).
• Demonstrated state-of-the-art performance improvements over existing methods.

Conclusions:

• Ensemble methods, by dividing deep networks into smaller, diverse components, can effectively reduce overfitting.
• The proposed approach enhances embedding robustness and retrieval accuracy without increasing test-time complexity.