Meta-Learning Initializations for Image SegmentationDownload PDF

25 Sep 2019 (modified: 24 Dec 2019)ICLR 2020 Conference Blind SubmissionReaders: Everyone
  • Original Pdf: pdf
  • TL;DR: We show that model agnostic meta-learning extends to the high dimensionality and dense prediction of image segmentation.
  • Abstract: While meta-learning approaches that utilize neural network representations have made progress in few-shot image classification, reinforcement learning, and, more recently, image semantic segmentation, the training algorithms and model architectures have become increasingly specialized to the few-shot domain. A natural question that arises is how to develop learning systems that scale from few-shot to many-shot settings while yielding human level performance in both. One scalable potential approach that does not require ensembling many models nor the computational costs of relation networks, is to meta-learn an initialization. In this work, we study first-order meta-learning of initializations for deep neural networks that must produce dense, structured predictions given an arbitrary amount of train- ing data for a new task. Our primary contributions include (1), an extension and experimental analysis of first-order model agnostic meta-learning algorithms (including FOMAML and Reptile) to image segmentation, (2) a formalization of the generalization error of episodic meta-learning algorithms, which we leverage to decrease error on unseen tasks, (3) a novel neural network architecture built for parameter efficiency which we call EfficientLab, and (4) an empirical study of how meta-learned initializations compare to ImageNet initializations as the training set size increases. We show that meta-learned initializations for image segmentation smoothly transition from canonical few-shot learning problems to larger datasets, outperforming random and ImageNet-trained initializations. Finally, we show both theoretically and empirically that a key limitation of MAML-type algorithms is that when adapting to new tasks, a single update procedure is used that is not conditioned on the data. We find that our network, with an empirically estimated optimal update procedure yields state of the art results on the FSS-1000 dataset, while only requiring one forward pass through a single model at evaluation time.
  • Keywords: meta-learning, image segmentation
12 Replies