Covariance-based metric model for cross-domain few-shot classification and learning-to-generalization

Machine learning has successfully achieved the ultimate goal to mimic intelligent human behavior in data-abundance problems but is often hampered in data scarcity problems. To tackle this, computer vision presents a low-cost solution by introducing a fundamental concept called “Few-Shot Learning (FS...

Full description

Saved in:
Bibliographic Details
Published in:Applied intelligence (Dordrecht, Netherlands) Netherlands), 2023-11, Vol.53 (22), p.27374-27391
Main Authors: Khanday, Nadeem Yousuf, Sofi, Shabir Ahmad
Format: Article
Language:English
Subjects:
Algorithms
Approximation
Artificial Intelligence
Classification
Computer Science
Computer vision
Covariance
Datasets
Hypotheses
Machine learning
Machines
Manufacturing
Mechanical Engineering
Modulation
Neural networks
Optimization
Processes
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Machine learning has successfully achieved the ultimate goal to mimic intelligent human behavior in data-abundance problems but is often hampered in data scarcity problems. To tackle this, computer vision presents a low-cost solution by introducing a fundamental concept called “Few-Shot Learning (FSL)” which helps to bridge the gap between AI, ML, and human-like learning. FSL makes learning models to generalize in low-shot regimes, hence avoiding strenuous human tasks like the collection of huge datasets, arduous labeling, or preprocessing. Data modality, domain adaptation, domain shifts, and discrepancy of the implementation details among multiple FSL algorithms are major challenges and of great interest in FSL models. Metric-based models successfully attempt to learn within tasks, however, they often fail to generalize across tasks due to the large disparity of feature distributions. In our study, we introduce a class-covariance based metric-learner that adopts Mahalanobis distance as a metric function to enhance few-shot classification performance significantly within and across domains. Feature-wise modulation layers are also introduced and integrated with a feature encoder at multiple steps to augment and diversify the distribution of image features which makes our model learning-to-generalize across domains as well. Meta-learning based optimization algorithm is adopted to fine-tune the hyper-parameters involved. Evaluation and analysis of experimental procedures compare favorably against other state-of-the-art FSL methods and justify the use of a covariance-based metric model and learned feature-wise modulation layers. Graphical abstract
ISSN:0924-669X
1573-7497
DOI:10.1007/s10489-023-04948-z