Loading [MathJax]/extensions/MathMenu.js
Unsupervised Learning of Shape-invariant Lie Group Transformer by Embedding Ordinary Differential Equation | IEEE Conference Publication | IEEE Xplore

Unsupervised Learning of Shape-invariant Lie Group Transformer by Embedding Ordinary Differential Equation


Abstract:

Humans can recognize geometric transformations, such as translation or rotation, regardless of the original patterns. Transformation recognition is subsumed under the Lie...Show More

Abstract:

Humans can recognize geometric transformations, such as translation or rotation, regardless of the original patterns. Transformation recognition is subsumed under the Lie group theory. As is known, as humans grow, they acquire the transformation recognition ability and mental transformers to apply to any imagined objects. However, the underlying mechanism of transformation recognition remains unclear. Although transformation discovery models have been researched extensively, many of these studies used an unrealistic amount of dataset, which is unsuitable for modeling the transformation recognition ability of humans, who can immediately recognize transformations using only a few examples. Therefore, in this paper, we propose a new approach to learning Lie group transformers from a few sequences of images in an unsupervised way by embedding ordinary differential equation dynamics in the transformer, which can naturally be derived from the Lie group theory. Unlike the conventional models, our model considers objects as a collection of tiny points, to which it locally applies a transformation to ensure the invariance to the shape of objects. We train our model to discover the transformation in a custom dataset of translation, rotation, or scaling transformation and successfully acquire a Lie group operator that is applicable to unseen objects outside the dataset.
Date of Conference: 23-26 August 2021
Date Added to IEEE Xplore: 20 August 2021
ISBN Information:
Conference Location: Beijing, China

Funding Agency:


References

References is not available for this document.