Vahid Shahverdi: Geometry of Linear Convolutional Networks with Arbitrary Strides
Tid: Fr 2023-02-03 kl 15.15 - 16.00
Videolänk: Zoom meeting ID: 686 7101 5535
[Note: this seminar has been moved from January 27th to February 3rd]
In this presentation, we will explore the geometric properties of Linear Convolutional Networks (LCNs) that use arbitrary strides. These networks have become essential tools in image processing and computer vision, yet their geometric properties are not fully understood. By delving into the non-convex function spaces and singular points generated by these networks, we aim to improve the design and optimization of LCNs. Through rigorous analysis, we will reveal that the function space can be represented as a collection of polynomials forming a semi-algebraic set. By presenting the defining equations, inequalities, boundaries, and singular points, we will provide a comprehensive understanding of the geometric characteristics of LCNs that use arbitrary strides, which can lead to new ways of improving their performance.