Learning from One and Only One Shot

Haizi Yu, Igor Mineyev, Lav R. Varshney, James Evans

Abstract: Humans can generalize from one or a few examples, and even from very little pre-training on similar tasks. Machine learning (ML) algorithms, however, typically require large data to either learn or pre-learn to transfer. Inspired by nativism, we directly model very basic human innate priors in abstract visual tasks like character or doodle recognition. The result is a white-box model that learns transformation-based topological similarity akin to how a human would naturally and unconsciously ``distort'' an object when first seeing it. Using the simple Nearest-Neighbor classifier in this similarity space, our model approaches human-level character recognition using only one to ten examples per class and nothing else (no pre-training). This is in contrast to one-shot and few-shot settings that require significant pre-training. On standard benchmarks including MNIST, EMNIST-letters, and the harder Omniglot challenge, our model outperforms both neural-network-based and classical ML methods in the ``tiny-data'' regime, including few-shot learning models that use an extra background set to perform transfer learning. Moreover, mimicking simple clustering methods like $k$-means but in a non-Euclidean space, our model can adapt to an unsupervised setting and generate human-interpretable archetypes of a class.