Tangent Space

Tangent space, a concept from differential geometry, is used in various fields to analyze and manipulate data residing on curved spaces, rather than flat Euclidean spaces. Current research focuses on applying tangent space methods in machine learning, particularly for improving neural network training and optimization (e.g., through task arithmetic and Riemannian optimization), and in computer vision tasks like 3D reconstruction and optical flow estimation from omnidirectional images. These applications aim to address limitations of traditional Euclidean methods when dealing with data exhibiting inherent geometric constraints, leading to more robust and accurate algorithms in diverse domains.