Few-shot and meta-learning methods for image understanding: a survey

Abstract State-of-the-art deep learning systems (e.g., ImageNet image classification) typically require very large training sets to achieve high accuracies. Therefore, one of the grand challenges is called few-shot where only a few samples are required for good performance. In this survey, we illuminate key paradigms in meta-learning. These meta-learning methods, by simulating tasks which will be presented at inference through episodic training, can effectively employ previous prior knowledge guide new tasks. paper, provide comprehensive overview and insights into approaches categorize them three branches according their technical characteristics, namely metric-based, model-based optimization-based Due major importance evaluation process, also present an current widely used benchmarks, as well performances recent methods on these datasets. Based over 200 papers conclude with future directions