Is Meta-Learning Always Necessary?: A Practical ML Framework Solving Novel Tasks at Large-Scale Car Sharing Platform

Hyunhee Chung, Kyung Ho Park; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2023, pp. 421-429

While the deep neural networks achieved superior performance in various tasks under the supervised regime, the ML practitioners in the real world frequently encounter a novel task that cannot acquire the labeled dataset shortly. Even if they have become available in acquiring the target samples from the unlabeled dataset, conventional labeling procedures require the practitioners to invest in resource consumption. Pursuing an effective solution to these problems, our study proposes a practical ML framework that efficiently enables practitioners to solve novel tasks. Our ML framework consists of two solutions consisting of early and mature stages. First, the early stage solution lets the practitioners solve the novel task under the few-shot classification setting. Second, the mature stage solution enhances the labeling efficiency by retrieving samples that seem relevant to the target. Upon these solutions, acquiring a qualified representation power is the most important job. Under the public benchmark datasets and image recognition tasks in a large-scale car-sharing platform, we examined that the paradigm of supervised learning, surprisingly not meta-learning, produces the most beneficial representation power to solve novel tasks. We further scrutinized the supremacy of supervised representation derives from broader, nourished high-level representations in the neural networks. We highly expect our analyses can be a concrete benchmark to the ML practitioners who solve novel tasks in their domain.