Adaptive Neighborhood Metric Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Mining

Background:

• Metric learning algorithms often struggle with inseparable samples, leading to training instability.
• Existing methods for handling inseparable samples can introduce undesirable side-effects.
• Informative sample mining is crucial for effective metric learning.

Purpose of the Study:

• To propose a novel distance metric learning algorithm, Adaptive Neighborhood Metric Learning (ANML), to overcome the inseparable sample problem.
• To develop a continuous formulation for ANML using a log-exp mean function for efficient gradient-based optimization.
• To demonstrate the versatility of ANML in learning both linear and deep embeddings.

Main Methods:

• ANML adaptively identifies inseparable similar and dissimilar samples using two thresholds during training.
• A log-exp mean function is introduced to create a continuous surrogate for the non-continuous ANML.
• The algorithm is optimized using gradient descent methods.

Main Results:

• ANML effectively removes inseparable samples and learns metric parameters simultaneously.
• ANML unifies several existing metric learning algorithms (LMNN, NCA, Triplet loss, etc.) as special cases.
• The proposed log-exp mean function offers a new perspective on deep metric learning methods.

Conclusions:

• ANML provides an effective and stable solution to the inseparable sample problem in metric learning.
• The proposed method demonstrates broad applicability across linear and deep embedding learning.
• Experimental results validate the superior performance of ANML.