Abstract
In many machine learning tasks, the available training data has a skewed distribution- a small set of training classes for which a large number of examples are available (“base classes”), and many classes for which only a limited number of examples are available (fewshot classes). This is known as the long-tail distribution problem. Few-shot learning refers to understanding new concepts from only a few examples. Training a classifier on these fewexample classes is known as the few-shot classification task. Techniques disclosed herein improve classification accuracy for few-shot classes by leveraging examples from the base classes. A generative machine-learning model is trained using the base class examples and learns essential properties of the base classes. These essential properties, representing the intersection between base and few-shot classes, are applied to fewshot classes to generate additional few-shot examples. The generated few-shot examples are used to train a machine classifier to achieve better classification of inputs from few-shot classes.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Leung, King Hong; Toshev, Alexander; Hegde, Narayan; and Movshovitz-Attias, Yair, "Few-shot learning using generative modeling", Technical Disclosure Commons, (January 02, 2018)
https://www.tdcommons.org/dpubs_series/1024