Combining various training and adaptation algorithms for ensemble few-shot classification | JoVE Visualize

Area of Science:

Computer Science
Artificial Intelligence
Machine Learning

Background:

Few-Shot Classification (FSC) methods train deep neural networks (DNNs) on base datasets and adapt them to new tasks with limited data.
Single FSC models often suffer from high variance and low confidence, motivating the development of ensemble FSC methods.
Existing ensemble FSC methods face challenges due to limited labeled data and high computational costs of DNNs.

Purpose of the Study:

To propose a novel ensemble method for Few-Shot Classification (FSC) that mitigates challenges of data scarcity and computational cost.
To generate diverse FSC models efficiently by reusing training phases and employing a unique pseudo-labeling strategy.
To improve the performance and confidence of FSC models through ensemble learning and noise reduction.

Main Methods:

The proposed method generates multiple FSC models by combining various training and adaptation algorithms.
Training phases are reused to significantly reduce learning costs and enhance base model diversity.
A novel "one-vs-others" pseudo-labeling strategy, using majority votes from other models, minimizes reliance on labeled data and reduces label noise and confirmation bias.

Main Results:

The ensemble FSC method significantly reduces learning costs while generating diverse base models.
The "one-vs-others" pseudo-labeling strategy effectively minimizes pseudo-label noise and confirmation bias compared to self-training methods.
Extensive experiments on miniImageNet, tieredImageNet, and CUB datasets demonstrate superior performance over state-of-the-art FSC methods, with notable improvements in base model performance.

Conclusions:

The novel ensemble FSC method effectively addresses data scarcity and computational challenges in few-shot learning.
The proposed approach enhances model diversity and reduces noise, leading to state-of-the-art performance.
The method shows significant potential for improving the reliability and accuracy of deep neural networks in low-data regimes.