Zero-shot learning
In machine learning, zero-shot learning is the ability of an intelligent agent to exhibit learnt behavior in an unknown environments (or) unseen classes [1]. Zero-shot learning is also known as zero-data learning [2]. It cleverly combines the prior knowledge in one domain and extends its knowledge to other undiscovered domains. Geometric generalization is considered as one of the precisely evaluable tasks for zero-shot learning. Dataset Infinite World[1] was introduced to test the geometric generalization abilities of machine learning algorithms.
Background
A simple example of the human brain performing zero-shot learning lies in its ability to perform geometric generalization. A more advanced example includes prediction of gravitational waves via thought experiments [3]. The human brain performs zero-shot learning at ease through generalization and abstract reasoning over previously learnt classes / domains.[1]
Geometric generalization
In two dimensional geometry, the concept of connecting three lines to form a triangle and four lines to form a square can be generalized to the concept of polygon. This generalization can be extended to any number until infinity. Though such generalization looks seemingly simple for human beings, it poses a challenging task for machine learning algorithms.
Dataset Infinite World
Infinite World provides a multi-modal, light-weight and robust dataset for tasks such as image classification and text-to-image synthesis. It also defines a Zero-Shot Intelligence metric ZSI to evaluate the performance of Deep learning and other Machine Learning algorithms.
Prediction of gravitational waves with zero-data
Prediction of gravitational waves in the 20th century preceded its experimental confirmation in the 21st century [4]. Albert Einstein was one of the pioneers who mathematically formulated the predictions of gravitational waves. Though the prediction involved several rectifications and editions, since the mathematical formulation preceded the discovery, the learning is considered as zero-shot learning.
References
- ↑ 1.0 1.1 1.2 Chidambaram, Rajesh; Kampffmeyer, Michael; Neiswanger, Willie; Liang, Xiaodan; Lachmann, Thomas; Xing, Eric (2018). "Geometric Generalization Based Zero-Shot Learning Dataset Infinite World: Simple Yet Powerful". arXiv:1807.03711.
- ↑ http://www.deeplearningbook.org/contents/representation.html
- ↑ Einstein, A; Rosen, N (1937). "On gravitational waves". Journal of the Franklin Institute. 223: 43–54. doi:10.1016/S0016-0032(37)90583-0.
- ↑ Abbott, B. P; et al. (2016). "Observation of Gravitational Waves from a Binary Black Hole Merger". Physical Review Letters. 116 (6): 061102. doi:10.1103/PhysRevLett.116.061102. PMID 26918975.
This article "Zero-shot learning" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:Zero-shot learning. Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.
|
This page exists already on Wikipedia. |
