Adaptive Learning for Dynamic Features and Noisy Labels.

Area of Science:

• Machine Learning
• Signal Processing
• Data Science

Background:

• Machine learning in dynamic environments faces challenges due to coupled changing elements and scarce training data.
• Activity recognition tasks are susceptible to sensor displacement, causing feature space shifts and label noise.
• Learning from dynamic features with noisy labels, especially with limited new noisy samples, is an understudied problem.

Purpose of the Study:

• To propose a novel two-stage algorithm, Adaptive Learning for Dynamic features and Noisy labels (ALDN), to address coupled dynamic features and noisy labels.
• To develop a method that effectively maps prior models to current stages using optimal transport.
• To provide theoretical guarantees for risk minimization in the proposed algorithms.

Main Methods:

• A two-stage algorithm, ALDN, is proposed, utilizing modified optimal transport to map previous models to current stages.
• A consistency constraint regularizer is introduced to aid noise transition matrix estimation and model training.
• Two implementations, ALDN-D (direct) and ALDN-ID (indirect), are presented for investigation.

Main Results:

• Extensive experiments demonstrate the effectiveness of the proposed ALDN algorithms.
• The algorithms successfully handle dynamic features coupled with noisy labels in data-scarce scenarios.
• Theoretical guarantees for risk minimization are provided for ALDN-D and ALDN-ID.

Conclusions:

• The ALDN algorithm offers a robust solution for machine learning in complex, open environments with noisy labels and dynamic features.
• The proposed method shows significant improvements in activity recognition and similar tasks.
• ALDN provides a valuable contribution to the field of robust machine learning under challenging data conditions.