MetAdapt: Meta-learned task-adaptive architecture for few-shot classification

• We propose a feed-forward model rewiring its architecture according to few-shot task. meta-learned NAS-cell controllers for predictive rewiring. show performance preserving way of pruning feed-forward-adaptive NAS-cells. Using the proposed approach, we obtain strong results on two popular FSC benchmarks. Recently, great progress has been made in field Few-Shot Learning (FSL). While many different methods have proposed, one key factors leading higher FSL is surprisingly simple. It backbone network used embed images tasks. first works resorted small architectures with just few convolution layers, recent that large pre-trained training portion datasets produce features are more easily transferable novel tasks, thus attaining significant gains using them. Despite these observations, little no work done towards finding right FSL. In this paper MetAdapt not only meta-searches an optimized Network Architecture Search (NAS), but also can adaptively ‘re-wire’ itself predicting better given approach observe benchmarks: mini ImageNet and FC100.