Proxy-AN loss for deep metric learning | JoVE Visualize

Area of Science:

Computer Science
Machine Learning
Artificial Intelligence

Background:

Deep metric learning utilizes proxy representations to approximate class distributions, aiming to simplify training and speed up convergence.
Existing proxy-based losses are categorized as sample-centric or proxy-centric, each with limitations in balancing sample and proxy representation quality.
Sub-optimal embedding spaces arise from the neglect of proxy fidelity in sample-centric losses and lack of sample discriminability in proxy-centric losses.

Purpose of the Study:

To introduce a novel loss function, Proxy-Anchor-Negative (Proxy-AN), that reconciles the divergent focuses of sample-centric and proxy-centric losses.
To combine the advantages of both approaches to achieve a holistic enhancement in representation quality for both proxies and samples.
To facilitate the learning of more discriminative metrics in deep metric learning.

Main Methods:

The Proxy-AN loss employs a proxy-centric approach for positive pairs to enhance intra-class compactness by aligning anchor proxies with positive samples.
For negative pairs, a sample-centric approach is utilized to improve sample discriminability by distancing samples from negative proxies.
This synergistic strategy ensures balanced improvements in both proxy and sample representations.

Main Results:

Extensive experiments on mainstream image retrieval benchmark datasets show substantial improvements over leading metric learning algorithms.
The method demonstrates superior performance in diverse scenarios, including partial training data and class imbalance settings.
Proxy-AN loss effectively enhances both proxy fidelity and sample discriminability, leading to improved embedding spaces.

Conclusions:

The proposed Proxy-AN loss effectively balances sample-centric and proxy-centric strategies in deep metric learning.
This novel approach leads to significant performance gains in image retrieval tasks, outperforming existing state-of-the-art methods.
Proxy-AN loss exhibits robustness and superior performance across various data conditions, including class imbalance and partial data settings.