The Omniglot Challenge: A 3-Year Progress Report

New larger datasets contributed to the resurgence of interest in neural networks, Three years ago, we released the Omniglot dataset for such as the ImageNet dataset for objection recognition developing more humanlike learning algorithms. Omniglot that provides 1,000 classes with about 1,200 examples is a one-shot learning challenge, inspired by how each (Deng et al., 2009; Krizhevsky et al., 2012) and the people can learn a new concept from just one or a few Atari benchmark that typically provides 900 hours of experience examples. Along with the dataset, we proposed a suite playing each game (Bellemare et al., 2013; Mnih of five challenge tasks and a computational model based et al., 2015). These datasets opened important new lines on probabilistic program induction that addresses them. of work, but they offer far more experience than human The computational model, although powerful, was not learners require. People can learn a new concept from just meant to be the final word on Omniglot; we hoped that one or a handful of examples, and then use this concept the machine learning community would both build on for a range of tasks beyond recognition (Figure 1). Similarly, our work and develop novel approaches to tackling the people can learn a new Atari game in minutes rather challenge. In the time since, we have been pleased to than hundreds of hours, and then generalize to game variants see the wide adoption of Omniglot and notable technical beyond those that were trained (Lake et al., 2017).