Jorge Daniel Aguero Quinteros | 2025 I.S. Symposium
狈补尘别:听Jorge Daniel Aguero Quinteros
罢颈迟濒别:听Exploiting Symmetries in Training Convolutional Neural Networks
惭补箩辞谤蝉:听Mathematics, Computer Science
础诲惫颈蝉辞谤蝉:听Robert Kelvey; John Musgrave
The fields of Deep Learning and Abstract algebra are connected by the theory of symmetry groups. Convolutional Neural Network Layers are equivariant with respect to translations of an input image. Group Convolutional Networks are a generalization of regular CNN’s that are equivariant to the action of a general Group. I reproduce experiments that confirm an increase in performance of a CNN architecture by replacing convolutions with group convolutions.
Posted in Symposium 2025 on May 1, 2025.
