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

Following the metric-based between human and artificial intelligence is the ability to methods, MetaDelta firstly adopts pretrained convolutional learn from small samples, e.g., learning to recognize objects networks as backbones to project images to latent vectors from limited examples. Inspired by human's ability of learning and trains the backbones with linear classifiers in a nonepisodic to learn from experience, meta-learning (Vanschoren way on the training classes. To improve the system's 2018) aims to transfer the generic experience learned from generalization capacity to any unknown datasets under multiple tasks of limited data to efficiently complete new time and memory budgets, we employ multiple metalearning tasks. As one of the most successful applications for metalearning, models with multi-processing, while managing the few-shot learning targets at learning from a limited time and resources with a central controller in the main number of labeled examples, which has become a research process at the same time. Moreover, we implement a latefusion trend recently. Few-shot image classification is a task where meta-ensemble mechanism to improve the generalization the classifier must learn to accommodate new classes not ability by taking the prediction from each model seen during training with limited examples.