Core Concepts
Addressing the Feature-Extractor-Maladaptive problem in few-shot learning through Isolated Graph Learning and Graph Co-Training.
Abstract
Introduction to Few-Shot Learning:
FSL tackles data scarcity with base and novel sets.
Phases: pre-train feature extractor, meta-test classification.
Semi-Supervised Few-Shot Learning:
SSFSL corrects data distribution with unlabeled data.
Challenges of Feature-Extractor-Maladaptive (FEM) problem.
Isolated Graph Learning (IGL):
Encodes samples to graph space for label prediction.
Flexibility in training and testing procedures.
Graph Co-Training (GCT):
Multi-modal fusion approach to tackle FEM.
Exploits unlabeled samples for classifier enhancement.
Methodology Overview:
Graph Structure Encoder, IGL, GCT explained.
Comparison of IGL and GL:
IGL's advantage in independent training and testing.
Comprehensive Analysis of GCT:
Positive influences: IGL, multi-modal features, co-training strategy.
Conclusion:
Addressing FEM in few-shot learning effectively with IGL and GCT.
Stats
"Aiming at the Feature-Extractor-Maladaptive (FEM) problem in semi-supervised few-shot learning..."
"The categories contained in the base set are entirely different from those in the novel set..."
Quotes
"IGL can weaken the negative influence of noise from the feature representation perspective."
"GCT is capable of further addressing the FEM problem from a multi-modal fusion perspective."