Incremental Embedding Learning With Disentangled Representation Translation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Neural networks struggle with catastrophic forgetting in sequential task learning.
• Existing incremental learning methods primarily address classification networks, neglecting embedding networks.
• Embedding networks, crucial for metric learning, also face catastrophic forgetting, destroying latent feature relationships.

Purpose of the Study:

• To propose a novel incremental learning method for embedding networks to overcome catastrophic forgetting.
• To preserve discriminative and representative latent features without reusing past task data.
• To avoid perturbing task-related information during continual learning.

Main Methods:

• Introduced the disentangled representation translation (DRT) method for embedding networks.
• Employed a mask-guided module to adaptively manage latent feature information.
• Integrated an optional regularization item for enhanced incremental learning performance.

Main Results:

• The proposed DRT method effectively alleviates catastrophic forgetting in embedding networks.
• Experiments on four datasets demonstrate the method's efficacy.
• The approach successfully preserves discriminative class-disentangled features without sample replay.

Conclusions:

• The DRT method offers a robust solution for continual learning in embedding networks.
• It successfully tackles the challenge of preserving relationships between latent features and prototypes across tasks.
• This work advances incremental learning for metric learning applications by mitigating catastrophic forgetting.