Abstract:Humans can solve image classification tasks by learning from a few images and reusing prior-knowledge. In Artificial Intelligence, deep-learning models have been implemented to simulate human learning and face problems with little data available, few-shot learning. Nevertheless, one crucial problem of deep-learning is the selection of architectures and initial parameters that accomplish the requirements for a specific task. Therefore, we propose a novel methodology based on the meta-learning paradigm, which reuses prior- and meta-knowledge to select the best architecture and its initial parameters to solve different few-shot image classification tasks, as humans do. Our methodology was designed to understand the knowledge flow of the meta-learning paradigm by dividing the learning into prior-models, meta-models, meta-unsupervised algorithm, and base-models. We considered 9 architectures of deep convolutional neural networks as prior-models. Also, we propose a meta-unsupervised algorithm inspired by the human-cognitive problem-solving process, which acquires knowledge by solving tasks to recommend initial parameters to solve other new tasks. Furthermore, we propose a New Task Distribution Scheme to better evaluate few-shot learning models and analyze the difficulty of the new tasks. We evaluated our meta-learning methodology by applying the NTDS to the Mini-ImageNet and Caltech-UCSD-Birds-200-2011 (CUB) datasets. The achieved Adjusted Mutual Information average scores are 0.955 (zero-shot) and 0.957 (five-shot) for Mini-ImageNet, and 0.822 (zero-shot) and 0.848 (five-shot) for CUB. Moreover, we demonstrated the applicability of our meta-learning methodology in a real-world scenario using a new copro-parasite dataset. The performances of our meta-learning methodology are competitive regarding the state-of-the-art unsupervised few-shot learning models.

Lessons Learned from the NeurIPS 2021 MetaDL Challenge: Backbone Fine-Tuning Without Episodic Meta-Learning Dominates for Few-Shot Learning Image Classification

MetaDelta: A Meta-Learning System for Few-shot Image Classification

Meta-Learning Adversarial Domain Adaptation Network for Few-Shot Text Classification.

Meta-Transfer Learning Through Hard Tasks

Meta-learning methodology based on meta-unsupervised algorithm for meta-model selection to solve few-shot base-tasks

When Low Resource NLP Meets Unsupervised Language Model: Meta-Pretraining then Meta-Learning for Few-Shot Text Classification (Student Abstract)

Beyond Simple Meta-Learning: Multi-Purpose Models for Multi-Domain, Active and Continual Few-Shot Learning

Winning Solutions and Post-Challenge Analyses of the ChaLearn AutoDL Challenge 2019

Deep Meta-Learning: Learning to Learn in the Concept Space

Is the Meta-Learning Idea Able to Improve the Generalization of Deep Neural Networks on the Standard Supervised Learning?

Revisiting Unsupervised Meta-Learning via the Characteristics of Few-Shot Tasks

MetaComp: Learning to Adapt for Online Depth Completion

Boosting Meta-Training with Base Class Information for Few-Shot Learning

Improve Meta-learning for Few-Shot Text Classification with All You Can Acquire from the Tasks

Designing Neural Speaker Embeddings with Meta Learning

Few-Shot Classification Based on Sparse Dictionary Meta-Learning

Meta-learning Approaches for Few-Shot Learning: A Survey of Recent Advances

Task Attended Meta-Learning for Few-Shot Learning

Incremental Meta-Learning via Indirect Discriminant Alignment

Meta Learning for Few-Shot One-class Classification